Nishi Frontier Board

nx_frontier_board (F-202) from frontier_queue.tsv. span=14 sessions · ready=335 blocked=157 done=252 external=10

idwdurESEFfloatownerlanestatustask
F00180002operatorday0DONEVast.ai GPU key
F002800013operatorday0EXTERNALULX3S board
F00381015engineerday0DONEverify tree deletions
F004810113modelwrightday0DONEserve rebuild + remeasure
F005810113supervisorday0DONEreader-breaker deploy
F00681015censusday0DONEecomat re-seed
F007710113teamday0DONEautofix intake sweep
F20181125pmautoDONEfrontier store + queue
F20283255pmautoDONEboard organ CPM + emit
F20371237pmautoREADYws_put F-rows to WMS
F20481565conductorautoDONEcron + cadence hook
F204b71677conductorautoDONEfrontier pages crontab end-state
F20571568pmautoDONEMCP/API tool row
F20671238pmautoDONERACI store rows
F20772358hrautoREADYHR staffing join
F20882685pmautoDONEautonomous backlog feeder
F20971566librarianautoREADYevidence check-in
F21072029engineerautoDONEtools-plane admin API
F210b712311engineerautoDONEtools-register offc path fix
F211720211wardenautoDONEvault + ssh to MCP
F21281016conductorautoDONEkickoff-sync capture
F213712311wardenautoREADYsecrets portal vault surface
F214810112conductorautoDONEseat lease orphan reap
F21582136conductorautoDONEloop R1 cycle+debt stores + ws_cycle
F21681346conductorautoDONEloop R2 evidence advance + debt gate
F21782466librarianautoDONEloop R3 actlog + workflow miner
F21882686conductorautoREADYloop R4 role daemons zero-claude
F21972359researcherautoREADYloop R5 research-exceed arena
F101a81012engineer-gpugpuREADYGPU provision + deploy
F101b81122engineer-gpugpuBLOCKED F101afirst sovereign submit
F101c81232engineer-gpugpuBLOCKED F101bC0 FP32 GEMM
F101d82352engineer-gpugpuBLOCKED F101cC1 tensor-core BF16
F101e82572engineer-gpugpuBLOCKED F101dC2 90pct cuBLAS
F101f81782engineer-gpugpuBLOCKED F101eC3 determinism exceed
F103a91010engineer-ososDONENative Contract v1
F103b93140engineer-ososDONEpabi-native core byte-exact
F103c92460engineer-ososDONELinux guest lane
F103d93690engineer-ososREADYUEFI image on metal
F103e939120engineer-ososBLOCKED F103dsovereign transport
F103f9212140engineer-ososBLOCKED F103eos-compare measured
F104a72027engineer-compilercompilerREADYinliner miscompile fix
F104b71237engineer-compilercompilerBLOCKED F104ainliner enable + measure
F104c74377engineer-compilercompilerBLOCKED F104baarch64 backend
F102a62029engineer-browserbrowserREADYw=0 collapse + float
F102b61239engineer-browserbrowserBLOCKED F102aOCR recall ratchet
F102c610110engineer-browserbrowserREADYTLS session reuse
F102d62359engineer-browserbrowserBLOCKED F102bfont fidelity
F105a72020engineer-swarmswarmREADYworker auth/enroll
F105b71011engineer-swarmswarmDONEVRAM beacon
F105c71230engineer-swarmswarmBLOCKED F105aswarm multinode proof
F106a610110racingexperientialDONEconsole/error capture
F106b611210racingexperientialDONECI sweep orchestration
F106c622410racingexperientialDONEhead-to-heads vs SOTA
F107a73145engineer-siliconsiliconDONEcsr/clint/uart emitters
F107b72465engineer-siliconsiliconREADYtop CPU netlist
F107c72685engineer-siliconsiliconBLOCKED F107bPnR + bitstream
F107d71895engineer-siliconsiliconBLOCKED F107cULX3S boot
F301720210pmvalueREADYrevenue funnel measured
F302712310finvalueBLOCKED F301reinvest ledger
F30371568hrvalueBLOCKED F302living-wage model
F22072028engineersovgitDONEsovgit X2 multi-repo host + pack scale
F221712310conductorsovgitREADYark+pulse push lane to sovgit host
F22273258engineersovgitDONEsovereign git client legs retire usr-bin-git
F22361568pmsovgitBLOCKED F221gitea sunset after measured parity
F310820212researcherdmktDONEatlas recombination product ideator
F311720212engineerdmktREADYCAD parametric plus STEP export
F31282029engineerdmktREADYlive electronics sourcing connector
F31382029conductordmktREADYprint and quote connector home plus industrial
F314712311conductordmktBLOCKED F313prototype orchestration router
F31571239findmktBLOCKED F312finance cost-margin-viability rollup
F316820210pmdmktREADYpropensity plus VHECT demand scorer
F31781349pmdmktBLOCKED F315channel viability GO NO-GO gate
F31871459librariandmktBLOCKED F317expose board tool plus page plus raci ws rows
F501810113pmofficeREADYNishi Office/Legal suite program card
F108a71239engineer-browsercleanserveDONEcleanserve MCP receipt tool nx_cleanview
F50282027counselofficeDONEMatter and client spine parties conflicts engagement
F50373257counselofficeDONENishi Box legal DMS matter-scoped privilege Bates retention
F50483257counselofficeDONENishi Sign e-signature envelope hash-chain CID audit
F50572577counselofficeREADYDocument assembly legal templates clause library
F108b710113engineer-browsercleanserveDONEcleanserve durable supervision respawn checked
F50672248finofficeDONETrust accounting billing IOLTA RPC1.15 LEDES 3-way
F22481016architectatlasDONEatlas ingests new organs on cadence
F22582138researcheratlasREADYatlas recombine discovery organ
F22671895pmatlasBLOCKED F225discovery to frontier autotriage wire
F50772248counselofficeDONEDocketing deadline engine URCP rule-based SOL tracking
F50862577counselofficeREADYClient portal secure messaging per-client gated realm
F50962577counselofficeEXTERNALUtah court e-filing connector EFSP adapter
F51062577librarianofficeREADYRecords retention RPC compliance dashboard privilege log
F51181568researcherofficeREADYAtlas card + recombination register office nodes to ideator F310
F108c71349engineer-browsercleanserveDONEcleanserve atlas node pedigree card honesty
F108d720212engineer-browsercleanserveREADYcleanserve JS-exec media resolve framework sites
F108e71459researchercleanserveREADYcleanserve recombination hooks into atlas ideator F310
F227820210refereeautoDONEscale-envelope audit sweep of tool fleet
F228821311architectatlasREADYcoupling agnosticism flag in atlas cards+census
F103g820212engineer-ososREADYboot-verify payload NXE verify-then-execute never-brick loader K-R1
F23081014librarianrecallDONEsearch-SOTA IR eval ruler nDCG/MRR/Recall
F23182134librarianrecallREADYwire RRF hybrid fusion BM25+dense into SERP
F23283364librarianrecallBLOCKED F231sovereign semantic reranker PPMI-SVD+learned
F23382684librarianrecallBLOCKED F232nishi_recall MCP tool memory+source+store
F234828104librarianrecallBLOCKED F233atlas GraphRAG recall recombination retrieval
F23572686modelwrightrecallBLOCKED F232LLM-reranker escalation via no-float forward
F601810113engineer-videovideoDONEvideo: kill vroom verify-trap + vqoe standing gate
F602811211architectvideoDONEatlas-ingest video codec organs
F603813410researchervideoBLOCKED F225recombine: discover unproven video opportunities
F604810110engineer-videovideoDONEexperience: paced-keyframe slow-start kill 15s freeze
F605711210engineer-videovideoDONEexperience: link-aware capture-fps 15 to 30
F60672359engineer-videovideoBLOCKED F105c720p tile-parallel encode
F607720210engineer-videovideoDONEcodec: ship motion-partition vcv-7 + external BD-rate
F608622410engineer-videovideoDONEcodec: CABAC-class entropy generation
F609848122researchervideoBLOCKED F101fneural learned-representation codec DCVC-class
F610720212engineer-videovideoREADYsovereign STUN/TURN cross-network calls
F250720212librarianatlasREADYlivingdocs compare domain LIVE; atlas card-wiki + staleness next
F251820212conductorautoREADYhub the claims plane to NAS store
F290710110researcherresearchREADYresearch fetch sweep sprint-telemetry SOTA
F611710113researchervideoREADYresearch-fetch video SOTA frontier: DCVC-FM neural, learned entropy/hyperprior, VVC/AV1 tools, perceptual/GAN, sovereign STUN-TURN
Gx0800014engineer-graphics-READYsovereign graphics SOTA ruler + nx_graphics_board LIVE on MCP (644, research-opps)
Gx1800014engineer-graphics-READYSW-GPU into Nishi-Browser framebuffer + NishiOS (sovereign in-browser)
Gx2700014engineer-graphics-BLOCKED Gx1hardware-agnostic ~25-op IR + aarch64 (mobile+desktop)
Gx3800014engineer-graphics-READYreal-silicon GPU accel (rides gpu-lane; Vast credit ready)
Gx4700014engineer-graphics-BLOCKED Gx1d3d11-compat game path + flagship sovereign browser game
F26081126architectatlashygDONEAtlas-hygiene engine: program card + provenance/confidence plane schema
F26183366librarianatlashygBLOCKED F231Tri-layer dedup: SHA-256 collapse then semantic vector-cluster then ontological merge/link
F26272248refereeatlashygDONEContradiction/decay resolver: CONTRASTING_EVIDENCE edge + rigor-scored LEGACY/CANONICAL flip
F26382246wardenatlashygDONEZombie-loop firewall: source_type provenance ledger; block AI_SYNTHESIZED as primary validation
F26472466refereeatlashygDONEConfidence-weight gating: 0-1 edge weights; downstream exec blocked below 0.75
F26573259researcheratlashygDONEHallucination/citation firewall: Crossref/SemanticScholar/USPTO/SEC verify then quarantine unverifiable
F266722410pmatlashygDONECompute/token circuit-breaker: per-subgraph daily budget; pause+serialize+alert operator
F26782686conductoratlashygBLOCKED F261Nightly self-cleaning cycle: merge-phase then re-weight-phase then garbage-collect (capstone)
F107f73477engineer-siliconsiliconREADYboot emulator JIT maturity: MMIO-in-JIT + cross-block trace chaining + JALR calls
F107g73477engineer-siliconsiliconREADYboot RV64 SMP multicore (scale beyond the single-core envelope)
F700810113pmcoeREADYChain-of-Evidence suite program card: 5-pillar signed manifest across Evidence-Atlas-Scientist-Builder hubs
F70193036librariancoeDONEnx_coe_frame organ: emit+verify signed 5-pillar evidence frame over CID + ed25519-C2PA + tamper-store
F70282359wardencoeREADYCustody pillar: cap-chain + consent-ledger binding; every hub hop cap-gated + audited into the frame
F70382357refereecoeREADYCorroboration pillar: deny-by-default multi-source verify gate >=2 independent attestations + adversarial refute
F70482359engineercoeREADYMCP+API expose nx_coe_frame/nx_coe_verify via register-mint-call + /api/coe REST
F70582359conductorcoeDONEWorkflow+Agent wiring: frame emission as nx_plan_run step + auto-beat onto ws_sync per closed rung
F70682356architectcoeDONEAtlas sync: CoE organs ingested on F224 cadence as nodes+card; manifest lineage = trace-to-god
F70782576researchercoeBLOCKED F225Atlas recombination + beyond-SOTA research: PQ sigs, transparency-log inclusion proofs, ZK provenance, SLSA L4, C2PA 2.x
F70871786researchercoeBLOCKED F707Fetch watch-list: C2PA 2.x, SLSA/in-toto/witness, Sigstore/Rekor, PROV-O/VC, eIDAS/ETSI, NIST FIPS-204 ML-DSA, Cert Transparency
F70982577pmcoeBLOCKED F703Honest SOTA+maturity benchmark /compare/coe vs Sigstore/Rekor, in-toto/SLSA, C2PA, Hyperledger, OpenLineage/PROV-O, AWS QLDB
F710713410refereecoeREADYScale-envelope + never-brick tooth: frame emit declares production-shape envelope in output; append-only no-delete
F711710113conductorcoeREADYHub + coordinate: absorb FLOATING provenance stubs (content-credentials/lineage/secure-boot-attest/container-signing) + wire office F502, librarian F250, security
F71272359counselcoeREADYLegal admissibility profile: frame -> court-admissible chain-of-custody export (Nishi Sign CID hash-chain, Utah e-filing)
F236721311researcherrecallREADYBEIR external suite through nx_recall_eval honest C5 row
F23762359researcherrecallBLOCKED F231deterministic learned-sparse research spike beyond SOTA
F107h73477engineer-siliconsiliconREADYRV64 privileged S/U modes + A-extension atomics (SMP + real-OS prerequisite)
F107i62468engineer-siliconsiliconREADYGDB debug stub + snapshot/restore for the sovereign emulator
F730810110pmeffDONEefficiency telemetry program: spend planes energy/time/tokens from-god
F731811211censuseffDONEnx_eff_scan measured spend scanner worklog+transcripts laptop lane
F732811210pmeffDONEnx_eff_board organ on MCP register+mint+call spend board JSON
F733712311conductoreffDONEeff cron beat zero-claude daily spend rollup publish
F734712311librarianeffREADYatlas ingest eff organs + card resolves rides F224
F735822410pmeffREADYtoken+compute budget circuit-breaker wired into conductor+graph loops eats atlashyg seq53
F736711212researchereffREADYefficiency research-exceed fetch sweep beyond SOTA ties F290 F219
F320820211pmpmdashDONEPM dashboard organ compose frontier+ecomat+debt+roi
F321820212finpmdashREADYROI computer honest money-time-energy tagged
F322720212pmpmdashREADYfeed nishi_project the real program EVM live
F323712311pmpmdashREADYfeedback loop dashboard flags to backlog rank
F324712311librarianpmdashREADYexpose /pmdash page + nx_pm_dashboard tool
F252810113pmvalueDONEPM cockpit exec/mgmt/ops dashboard + honest ROI LIVE at /pm
F737720212pmpmDONEPM oversight dashboard exec-mgmt-ops + honest ROI ledger nx_pm_deck LIVE at /pm
F740070772pmdash-bench SOTA gap - less-than-SOTA-is-never-designDONEpmdash
F310b820211researcherdmktDONEnx_cap_harvest broad catalog from tool-registry + atlas 16575 nodes (SCALE)
F310c712311researcherdmktREADYrecombine richer multi-keyword weighted scoring
F741060682pmdash-bench SOTA gapBLOCKED SPACE DevEx signal - developer satisfaction and flow survey into the dashboardpmdash
F742070772pmdash-bench SOTA gap rides EVM feedBLOCKED cost-per-initiative and R-and-D capitalization timesheet-to-dollar rolluppmdash
F743060683pmdash-bench SOTA gapBLOCKED external source ingest connectors - git CI Jira incident into the intelligence planepmdash
F40181010modelwrightmodelDONEK3 arch census measure-first
F402831410modelwrightmodelREADYK3 KDA linear-attention no-float
F403721311modelwrightmodelREADYK3 Attention Residuals no-float
F40482130modelwrightmodelREADYK3 LatentMoE quantile-balance router
F405831410modelwrightmodelREADYK3 Per-Head Muon integer optimizer
F406721311modelwrightmodelREADYK3 MXFP4/MXFP8 microscaling dequant
F407721311modelwrightmodelREADYK3 1M-ctx long-RoPE YaRN/iRoPE
F40874159modelwrightmodelREADYK3 native multimodal vision tower
F41184377modelwrightmodelBLOCKED F404K3-mu train local on GPU seat
F744720210engineeraclDONEnx_clobber_guard: cross-tree source reconcile + pre-push OCC refuse (git-non-fast-forward for the ship path); eats seq71/68/96 dual-copy clobber class
F745720212wardenaclDONEnx_store_collision_check: detect duplicate-key/id overwrites across seg-store segments (the seq71 docs-vs-designmarket id17 signature) - scan + report + fail-closed refuse
F746730311wardenaclREADYclaims plane on MCP + hubbed to NAS: expose claim/beat/release/walk as MCP tools (peer to ws_sync) + move leash store NAS-side so any agent coordinates; eats seq37/seq39
F747620212engineeraclREADYsingle-writer hub regen server-side (flock) to kill the laptop-publish hub race; eats seq68 lost-update class
F748720212pmaclREADYanti-clobber PROGRAM CARD: compose guard+drift_watch+claims+autoid into ONE coordination-integrity capability; benchmark vs git-nonff/etcd-CAS/K8s-resourceVersion/CRDT; maturity domain + atlas card + RACI
F225a080862F225 real unblock diagnosed live by claude_f749ee6c; example lead=perceptual video recallBLOCKED nx_atlas_discover must WALK ecograph_full 16580-node graph not the saturated 7-verb commontask catalog - native graph-adjacency recombination is the unblock (alt-catalog via nx_store_put fails-closed W002 seg-vs-plane)atlas
F612720212engineer-videovideoEXTERNALvideo fps WIN: fix reactive ladder to climb toward cap on good links (F605 capture-raise shipped but no-op until this; the real fps limiter)
F749622410engineeraclREADYwire nx_clobber_guard into nx_ship pre-upload + ecomat/roi seed paths (call cmp; abort push on DIVERGED) = the OCC refuse ENFORCEMENT half of F744
F108f731410engineer-browsercleanserveBLOCKED F102ccleanserve TLS session resumption crypto perf (certloop ~1s to <100ms)
F108g620212engineer-browsercleanserveREADYcleanserve filter-list breadth to EasyList-scale (tracker/miner)
F108h630311wardencleanserveREADYcleanserve fingerprint-defense canvas/WebGL + CF-challenge passage
F108i630311wardencleanserveREADYcleanserve WARC link-rot archive at Browsertrix/Wayback scale
F268712310wardenatlashygDONESafety guardrail: block restricted biosec/finrisk footprints on ontology insertion
F613810111researchervideoREADYSOTA test-corpus ingest: HEVC conformance streams (libde265) + DCVC test sequences (UVG/MCL-JCV/HEVC-ClassB) as sovereign benchmark ORACLES (their test data + published numbers, never their code)
F614721311engineer-videovideoBLOCKED F613adopt SOTA integer-native techniques from the repo review: classical (partition-tree, CABAC-class entropy, compound low-delay-B) + neural (DCVC-RT, integer entropy per Balle-2019) - each gated by BD-rate vs the ingested oracles
F26981677wardenatlashygDONEUnified admission-control gate: safety+confidence+provenance in one insertion decision
F253810113librarianatlasDONER6 walkable card-wiki LIVE (16599 organ cards, doc==unit) /compare/atlas/card
F71382359refereecoeDONEtransparency-log: Merkle inclusion + consistency proofs over coe_signed.log (Rekor-class STH witness) - the biggest SOTA gap
F750720212supervisorgovDONEnx_govern_sweep: supervisor crash-loop detector (respawn-without-backoff vs K8s reconcile-loop SOTA) -- built gate 2/2 + real-log dogfood respawns=23 UNGOVERNED; deployed MCP callable+cap
F751820212supervisorgovDONEreader-keep crash-loop FIX: add exponential-backoff circuit-breaker/quarantine (torrent guard already BACKS OFF; reader-keep respawns every time = the live resource failure); eats seq30
F752720212supervisorgovDONEsupervise-proc reconcile-to-desired-count: ~13 nx_hostctl supervise procs vs doctrine ~2; declare desired state + reconcile (K8s replica-reconcile pattern)
F753630311conductorgovDONEcron/clock governance: every scheduled job reconciled to cron.reg + heartbeat-liveness (no dead cron.d, no corrupt clock rows seq83) + govern-sweep on cadence
F754720212pmgovREADYgovernance PROGRAM CARD: the reconciler (observe->compare-desired->correct+backoff) benchmarked vs K8s control-plane; + /governance page + maturity domain + atlas card
F755720212engineerbuildroot-guardDONEbuildroot integrity guard nx_buildroot_guard self-healing lib-drift LIVE MCP+cron
F251b070772coordination bench gap: claim-beat-on-MCPBLOCKED claim/beat/release WRITE verb on MCP - flock-safe ws_sync append tool (today write=nx_fs_write editauto
F756060682coordination bench gap: human-UIBLOCKED human-in-loop coordination UI - interactive assign/comment on /standup or /org (today read-only)auto
F757070771supervised surface-liveness sentinel - DONE proven 5of5 200BLOCKED surface-sentinel workflow LIVE: plan-surfsentinel- verifies all published surfaces 200 via one nx_plan_run; hourly cron :15 supervises the suite; catches edge regressions zero-Claudeauto
F760800014pm-DONEpublishing program: elite-publisher research + program card
F761800014librarian-DONEpub- registry seeded: every live page owner+section+status
F762800014engineer-DONEnx_pub_desk organ: census+board gate 15/15 MCP-live /publishing
F763700014engineer-READYpublish gateway: registration-refusal + link/schema check in pipeline
F764700014conductor-READYpublishing cron beat: zero-claude board re-emit + orphan auto-debt
F765600014warden-DONEdebris eviction: 18 debris rows out of docroot w/ redirects
F766700014pm-READYfronts consolidation: index/home/hub as data containers + redirect table + withdrawals
F767700014librarian-READYfreshness: last-publish stamps + review-by dates + stale auto-debt sweep
F768600014researcher-READYfeeds+sitemap from registry + publish metrics + AB-testing exceed research
F27082686architectatlashygDONEAuto-sweep: judge every real node in a plane by its own stored confidence+provenance (guards over the real graph, not args)
F71472359engineercoeDONEin-toto/SLSA attestation export: emit CoE frame as ITE-6 Statement for external verifiers (interop boundary, export-only)
F769720212engineerapi-contract-guardDONEAPI-contract regression guard nx_api_contract detects dropped routes rule-19 LIVE MCP+cron+/pm
F770720212engineerdeploy-readyDONEPre-deploy safety gate nx_deploy_ready aggregates buildroot+fail-loud+contract into one DEPLOY-SAFE/BLOCKED verdict LIVE MCP
F350310113engineerplatformDONEnx_job_run async job-runner (fetch-scoped, 4-gate guarded) LIVE+PROVEN: eats async-fetch >15s edge-timeout debt; nx_plan_run jobtest step1 rc=0 JOB-STARTED pid8037 nx_supply_research_fetch + step2 rc=6 REFUSED non-fetch
F771720210censusaclREADYnx_ecomat_put: store-native autoid domain adder (mirror nx_frontier_put addauto: atomic ec_key(max+1) under flock -> ss_add+ss_commit) + relax nx_ecomat_seed ver-check ==N->=N = store-as-SSOT transition (F215 pattern). Kills the ecomat two-edit-surface clobber AT THE PRIMITIVE; unblocks governance->maturity + all future domains collision-safe. Touches the contended seeder = careful fresh-session arc, use nx_clobber_guard to verify convergence.
F772070772debt-hygiene organ - DONE built+promoted+registered+MCP-live+used (195/144/51/51)BLOCKED nx_debt_hygiene MCP tool LIVE: reads mixed-schema debt- plane -> total/open/eaten/gating[sev>=6]/histogram; replaces one-off analysis; built API-pure first-try 86875Bauto
F773620212supervisorcron-watchDONECron freshness supervisor nx_cron_watch dead-mans-switch detects silently-stopped automations LIVE MCP+cron+research
F774722410censusaclBLOCKED F771F771b: nx_ecomat_seed -> RMW (load-existing+merge, NOT replace-wipe) OR retire to store-SSOT, so nx_ecomat_put additions are DURABLE against a manual reseed (nx_ecomat_beat is READ-only = cron already safe, verified). THEN add governance->maturity domain via the put (collision-safe). Gate on throwaway first (seeder replace-wipe already caught there).
F775720212supervisorreliability-planeDONEReliability plane nx_deploy_ready aggregates 4 silent-failure guards into ONE managed gate + atlas cards + cron-published verdict
F351210113engineerplatformDONEnx_mkdirp sovereign mkdir-p MCP primitive (guarded knowledge/|sites/|/tmp/, no dotdot, idempotent) LIVE+PROVEN: filled the missing-mkdir gap -> created knowledge/fetched -> eats seq191 fetch SAVE-FAIL; 21+ real .raw bodies now persist
F776620212engineersendingREADYsending SOTA climb beyond 514: gift-reco + variable-data mail + send-time optimize
F777620212conductorautomationREADYautomation SOTA climb beyond 676: managed-scale runners + connector marketplace breadth
F778620212janitorjanitorREADYjanitor SOTA climb beyond 583: dep-update automation + secrets scan + SBOM + taint rules
F779620212librariangenealogyREADYgenealogy SOTA climb beyond 444: semantic code-graph + dataflow + typed call-graph
F780620212engineerdataengREADYdataeng SOTA climb beyond 586: close 12 PARTIAL + 6 GAP axes (SQL surface, lineage depth, viz)
F78182025librarianprimDONEprimitive registry BOTTOM-UP: organ-emitted knowledge/store/prim- capability rows (id/domain/capabilities/signature/deps) - tier1 MCP tools from tool_allowlist+tool_schemas typed grammar, tier2 16k organs via ecograph deps + nx_cc front-end exports parse (OUR AST, no tree-sitter); embeddings ride F231
F78282245engineerprimDONEquery_primitives MCP verb: capability-faceted search over prim- + ecograph adjacency = the what-do-I-have half of meet-in-the-middle (today only name/text via nishi_search/eco_graph_card)
F78383475researcherprimDONETOP-DOWN intent decomposer: goal -> required-capability list -> F782 match -> MISSING-PRIMITIVE EXCEPTION auto-files debt+frontier rung with structural spec (inputs/outputs/placement) = the meet-in-middle collision; composes F225 recombine + F310 ideator
F78472795conductorprimDONEsynthesize_workflow: matched primitives -> plan- rows glue (workflow-as-data via nx_plan_run/nx_workflow, organ composition = the OO chain) -> execute+verify; sovereign orchestration NOT LangGraph/Temporal
F78582029engineermodelcardDONEnx_model_card organ: emit /modelcard from the modelcard- plane (Inkling-parity 8-section card, 25-bench 9-model matrix, honest NISHI col MEASURED/ANALOG/UNMEASURED) + freshness re-emit beat + site_publish + pub- registered
F78673259refereemodelcardREADYseat-eval harness: HLE/AIME/GPQA/IFBench/Global-MMLU-class reasoning+chat battery over engine seats, fresh-run scored, writes measured NISHI col into modelcard- plane
F78784047engineermodelcardREADYagentic-coding parity climb: SWE-bench Verified-contract harness onto the PUBLIC 500-instance set + TerminalBench-class terminal analog (from 71pct sovereign foreign-bug analog; agent-parity honest, contamination-zeroed scoring)
F78873259refereemodelcardREADYagentic-general harness: MCP-Atlas/Tau/Toolathlon-class eval over the OWN 111-primitive MCP fleet (native testbed) + BrowseComp-class browse tasks via sovereign browser lane; measured NISHI col into modelcard- plane
F789622410librarianmodelcardREADYfactuality harness: SimpleQA-class graded factual QA wired to nx_recall_bench/BEIR planes; measured NISHI col into modelcard- plane
F790622410wardenmodelcardREADYsafety battery: FORTRESS/StrongREJECT-class adversarial+benign refusal eval over seat+cap layer (structural never-brick/ocap already EXCEED; measure the behavioral half); measured NISHI col into modelcard- plane
F791430311modelwrightmodelcardREADYvision+audio perception harness: MMMU/Charxiv/MMAU-class eval; media perception currently out of ecosystem scope (served not perceived) - honest UNMEASURED until a perception lane exists
F79283360modelwrightswarmBLOCKED F105csovereign split-model serve: layer/expert-shard no-float LLM across fleet nodes (exo/Petals = ruler not design; LAN-first chunky pipeline stages per Amdahl law; composes swarm place+job + no-float serve)
F793720212conductorswarmREADYfusion council serve: N heterogeneous model seats on different nodes + independent judge synthesize (OpenWebUI-fusion = ruler; composes agent-seat S2 + swarm coord llm lane + autofix best-of-N judge)
F794822410engineer-swarmswarmBLOCKED F105arented-node fleet enrollment: Vast/cloud box germinates spore -> enroll attenuated cap -> beacon rows -> placement-eligible worker (own+rent unified over WAN, never-brick, scale-to-zero on $ meter)
DR-052020researcherdeepresearchREADYExternal ruler run real DRBench+DEER measurement-first
DR-631230pmdeepresearchDONEOpportunity-cost research ranker organ OCM
DR-1413410engineer-recalldeepresearchDONESovereign Elo tournament ranker Co-Scientist core
DR-342350researcherdeepresearchDONECitation-entailment verification loop DEER back-track
DR-253580researcherdeepresearchDONEReflection Critic refutation upgrade 0-vs-9 gap
DR-4428100conductordeepresearchDONEAsync research conductor topology + FutureWeaver budget
DR-73210122researcherdeepresearchDONEWeb-vs-private source routing DRBench #1 failure
DR-55410140modelwrightdeepresearchREADYSovereign LLM synthesis stage 0-vs-7 nofloat-gated
F615710113engineer-videovideoREADYvideo: BWE default-on A/B under induced congestion (CDP network-throttle harness driving probe pairs bwe=0 vs bwe=1; default-on only on measured win; probe834 already proves non-harmful on clean links)
F795820212refereeeditorDONEeditor failure-mode COVERAGE manifest (ecosystem-as-editor law): store plane editcover- mapping each reported LLM failure mode (dithering/fake-DONE/patch-cascade/silent-truncation/stale-memory-assert/sycophancy/over-ask) -> covering structural organ -> evidence log path; gap rows auto-file F-rungs; board emit to /editor
F796820212refereeeditorREADYedit-lift meter + per-seat trust gauges: measure draft-vs-post-editorial delta per task (generalize autofix best-of-N+judge over the seat contract); per-seat gate-pass rate / rejected-DONE-over-debt rate / tokens-per-closed-rung -> evidence layer + /standup + modelcard NISHI col; trust MEASURED never asserted
F79751017netobsnetobsDONEnetobs sovereign observability stack v1: nx_netobs organ (beat/burst/page/bench/verdict/selftest) + METRIC ledger w/ rotation + /netobs correlative dashboard + loadgen percentiles + data-driven verdict gate; wired clock/cron_watch/deploy_ready
F79842137supervisornetobsREADYnetobs sub-second STANDING sampler: supervised daemon (hostctl-guarded) sampling cpu/net at 250ms continuously w/ ring-buffer ledger segment; retires the bounded burst verb for load windows
F79953147edgenetobsDONETLS-edge handshake observability + certloop cure: MEASURED certloop=1385-1606ms per sovereign TLS handshake (nx_https_get probes 07-20) = the dominant edge latency; add tls-handshake timing probe mode to nx_netobs + a certloop cache/fix arc on the managed edge lane (crown-jewel change, never-brick deploy)
F800431410fabricnetobsREADYcross-node GPU/laptop metrics into netobs: RTX-5080 VRAM/util + laptop cpu ride the swarm fabric beacon feed (F792-F794 lane) into the netobs ledger = the DCGM role, honest per-node attribution
F80134377netobsnetobsBLOCKED F798kernel-queue latency attribution (eBPF-class, sovereign): research + build accept-to-first-byte vs socket-queue-time split via SO_TIMESTAMPing/poll instrumentation in the serve organs; answers WHERE a slow request waited
F80243477netobsnetobsREADYclosed-loop auto-tune actuator (never-brick BY CONSTRUCTION): verdict breaches drive DATA-row changes only (svc-config/registry rows, reversible, deploy_ready-gated) - never sysctl/fw writes; loop = bench -> verdict -> tune-row -> re-bench w/ auto-rollback on WORSE trend
F803720212refereeeditorDONEpatch-cascade structural tooth (editcover row4 GAP): rule-3 mechanized - the cycle/journal layer flags N>=2 consecutive fix-beats on the same target without a rewrite/gate artifact and the debt-gate refuses the 3rd patch until a rewrite note or gate-GREEN lands; composes ws_sync beats + evidence-phase R2b
F616710113wardenvideoREADYbench-only quarantine tooth (OPERATOR LAW 07-20): own-versions-of-external organs marked BENCH-ONLY ORACLE; sweep flags any marked source appearing in tool_allowlist.conf / proxy_routes.conf / daemons.reg / client wasm manifest; violation -> sheriff + deploy_checks RED; applies to ANYTHING we have, not just codecs
F804830311engineergovREADYseq162 auto-rollback half: on a TRUE api-contract REGRESSION (fail-static detector now trustworthy), auto-remediate ONCE via the existing never-brick /api/rollback then re-probe + alert-if-still-broken (bounded, no loop). SAFEST design = fold the route-probe INTO the mgmtapi never-brick health check (nx_deploy_lib dep_health) so a route-dropping deploy fails health -> the EXISTING rollback fires (no new authority, no cron escalation). Careful crown-jewel supervisor arc = dedicated fresh session
NB1800014engineernsbuilderDONEsymbolic judge organ nx_symjudge: property-contract sweep from symprop- plane, EXH bounded check, overfit-kill, crash-detect, gate 9of9 GREEN
NB2800014engineernsbuilderDONEautofix 5b SYMJUDGE stage wired: tests-green maker fix must also hold property contract else DEMOTED to MISS; ledger symj field; validated 2of2 live episodes
NB3800014examinernsbuilderDONEsymprop contracts for the FULL foreign-bug instance set + official sweep re-measure under the stronger neuro-symbolic judge (honest: resolve pct may DROP = truth)
NB4600014engineernsbuilderDONEsymjudge harness watchdog timeout (hang class) + sov_build_run rc surfacing
NB5600014engineernsbuilderREADYmulti-fn closure extraction: judged fn may call sibling candidate fns (v1 refuses/self-contained only)
NB6700014engineernsbuilderREADYship nx_symjudge to NAS + MCP register (fs_write update path per seq124 page-cache/treepack-no-overwrite debt; register+mint pair per audit law)
NB7800014modelwrightnsbuilderREADYmaker climb: 7B-class coder gguf seat (Qwen2.5-Coder-class, 16GB VRAM budget) under the SAME judge cascade + best-of-N; agent-seat S2 composes; operand-binding class unblocks
NB8700014researchernsbuilderREADYproperty VM v2: metamorphic pair sweeps + coverage-guided input mutation (AFL-class, deterministic seeds) + DaeDaLus-class parser-contract rows, all data-driven
F805720212engineertoolsplaneREADYcompiled-in shrink enforcement (seq162 final half, mgmt-owned per route_diff design note): ma_do_deploy/promote runs route+tool snapshot->deploy->check INSIDE the transaction and AUTO-ROLLBACKS on shrink (the transaction knows intent, a cron cannot); rides next mgmt touch (migrate-on-touch); detection pair nx_route_diff+nx_tooldiff live */10 + deploy_checks
F617810113engineer-videovideoREADYvideo: vcv-10 SHIP to client — vv_seed_rctx gains rctx[6] MV row-plane (CX+0x32000, clear of 150KB worst rc bytes), wasm build from C:/nishi-core/nxc2 (the codec SSOT), client rct9All generation gate + txMode wire, probe pair, vqoe GREEN; MEASURED native win -11.9/-19.9/-20.6 pct matched-PSNR bytes vs rct8 (bgop 2026-07-20, par=0 all arms)
F80673477engineermodelcardBLOCKED F787SWE-bench pytest ORACLE runner (oracle-only law: python executes FAIL_TO_PASS/PASS_TO_PASS as ground-truth oracle, writes verdict rows to swebvres- via nx_store_put; Nishi owns harness+scoring via nx_swebv_grade) - env options: laptop WSL py-env / rented compute seat
F830810111pmseatsDONEseat S0 boot-pack organ nx_seat LIVE: one-call 1195B capped digest (phase+debts+active-ws+frontier+meter) replaces ~40K-token session boot; gate GREEN on-hub; cap in token-engine memory
F831721311pmseatsDONEseat S3 metering pipeline: per-session transcript harvest cron feeds nx_seat meter; tokens-per-closed-rung on /standup; trend down release-over-release
F832720210engineertoolsplaneDONEenvelope caps on fat tool surfaces per seq28 scale-law: frontier_board 33.7KB/call digest verb + nx_status brief verb + debt list capped default; nx_seat boot = the shipped exemplar
F833821311engineercoordinationREADYseat S1 checkin/checkout verbs on nx_seat + claims to MCP with NAS hub claims store (eats seq37+seq39): claim+kickoff one call, done+release one call
F820310112pminfoplaneDONEdecision-ADL registry: decreg- plane (id title driver context alternatives supersedes epoch) + supersedes-chain verbs on nx_info_registry
F821420212supervisorinfraREADYinfo-plane replication: mirror knowledge/store planes to a second node (RK008 NAS-SPOF mitigation, riskreg-sourced)
F822210113sitesinfoplaneREADYregistries freshness beat: cron re-emit /registries page + selftest sanity zero-Claude (frontier-pages pattern)
F83432468netobsnetobsDONEtlscache rollout to remaining TLS-client callers: wire nx_tls_cert_cache (tcc_load/arm/save) into nx_mgmt_client + nx_health_eval he_get_lo + nx_content_ship self-mint fetches (each builds its own TlsValidationContext); plus nx_netobs tls-handshake probe mode so certloop is a tracked series
F823310113pminfoplaneREADYKPI-goal guardrail registry: kpireg- composing maturity-rollup + honest-ROI targets w/ tolerance (no-fake-numbers law; blueprint concept mapped sovereign)
F319710113findmktDONEproduct viability scorecard nx_product_score (coverage+readiness+GO-NO-GO, no-fake-numbers) DONE; next=multi-product ranking + run the research-fetch caps for real demand/cost
F836522410engineertoolsplaneREADYremaining fat-surface caps (frontier digest DONE via nx_seat ready 33.7KB->0.65KB): nx_status brief verb + nx_debt list default-cap + https_get->page_verify verdict nudge; each declares envelope
NB9700014engineernsbuilderDONEpoint-anchor oracle (anchor:x:want / anchor2:a:b:want) checked OUTSIDE the sweep=always-fires regardless of stride; combined w/ odd+lin2 pins whole fn; closes neg-vs-identity gap that odd/lin2/fix0 alone missed; gate 13of13; live neg contract strengthened+proven identity=RED negation=GREEN
F837720212conductoreditorREADYpatch-guard ENFORCEMENT wiring (F803 second half): nx_ws_cycle next/board consult nx_patch_guard check alongside the scoped debt-gate so a REFUSE actually blocks the next patch (today the tooth is provable but advisory); + seats append patch|rewrite|gate rows at fix time (nx_seat checkin/checkout hook)
F838846100modelwrightswarmBLOCKED F792F792b split-model SERVE implementation: layer-sharded LLM pipeline per the nx_swarm_gpu shard plan -- load model layers onto each shard node, pipeline activations over LAN (binary-framed /api/rpc floor), compose nx_nofloat_serve as the per-shard executor. Keystone (inventory+planner) DONE+LIVE 2026-07-20; this is the real-inference piece. Needs a 2nd fresh GPU (west re-probe OR F794 rented enroll)
DR-84110112conductordeepresearchDONEOne-call end-to-end chain judge-conduct-score + budget-cost measurement
F824311212pminfoplaneREADYADL-consult gate: surface in-scope accepted decisions (decreg- by scope) in nx_seat boot + nx_ws_cycle next so a session cannot re-litigate a settled decision - the DEC010/DEC011 class
F83952027netobsnetobsDONEHTTP keep-alive connection reuse (census gap #3 EATEN): nx_kaprobe organ proves + tracks 1-handshake-K-GETs vs K-handshakes; MEASURED 5-GET batch 653->106ms = 547ms/84% saved; server honors keep-alive; nx_https_req_complete session-reuse-after-cl-stop verified correct; /netobs section + verdict-gate 7th check keepalive_saved>=100 (regression tooth) + clock 600s beat
F84043259netobsnetobsREADYkeep-alive PRODUCTION ROLLOUT (F839 proven the primitive; adopt it): promote nx_http_client_build_request_keepalive + nx_https_get_complete_keepalive into the SHARED client, then reuse ONE session for the multi-fetch consumers = nx_health_eval he_check (9 edge fetches/pass -> 1 handshake) + deep-research crawler (N pages/host) + nx_content_ship verify loop; measured saving ~110ms per avoided handshake
F84165277edgenetobsREADYTLS 1.3 PSK session resumption / 0-RTT (the LAST big handshake win, census #1-remaining): server issues NewSessionTicket + client stores+presents pre_shared_key -> a resumed connection skips the CERTIFICATE and the mandatory CertificateVerify entirely (~0-RTT), beating even the F799 cache + F839 keep-alive for NEW connections. Two-sided crown-jewel crypto (psk binder HMAC + resumption key schedule on nx_tls13_server AND client); heavy KAT gating, never-brick deploy, brand-critical. Recon: nx_tls13_ext/server/schedule exist; check psk surface first
F845720212engineermgmtapiDONEseq236 ROOT-CAUSED by source: ma_do_build captures to FIXED shared /tmp/nx_ma_build.out then scrapes FIRST size= -> ANY concurrent /api/build corrupts bytes+BUILD-FAILED diag BY CONSTRUCTION. FIX (rule-19 identical JSON): uniquify capture path per call + derive bytes by stat of .sov.elf.new (both request-local like staged already is). RESIDUE: observed staged=OTHER target is NOT explained by this (source composes staged from request-local nm) -> transport response cross-wire suspected, ties D008 dual-443; do NOT close seq236 on the /tmp fix alone - needs a coordinated concurrency probe. CROWN-JEWEL: land under coordination, ship from buildroot SSOT never a stale tree (route-revert injury)
F846730311refereeeditorREADYenvelope-declaration remediation (F227 second half, MEASURED baseline 447permille): drive the 58 silent-cap candidates to declare their envelopes in output under MIGRATE-ON-TOUCH (any organ you edit gains its declaration - the D001 gate-lib pattern, never a unilateral 58-organ clobber sweep); reliability-plane guards FIRST (nx_deploy_ready nx_api_contract nx_cron_watch nx_debt_hygiene - a guard that truncates its own verdict is the worst case); ratchet floor set only after referee reviews the baseline
DR-94111122researcherdeepresearchDONEFile-backed corpus mode - real on-disk documents, the scale unlock
F847730311librariancoordREADYMEMORY.md ACTIVE-section conflict-free budgeted RAIL (SOTA memory-index fix, operator-flagged): put curated ACTIVE pointers on the same rail model the COINDEX LIVE block already proves -- append-only store plane memidx-active- (priority=stars, budget, auto-evict-oldest-low-star to overflow, dedup-on-emit merge same-file-link) + nx_memindex_emit v2 splices a DERIVED block between markers (never hand-edited) + MCP append verb (no Edit) + cadence beat re-emit+compact. Kills the 10-session clobber-collision + unbounded-growth class (evidence seq255: 4 failed compactions/one session). Separate workstream per operator.
F825410112librarianinfoplaneDONEmemindex curated-section rail: budget+auto-evict+ATOMIC nx_memindex_put verb for the hand-curated ACTIVE block (kills silent tail-truncation past the 24.4KB read limit + hand-Edit clobber races)
F851710113engineertoolsplaneDONEnx_fs `lines [count]` LINE-ADDRESSED read LIVE (D): the missing primitive - grep reports file:LINE but read took BYTES so the two could not compose; extracting ONE function from a remote file cost a subagent 70K tokens + 22 calls, now 1 call / 869B (~320x). Declared envelope (start/lines/next/scanned/scan_cap + FILE-EXCEEDS-SCAN-WINDOW + BYTE-CLIPPED flags) = scale-law, never silently windowed. Beyond-EOF fails LOUD. Inherits the secret deny-list (live-proven vs the real signing key = no key-theft primitive). Gate teeth T17-T19 added (19 checks, compiles); schema-bridge + reconcile done
F826311212librarianinfoplaneDONEatomic curated-append verb nx_memindex_put (same journal-lock + tmp/fsync/rename path as the derived splice) so sessions never hand-Edit the shared index - the seq256 contention half
F85281019pmosbenchREADYosperf matrix bench rig program card: 4 evidence classes MEASURED-TWIN-DIFFERENT_ALTITUDE-UNMEASURED
F853821311engineer-ososbenchBLOCKED F852nx_osbench_micro: our-side kernel primitives over nx_pabi seam + binary-SHA256 identity tooth
F854821311refereeosbenchBLOCKED F852nx_osbench_oracle: foreign static oracles inside nx_container rootless chroot BENCH-ONLY never run-lane
F85582139refereeosbenchBLOCKED F852nx_osbench_judge: 4-class verdict + superiority index; weights-as-data rule11; 3 neg controls
F856730311engineer-ososbenchREADYbare-metal x86 node bring-up (never-brick rule26 gated, removable media only) + RAPL energy axis
F857620212engineerosbenchREADYAndroid LAN device as third MEASURED target (real device not x86 emulator)
F858720212counselosbenchREADYnx_os_twin: lawful digital-twin plane for iOS-macOS (published figures + spec, content-addressed via nx_benchmark_ref, NEVER printed as MEASURED)
F85972359librarianosbenchBLOCKED F855/compare/osperf domain + nx_osbench_board on MCP (NEW domain; /compare/nishios stays emulator-vs-emulator)
F860813410censusosbenchBLOCKED F855osperf_extbench.proof = FIRST proof file in ecosystem; un-caps L5 maturity; supersede nx_os_compete narrative axes with measured rows
F861630311wardenosbenchREADYdifferential fuzz / mean-time-to-crash resilience axis (decoder 349k inputs traps 558-to-0 is the starting point)
DR-10512311librariandeepresearchDONESemantic judge tier - PPMI late-interaction over the real shipped model
F86242029netobsnetobsDONEnetobs config TSV->seg-store migration (no-tsv law, ATE MY OWN fresh debt created today): netobs_probes.tsv + netobs_thresholds.tsv now in sovereign planes knowledge/store/netobs-probes-/netobs-thresholds- via sts_seed/sts_load (THE canonical store-seed codec); nx_netobs reads plane-first w/ tsv-fallback; PROVEN plane seg-files on disk + seedcfg GREEN + production verdict GREEN 7/7 reads thresholds+probes from plane
F86333259netobsnetobsREADYnetobs config FULL TSV retirement (follow-on to F862): the netobs_probes/netobs_thresholds .tsv remain as demoted zero-downtime FALLBACK + seedcfg source. To fully retire per no-tsv law: add a native-edit path (verify nx_store_put putn writes sts-compatible q:seq/q:n keys like modelcard, OR add nx_netobs addprobe/addthresh verbs that sts-append), THEN delete the fallback .tsv. ⚠seedcfg footgun: never re-run against an EMPTIED tsv (sts_seed 0 rows wipes the plane) - safe only while the tsv keeps its data
F827310110censuslawDONEnx_magic rule-11 FIXER shipped: map/propose/apply const-hoist with byte-identical-rebuild neutrality proof (the missing half of L001 - warden detects, cap-autonomy files, this FIXES)
F828531410censuslawREADYL001 SWEEP under migrate-on-touch: each owner runs nx_magic apply+rebuild+byte-compare on organs they touch (start nx_law_warden 1 site, nx_archive_daemon 61/14) then RENAME mechanical consts to domain terms or config rows - never a unilateral 600-organ clobber
F864620212engineereditorDONE/editor adopts nx_ui_kit (seq247 operator #1-gap remediation applied to MY surface): nx_ui_kit + nishi_ds.css are now LIVE (sibling ui-design-system-u2 shipped; registered GREEN in the allowlist), so nx_editor_board should emit THROUGH the kit instead of its hand-rolled dev-grade table; DONE-WHEN nx_ui_audit scores /editor at the kit's band (measured before/after, not asserted) and the page still passes its own selftest teeth
F8658410140modelwrightswarmBLOCKED F838F838 REAL-MODEL sharding (seam FOUND 2026-07-20): wire the split-serve machinery (proven on the toy transformer) into the REAL no-float GGUF LLM. THE SEAM: nx_nofloat_llm.nx run_stack(buf,hdr,x,out,wb,...,N,rev) LOOPS layers 0..N + nf_load_layer_pool(buf,hdr,L,wb,...) loads ONE layer L by index. BUILD: (1) run_stack_range(a,b,x) runs layers [a,b) on an input hidden state loading ONLY those layers (mirror nf_load_layer_pool per-L in [a,b)); (2) each shard node loads ONLY its layer range's GGUF tensors = the REAL VRAM saving on a real model (a 30-layer Qwen split 0-14/15-29 across 2 nodes); (3) handle KV-cache per-shard + the activation transport at real hidden-dim (binary not hex); (4) prove split==whole bit-identical on the real model. Composes the DONE machinery: nx_swarm_gpu plan + nx_swarm_shardserve run/runf pattern + nx_split_pipeline driver. Deep integration = fresh focused session.
F866530311netobsnetobsDONEnetobs ramping loadgen = operator's Step A (progressive ramp to saturation): nx_netobs `ramp` verb, DRY bench-core refactor (bench+ramp share no_bench_run). PROVEN real saturation curve on gallery :18090: knee at conc=64 (p99 7.6->67ms breaches 50ms SLO), MAX SUSTAINED conc=32 @ peak ~1259 rps; past the knee latency AND throughput both regress (textbook). Emits netobs_ramp_max_sustained_conc/knee_conc/peak_rps series + /netobs capacity section + verdict floor-check (min conc 4). ⚠ramp is a REAL load event (conc=64 saturates host) -> ON-DEMAND ONLY, run sparingly, NOT a frequent clock beat (proven: repeated ramps spiked host cpu->100% load->134, analyzer correctly caught it; no worker leak)
F620910113supervisorvideoDONEOFF-BOX watchdog: nx_edge_watchdog LIVE on the laptop (schtasks NishiEdgeWatch every 5min, conhost headless) probing the PUBLIC WAN surfaces with CONTENT proof (200 + needle, never port-open); self-clearing latch + desktop escalation. Born from the 2026-07-20 NAS outage where EVERY monitor lived on the box being monitored so nothing alerted. PROVEN through a full real cycle: DOWN -> latch -> escalate -> RECOVERED -> auto-clear
F61894047engineer-compilervideoDONEvideo FPS NORTH-STAR: wasm SIMD in the sovereign toolchain. MEASURED: nx_compile_wat/nx_wasm/nx_wat_compiler emit ZERO v128 (grep v128/i32x4/i8x16 = 0 in compiler AND the shipped 207KB wasm) so the whole codec runs SCALAR in the browser — live enc 100-161ms at 416x320 where vc_res_bench measured 14ms (70fps) = a 7-11x SOFTWARE gap, not silicon. Emit v128 for SAD/SATD, motion search, transforms, quant, deblock, rgba<->yuv. 5 components: language builtins, IR, nx_wasm emitter, nx_wat_compiler encoder, nx_wasm_vm interpreter (so faithdiff can verify bit-exactness)
F61983477engineer-videovideoREADYvideo FPS: modern drift-safe parallel band encode. The K-way parallel encoder EXISTS (bandEncode, bandK=min(cores-1,4) workers, own wasm instance each) but was deliberately disabled in 829 because (a) it rides the LEGACY 4-mode codec so enabling it forfeits the -20pct rct8/rct9 gains, (b) it is drift-prone (sender re-decodes on its MAIN thread to advance prev), (c) the stated justification was that single-frame is fast enough at 39ms vs THE 15FPS CAPTURE CAP — grading against our own throttle, not the silicon. REAL ARC: move bands onto vv_enc_rct8_band/vv_dec_rct8_band (already in the wasm exports), kill the main-thread re-decode drift surface, gate on cores. Sequence AFTER F618
F82942029architectstoresDONEnx_store_compact: collapse seg-store planes to ONE segment (newest is already a full snapshot incl tombstones) - verify byte-identical load BEFORE retiring, move superseded segs to retired/ under plock; debt- 442 segs/42MB -> ~200KB, restores O(rows) loads. RK013 sev9
F830b53259architectstoresREADYseg-store ROOT fix + janitor: sts_seed still commits a FULL snapshot per write so planes re-grow linearly - add a periodic compaction beat (nx_store_compact check/apply on a cron) AND a reclaim pass for superseded .docs/.idx left on disk (~42MB debt- alone); investigate a real append/delta write path
F86763037hostinghostingDONESOVEREIGN gzip/DEFLATE COMPRESSOR (the #1 hosting SOTA gap EATEN): nx_gzip = from-scratch RFC1951 DEFLATE (greedy LZ77 hash-chain + fixed Huffman) + RFC1952 gzip framing + zlib-exact CRC32. PROVEN byte-identical vs STANDARD gunzip (text 51%, 54KB src 71% cut, 25KB binary 71%, empty/1byte OK). LIVE on NAS + registered callable. Retires Accept-Encoding: identity dead-end
F868210113refereedebtDONEnx_debt add is CONTENT-IDEMPOTENT: exact-desc match => DUPLICATE-SKIPPED naming the existing row, exit 0, so a retry after a transport flake is a NO-OP (DEC014 mechanized). Proven with neg-control; deduped the seq332/333/334 triplicate
F86953368hostinghostingREADYsovereign INFLATE (gzip decompressor) - follow-on to F867: needed for (a) a SELF-CONTAINED KAT gate (round-trip without the external gunzip oracle) and (b) the CLIENT to ACCEPT gzip web content (currently sends Accept-Encoding: identity to avoid it). Inflate is mechanical (follow the encoded symbols) = simpler than deflate; oracle gzip(data)->our_inflate->data
F87054377hostinghostingREADYgzip EDGE-WIRING (crown-jewel, careful) - follow-on to F867: sites.elf serves Content-Encoding: gzip via nx_gzip when the request carries Accept-Encoding: gzip (+ the client sends it + inflates, needs F869). Every served page 60-80% smaller on the wire = the actual hosting bandwidth win. Never-brick deploy, KAT the served bytes decompress correctly
F871200014librarian-READYsovereign-formats migration arc: 135 registry TSVs + root .reg/.tsv -> seg-store planes via nx_tsv_migrate; per-registry rung = migrate + repoint decoders + tombstone TSV + rowset-identical NON-EMPTY gate; L008 lawreg row is the burn-down meter
F872200014conductor-READYfallback-engagement LOG: every backup-method use (transport fallback, ssh-once, manual promote, cat-pipe, .prev rollback, cap _cap fallback) appends a fallback- plane row {primary, backup, why, evidence}; miners rank frequency; each recurring fallback auto-files a primary-hardening rung -- operator NASA/DARPA-rigor directive 2026-07-20
F873200014referee-READYrigor ruler ingestion: NASA Power-of-10 + JPL coding standard + DO-178C objectives + FMEA/FTA as lawreg rows + a compare-plane rigor domain graded like the nist/swebench ingests; adversary role gets a standing IV&V beat (re-run closed-rung gates from wiped state)
NB10B600014engineer-modelnsbuilderREADYtokenizer-aware trie mask for grammar-decode (char-level automaton -> token-level; recovers masked-trajectory sampling losses)
NB10C500014engineer-modelnsbuilderREADYrejection-reason-fed round-2 prompting (gramdec rejection classes become anti-echo hints for the retry ladder)
S3BENCH800014pm+censusagent-seatREADYagent-seat benchmark board vs Claude Code/Gemini/Kimi3/GLM5.2/OpenAI: measured axes = crash+billing durability, cost-per-closed-rung (meter rows live), autonomy pct, verification soundness; external stacks bench-only per law
RANKW1800014pm+architectprogramREADYutility/enablement weight term in the ranker organs (gamebench queues, frontier ranking, F208 feeder, triage) per the 07-20 big-rocks sharpening: PARTIAL-blocking-use outranks fresh GAP; enablement multipliers rank at zero coverage delta; by construction not recall
SEATCAD700014conductoragent-seatREADYseat-lane cadence pulse: schtasks fires nx_seat_drive lanes on schedule (zero-Claude EXECUTE arm; daily maker-bench regression sentinel first) with hub frames + meter rows accruing
S2C400014operator+conductoragent-seatREADYoperator-assisted physical WAN-block tooth: run the S2 seat lane with the internet unplugged (LAN intact) and paste the closure — completes the offline-degradation law measurement (DNS half already proven LAN-only)
CLMEM1700014librarian+searchrecallREADYprivate search index over buildroot/runtime/clmem (1805-file Claude-side memory corpus, fed 2026-07-20): seats retrieve doctrine/laws/patterns via nishi_search-class query — PRIVATE index ONLY, never the public /doc corpus (family-suite never-public law); note nx_fs deny-list blocks token/key FILENAMES by construction
CLMEM2600014conductorrecallREADYcontinuous corpus sync beat: delta-pack the laptop memory dir to buildroot/runtime/clmem on cadence (schtasks/clock, zero-Claude — the proven pack→upload→unpack rail with the M5 mint recipe) so the ecosystem's copy of the Claude-side investment never goes stale
TS01800014modelwrighttoken-shiftDONEsplit-brain-router-organ: deterministic route(task-signature)->LOCAL-MAKER|CONTEXT-TRIM|CLAUDE-ARCHITECT from clawf- catalog; the 80pct edit-build-fix loops go to nx_autofix_ensemble at 0 API tok, architecture stays claude; measured by nx_seat shift
TS02600014conductortoken-shiftREADYhistoric-actlog-backfill: chunk-upload the 1.9MB claude_history.actlog (112944 hist tool-calls / 51387 transition frames / 221 sessions) into the sovereign actlog so nx_actlog mine + the F218 cron include ALL history, not just new sessions; laptop launcher = harvest_history.ps1 (nishi-first byte-move)
LADDER1800014census+pmprogramREADYmechanize the six-stage ladder (unstructured→structured→meaningful→actionable→reproducible→SOTA) into the census/ecomat instruments: per-asset stage GRADED BY EVIDENCE TESTS (plane-exists / digest-emits / seat-consumes / gate+attest / external-compare-row), board emitted, next-rung per asset auto-named — doctrine in clmem feedback-unstructured-to-sota-ladder-primacy.md; NISHI-OWNED build per the primacy rule
F874200014engineer-READYnx_stale_check sweep verb + clock beat: fleet mtime sweep (live elf vs buildroot source) emitted as JSON w/ per-organ owner join, fed to nx_triage -- the stale-mirror class measured continuously instead of found-by-injury (clock elf Jul-19, argecho 81h; 25-candidate surface 2026-07-20)
TS03800014conductortoken-shiftREADYsplit-brain live-dispatch loop: a seat/cron that reads a task signature, calls nx_clawf_route, and on route=LOCAL-MAKER actually dispatches to nx_autofix_ensemble (0 API tok) instead of a Claude turn; on ESCALATE keeps Claude; measured by nx_seat shift (local rows RISE, out/done FALLS). The router organ (TS01) is DONE; this wires it into the autonomous loop
F875200014conductor-READYwarden detector: unescaped % in a cron.reg command field -- cron reads % as newline and silently truncates the command (broke 3 reliability heartbeats 8.5h undetected 2026-07-20). Flag any cron.reg data row with an unescaped % in its command
TS04600014conductortoken-shiftDONEmigrate nx_seat (st_* helpers) + TS03 dispatcher onto nx_sovjson_lib (sj_*) = eliminate the 2nd inline-helper dup in the token subsystem, one canonical base. Crown-jewel careful window (nx_seat=boot/ready/shift heavily used); gate stays GREEN + byte-identical behavior. OO consolidation started on nx_clawf_route (proven 8/8)
F876300014referee-READYgate-verdict OO consolidation (D001 at scale): migrate 2601 custom-verdict gates onto the nx_gate_verdict base class, batched per owner lane, each behavior-preserved (gate still emits N/M+GREEN); 15 proven exemplars exist. Ratchet: DRY-adoption-permille (2 now) RISES, never falls; a warden detector flags any NEW *_gate.nx that rolls its own verdict
CONSOL1800014consolidator+pmatlasREADYatlas-consolidation architect (nx_consolidate LIVE on NAS, gate 6/6, mutation-proven): NAS-authoritative run over buildroot dup census -> seed consolq- plane -> a seat resolves each of the 32 divergent shadow pairs to ONE canonical + gated behaviour-preserving retire (identical=byte-equal-safe, divergent=prove-live-first); eats seq383/seq207/seq281/seq284 as they resolve
CONSOL2700014architect+consolidatoratlasBLOCKED CONSOL1OO/base-class consolidation from the atlas CALL/IMPORT graph (nx_eco_graph): cluster organs by shared function-shape -> propose base-lib/base-class extraction (rule 15: pattern in 3+ -> shared/) with a behaviour-preserving equivalence gate per extraction; the deeper 'object-oriented without losing functionality' rung beyond dup-shadow collapse
CONSOL1B800014consolidatoratlasBLOCKED CONSOL1nx_consolidate v2 REACHABILITY/TAXONOMY classifier: ONE tree-walk builds a basename->importer-count map, annotates each divergent pair as ORPHAN(0 importers=safe auto-retire) | STALE-SHADOW(same-purpose, retire non-canonical) | NAME-COLLISION(different organs, rename+repoint N consumers); blast-radius = importer count so a seat resolves lowest-risk first. Makes the 32 divergent hazards safely actionable
CONSOLRUN700014consolidator+seatatlasBLOCKED CONSOL1Breal gated consolidation runs (nx_consol_apply LIVE on NAS, gate 5/5 never-brick-proven): a seat points it at consolq- proposals lowest-blast-radius-first (the 9 zero-importer organs), buildelf=a nx_buildonly wrapper (nx_sov_build_run --build-only, stable artifact), artifact=/tmp/.sov.elf; CONSOLIDATED=real debt eaten, REFUSED=learns canonical, both never-brick. Eats seq383 pair-by-pair
TS05500014refereetoken-shiftREADYtree-wide JSON-emit dedup onto nx_sovjson_lib (D001-family growth): every organ reimplements cat/catn/vlen/cat_esc/le/col/lit_eq INLINE; migrate onto the proven shared base via DELEGATION (byte-identical, gate-per-organ, fail-loud edits). Token subsystem DONE (seat/clawf_route/actlog, base Ca=3); this is the tree-wide continuation = the base becomes the ecosystem standard
F877300014architect-READYnx_estr adoption + helperdup ratchet: migrate organs off local puts/putn/cat onto the nx_estr shared base (built+proven 2026-07-21); add a warden `helperdup` detector (L010) counting organs that reimplement the primitives instead of importing nx_estr, so line-dup adoption is a measured ratchet (mirrors L009 gatedry + nx_gate_verdict)
SHARP1800014architect+censusatlasREADYTHREE-PARTY small-sharp-tool ratifier (out-linux-linux grade): compose the MATH party (nx_sharpgrade LIVE, gate 6/6 — call-graph components + vocab cohesion, SHARP|SHARP-LIB|SPLIT-K) with the ECOSYSTEM party (a proposed split keeps all affected gates GREEN + nx_consol_apply byte-equiv) + the CLAUDE/judge party (component names a coherent responsibility); a decomposition ratifies ONLY 3/3, disagreement = filed question. Grades every organ, emits the split/merge backlog
SHARP2700014librarian+censusatlasDONEfunction-mining archive (rule-13 additive): nx_sharpgrade already emits the FN inventory (name|component); productize into a function registry store (name, organ, signature, body-hash, call-edges, lineage) so functions are ARCHIVED never deleted + MINEABLE — find duplicate BODIES across organs (true function-level dedup), mis-placed functions (belong to another component), and the natural clusters that ARE the small-sharp-tool boundaries; feeds SHARP1 the cross-organ merge candidates
F878200014architect-READYwire nx_presubmit into nx_ship as a pre-build gate (refuse a build when name is SHADOW-RISK/DUAL, exit before spending the compile) + extend it with the cron-unescaped-% check (F875 class) and a call-arity prepass -- one pre-submit correctness gate for every token-wasting class; MCP stub lands on next reconnect
SHARP3800014consolidator+architectatlasBLOCKED SHARP1dup-function EXTRACTOR (consumes nx_funcmine's 1634 groups): for a dup group, place the body ONCE in a shared base/lib, repoint each consumer to import it, gated by nx_consol_apply byte-equiv rebuild + three-party ratify (SHARP1); lowest-copy-count-first EXCEPT the gw-family-x427 which folds into the LIVE D001/F876 nx_gate_verdict base-class-adoption arc (do NOT double-drive — coordinate). Turns the measured 9495-dup debt (seq389) into extractions, behaviour-preserving by construction
RIGOR1800014referee+censusrigorREADYDYNAMIC gate-rigor harness (the STRONG DARPA/DO-178C proof beyond nx_gaterigor's static 525permil lower-bound): for each gate, MUTATE its subject organ (flip a byte / break a property) and CONFIRM the gate goes RED; a gate that stays GREEN under mutation is PROVABLY vacuous (junk). Triages the 1272 flagged: real-vacuous get a neg-control, false-positives widen the static markers. Ties F873 DO178C IV&V beat + nx_evidence_audit liar-catcher — compose don't duplicate
F879300014architect-READYmigrate the 10 _board organs onto the nx_board base (built+proven 2026-07-21; composes nx_estr) -- HTML scaffold + JSON envelope + atomic publish were reimplemented in each; batch per owner, board-output byte-identical; add a warden boarddup detector so board-base adoption is a measured ratchet (mirrors L009/L010)
CONSOLMC700014consolidatoratlasDONEmulti-consumer consolidation applier (extends nx_consol_apply beyond zero-importer): to retire a shadow of a LIB that N organs import, rebuild EVERY importer + byte-compare each artifact before/after — CONSOLIDATED only if ALL importers unchanged, else REFUSE+restore. Uses the nx_funcmine/histogram importer list; unlocks the ~28 remaining divergent shadows + the gw-family via F876. never-brick by construction
F880300014census-DONEwarden ratchet cap-window fix: LW_FILECAP=600 sees only ~10pct of the 6204-file tree, so L001/L006/L009/L010 sub-window migrations do not register and most offenders stay invisible -- raise the cap / recursive-walk / or sample-extrapolate so the ratchets reflect real burn-down; proven exemplar nx_modtest->nx_estr byte-identical does not move the metric today
F881700014architect-DONEENFORCE the guards at pre/at/post submit so non-compliant organs CANNOT be built/shipped (I bypass them today): wire L001 magic-number + L008 tsv + L011 self-ceiling checks into nx_presubmit src-lint mode AND into /api/build as a REFUSAL for NEW organs (existing backlog grandfathered) -- the ecosystem catches the violation at submission, not the operator re-sweeping after. sev high: this is the foundational check-and-balance
F882600014architect-READYF881 REMAINING (magic-number ratchet is LIVE+proven at /api/build; tsv+self-ceiling NOT yet enforced): extend the /api/build submission gate to also REFUSE a build that ADDS a .tsv write (L008) or a self-ceiling scanner (L011), same ratchet pattern (delegate to a canonical per-file counter, per-organ baseline, fail-open). Then the full pre/at/post enforcement the operator demanded is real, not just magic numbers
F883400014architect-READYrefusal-feedback ENFORCEMENT: a warden/lint check that flags any organ emitting a REFUSE/400/verdict without a why+fix field, so refusals-without-feedback become a mechanically-caught defect (operator: refusal without feedback is not acceptable)
FEEDBACK1700014referee+pmrigorREADYretrofit the verdict+reason+fix refusal contract across ALL organs/tools/APIs (operator law: no refusal without clear feedback + fix recommendation, no weird errors/blanks): sweep every refuse/abort/deny path — mgmt API 400s, cap -32001 denials, gate REDs, router/route refusals, applier refusals — to emit verdict + plain-language reason + actionable fix, each with a gate tooth ASSERTING reason+fix present (the nx_consol_apply_mc_gate T2 exemplar). A bare 400 or -32001 = a violation. LIVE exemplars: nx_consol_apply + nx_consol_apply_mc (both promoted, gate-enforced)
F884300014census-READYmechanize the magic-number sweep to zero: an nx_magic sweep-mode (or beat) that applies+byte-identical-verifies across ALL _hdl_build + buildroot/runtime organs, with a CONSUMER-BUILD verify path for libs (whose standalone build is empty) so they are not skipped, per-organ backup+revert-on-diff; ranges nx_e-nx_z + L006 corpus + the 47 skipped libs remain (26093 of the original 27266). PROVEN method this session (1173 eliminated byte-identical); this scales it to zero
OOEXTRACT800014consolidator+architectatlasREADYBYTE-PROVEN OO extractor/renamer (UNBLOCKED by the 2026-07-21 experiment: rename-lib+repoint-consumer is BYTE-IDENTICAL, md5 1393093682... identical before/after). Build a tool that (a) extracts a duplicated function to a shared base + repoints consumers, or (b) resolves a split-brain by renaming divergent copies to distinct names + repointing each consumer to its version — each step gated by a PER-CONSUMER byte-equiv rebuild (re-run the proof, never assume it scales). Eats funcmine's 9506 dups + the 26 live-divergent shadows; and makes D001's 2553-gate nx_gate_verdict adoption BYTE-PROVABLE (coordinate F876, don't double-drive). This is 'OO without losing functionality' made byte-provable, as sound as dedup
F-CC-NXINT800014engineer-compiler-READYnx_cc nx_int struct-field size-0 offset miscompile: root-fix layout (resolve type-alias size BEFORE field offsets) + sweep 337 exposed runtime/*.nx (2292 decls); proven minimal repro + workaround (i64 fields) already landed on the objective_beauty measuring suite 6/6 green; ties debt seq521
F885721311engineer-ososREADYNative Contract v1 gap witnesses: C07 supervision-trees, C08 async-batch ring, C10 kernel deterministic-mode, C11 NXE version solver, C12 event journal, C13 time/string ABI (card knowledge/registry/native_contract_v1.txt)
F886611212autoautoREADYseat-reap steady-state beat: schedule nx_seat reap on clock (clockjobs- PLANE only never TSV-edit; reap is idempotent+audited so interval-safe; exclude-reaped-from-boot-active nicety optional)
F887810113teamautoREADYdebt 1784480242: SYSTEMIC dual-copy hazard found + tooled 2026-07-19: nx_dup_source_check (DEPLOYED + self-test proven + live on NAS) scanned the build tree
F888810113teamautoREADYdebt 1784576686: LAW L001 BREACH (nx_law_warden, mechanically detected): 2348 inline numeric literals >=1024 on non-const non-comment lines across 600 scanne
F889810113teamautoREADYdebt 1784577890: LANDMINE REGISTRY FROZEN 36 DAYS -- and it has REGRESSED to the exact condition it was built to end (operator question 2026-07-20: 'are land
F890810113teamautoREADYdebt 1784578442: LAW L006 BREACH (nx_law_warden AUTOFILED by the hourly beat, zero-Claude): inline magic numbers in the CORE runtime tree - the _hdl_build co
F891810113teamautoREADYdebt 1784583453: MEASURED 2026-07-20: the PUBLIC edge (sites.elf :8443) is destabilised by oversized proxied POST bodies -- not merely capped. Evidence: push
F892810113teamautoREADYdebt 1784583610: CWE-798 LIVE HIGH-PRIVILEGE CAP EXPOSURE IN THE MEMORY CORPUS - a NEW class, distinct from seq317. seq317 turned out to be a FORGED negative
F893810113teamautoREADYdebt 1784608831: D001 MEASURED 2026-07-20 (atlas-scale gate-verdict DRY surface, ws=arch-consolidation): 5365 gate organs, only 15 INHERIT the nx_gate_verdic
F894810113teamautoREADYdebt 1784643013: ER-8 KEYSTONE: the gen engine serves only CFG-distilled Z-Image Turbo (cfg 1.0, no negative prompt) so it cannot control clothing/exposure/p
F895810113teamautoREADYdebt 1784644808: LAW L011 BREACH (nx_law_warden AUTOFILED by the hourly beat, zero-Claude): corpus scanners that walk a directory but HIDE their coverage (no
F896810113teamautoREADYdebt 1784661149: OPERATOR 2026-07-21 (CRITICAL SECURITY): external bug/fix ingestion for answer-key training is an UNTRUSTED-INPUT / DATA-POISONING / TROJAN-
F897810113teamautoREADYdebt 1784661346: OPERATOR 2026-07-21: protect EVERYTHING (not just LangIntel bug-ingest) against poisoning/trojans at ALL THREE stages BEFORE/DURING/AFTER su
F898810113teamautoREADYdebt 1784663080: X3 ASR long pole: real-time RU+EN speech->text has no acoustic model/decoder/LM (nx_asr_frontend.nx is only an N=32 toy FFT frontend, no mel
F899810113teamautoREADYdebt 1784664001: fix-diff/patch answer-key ingestion HARD-BLOCKED until 3-stage trojan fence LIVE: trust-tier PRE (merged-into-canonical-default-branch by ma
F900810113teamautoREADYdebt 1784668845: nx_cc STRUCT-FIELD MISCOMPILE (proven 2026-07-21, minimal repro runtime/_hdl_build/nx_segfield_repro.nx): struct fields declared with the nx
F901810113teamautoREADYdebt 1784670217: URGENT token hygiene: operator GitHub PAT ghp_BVId... is OVER-PRIVILEGED (X-OAuth-Scopes runtime header = repo,delete_repo,admin:repo_hook,w
F902070772mined 669/223 baselineBLOCKED Tgov
F903080853biggest honesty leverBLOCKED Tauto
F904070772operator law 07-21BLOCKED Tgov
F905070772DARPA rigor 07-21BLOCKED Tgov
F906060682bench-oracle fenceBLOCKED Tgov
F907070773lowest live matrixBLOCKED Treliability
F908060682feeds compare-regenBLOCKED Tresearch
F909610113conductorautoREADYbfeed beat wiring: zero-Claude cron/clockjobs row running nx_backlog_feeder run 8 5 team auto + heartbeat + cron_watch (clock plane contended; wire via clockjobs- plane only)
F910060682directive mechanizationBLOCKED Tgov
F911070773perceptual UNSOLVEDBLOCKED Texperiential
F912610113pmautoREADYbfeed source expansion: gap queues (gamebench gap_queue) + atlas maturity-gap x momentum + intake-TRIAGED promotion as additional nx_backlog_feeder evidence sources, same marker/dedup contract
F913810112pmcompanionREADYcompanion program card: Elder-AI to sovereign migration (SSOT=elara_card.json+companion_digest+companion_arch; parity checklist vs elder-ai-platform capabilities)
F914920212engineercompanionREADYER-8 controlled-model path: non-distilled checkpoint (cyberrealisticZImage_v40 / Realism_Engine_Klein on disk) at CFG>1 + negative prompt, model-select-by-intent in GenWorker = controllability keystone (clothing/exposure control)
F915830311modelwrightcompanionREADYER-5 identity LoRA: train Elara face LoRA from canonical refs (svc-lora port) + InstantID/IP-Adapter eval = lossless face clone for shot-to-shot identity
F91683039engineercompanionREADYER-1 prompt-grammar port: content.db outfits905/poses1564/locations361/arcs -> sovereign seg-store planes + nx_photo_grammar compiler (Director Shot-Sheet order, hub-spoke locations, weighted select)
F917720210engineercompanionDONEER-2 immersion chat: card-driven persona from elara_card.json + sampler params (min-p/DRY) + CJK logit-ban language lock + summarization memory + lorebook
F918720212engineercompanionREADYbeauty + anti-body-horror capstone: composite objective_beauty organ (png decode -> seg -> landmarks -> beauty/symmetry/aesthetics, suite 6/6 green) wired into score-and-rank + /api/batch best-first + anatomy_risk pre-render QC
F919620212engineercompanionREADYmedia-manager port: record/upload + download-manga into the gallery; verify gens/all-media tabs still live; hostctl-supervise the gen daemons (retire cron-only keeper)
F920730311engineercompanionREADYcompanion aliveness organs: proactive day-arcs (she reaches out on a beat) + dream-cycle (elara_mem.log -> long-term beliefs) + nx_companion_{rel,turn,render,voice,proactive} serving path behind /gen
F921946103modelwrightcompanionBLOCKED F792sovereign engine from the first byte: retire elder-sdcpp/sd-server - nishi-served image gen + companion LLM serving (rides split-model-serve F792 + nofloat lane)
F922730310engineer-graphicscompanionREADYavatar/VR export: genome -> 3D avatar (glTF/VRM) + online-store asset pipeline; composes graphics/game lane (nx_render3d, rasterizer, gamebench) for VR, stores, robotics, DnD, isekai
F923830311architectcompanionREADYphenotype-genome compiler + proc-gen: deterministic genome->prompt compiler (per-attribute edit = predictable change); sample genome space -> distinct adult cast (DnD party/isekai); digital-twin bridge to biotwin lane (rule-26 body fence)
F924720212researchercompanionREADYresearcher mine (nishi organs): Z-Image counterbalance registry from historic logs + 226k-gen learnings + render_bug_register 15063 rows -> companion_arch.md structured rows + companion_digest (self-structuring; naive edits never re-break the genome)
F925620211refereecompanionREADYcompanion SOTA ruler: capability checklist vs Replika/SillyTavern/character-ai-class + civitai fidelity bench -> measured gap_queue drives this lane (gamebench pattern; measure before claiming SOTA)
F926630311engineercompanionREADYcompanion voice + TTS lane: sovereign voice for Elara (gen-aud/Chatterbox parity path, nx_companion_voice) so she talks, not just texts; VR-ready
F927060682dAX08 distribution FIELD-DECISIVE (OPA bundles/Gatekeeper CRD sync/Config packs): each guard declares directive-ids it enforces; reconciler flags unbound directives + dead guards; dirregcmp- planeBLOCKED Tgov
F9280404101dAX09 authoring FIELD-led (git PR+CI / OTel weaver lint): putn mutations schema-linted + review-gated, kills hand-seed drift (seq550 class); dirregcmp- planeBLOCKED Tgov
F929050592dAX10 audit-trail FIELD-led (OPA decision logs / CloudTrail): every guard refusal/pass logs the directive-id it enforced; dirregcmp- planeBLOCKED Tgov
F930050591dAX11 exception-waiver FIELD-DECISIVE (InSpec waivers/Gatekeeper exemptions): absence PROVEN by dr_run negcontrol 0 permil vs plane bytes 07-21; sweeps become waiver-aware; dirregcmp- planeBLOCKED Tgov
F9310303111dAX12 lifecycle FIELD-led (ADR status lifecycle / OTel stability levels): state machine + close discipline for retired directives; dirregcmp- planeBLOCKED Tgov
F932720212engineershipcompleteREADYAT-submit ship-complete refusal: /api/tools/register refuses elf .new paths + schema-row-missing (extend ma_tools_register fail-closed; crown-jewel, quiet window ONLY, forbidden mid-swarm)
F933620212architectshipcompleteREADYPRE-submit integration-completeness: extend nx_presubmit with shipcheck-lite target checks (schema row present if registered, no stagedref for target, atlas advisory) -- fold beside F878b cron-pct+arity
F93482468engineer-videovideoREADYF618 FLIP LANDED 2026-07-21: SIMD SAD lives in the SHARED toolchain (nx_wasm intercept + nx_wat_compiler 0xFD atoms + nx_wasm_vm interp + vm_sad/vm_sad_capped T==16 seam via v128_sad16). PROVEN: nx_v128_gate 6/6, faithdiff 10/10, layout gate 6/6, bgop par=0 all qp, V8 15.60x @416x320 / 14.95x @640x480 checksum-locked. NEXT RUNG: rebuild nx_video_client.wasm through the flipped toolchain, stage web_assets, vcv gate + vqoe>=30 + BD re-sweep with EO armed (debt seq563), opt-in probe, ship
F935610113conductorshipcompleteREADYwire nx_shipcheck beat into the hourly clock: verify clockjobs cmd-with-args support or add nx_shipcheck.cron.sh wrapper (F872 recipe), put-writer row shipcheck 3600, then verify an UNATTENDED fire in knowledge/status/shipcheck.log before claiming zero-Claude
F93652686pmshipcompleteBLOCKED F209frontier D-row to ws join UNMEASURABLE: F-rows carry no closing-ws column so shipcheck frowjoin=-1; add closing-ws or evidence-ref on close (F209 family) so unboarded completions become a measured class
F93773477engineer-compilervideoREADYF618 SIMD ladder rung 2: vectorize SATD (vc_satd4_d/vc_satd16_cand) + subpel sp_sad + transforms/quant via the SAME fused-intercept pattern (one bit-exact scalar reference fn per kernel, intercepted by name in nx_wasm.nx, per-kernel teeth added to nx_v128_gate). Kernel order by measured hotness per the F618 design
F938540410engineer-compilervideoREADYF618 register-resident v128 SSA + parser __v128_* intrinsics (original design components 1-2, deliberately DEFERRED): only worth it when a kernel class needs cross-op fusion beyond the fused-call intercept. Touches nx_parse/nx_types = ecosystem blast radius, fresh-session-only per the F618 design warning
F939720212engineercompanionREADYsovereign-formats migration for the gen stack: cids.tsv sidecar + elara_*.txt state files + mem log + anatomy_audit.tsv -> seg-store planes (genscores- pattern; sovereign-formats-first law; card/persona/state as planes with history)
F940830311refereecompanionREADYrealism ABSOLUTE ruler: detector-graded realism gate - forensic axes (spectral slope, skin micro-texture, hands, crowd/background coherence) + external AI-detector batteries used as quality RULERS grading every gen toward indistinguishable; store-photo bar; benchmark corpus = old-system best renders + Replika-class SFW/NSFW competitor set (researcher-fed via F924/F925)
F94162468engineer-ososREADYpabi native mem+clock backends (nx_mmu_hw + CLINT wired as pabi seams) byte-exact increment
F942830311researchercompanionREADYsovereign PROMPT-SCIENCE engine: multivariate word-impact measurement in 4 dims (presence, RoPE position, dose/weight, combination+arc timing) over the 226k-gen corpus + render_bug_register 15063 + combination_scores (migrate_v83) + util-inject seed/word/phrase A/B -> measured coefficients feeding the counterbalance registry; controlled A/B via seed-held img-diff + beauty/anatomy gates as judges; generalizes to other RoPE systems (graphics/video/LLM)
F94372686engineer-ososREADYlinux guest lane depth: arbitrary-ELF guest breadth beyond /bin/sh + guest stdin/pipes + guest supervision-resource caps (C07 tie) + RV64 guest lane
F944720212engineercompanionREADYnx_model_get sovereign model/LoRA acquisition: civitai + civitai.red/archive mirrors + HF -> download to NAS with SHA256 verify (recipe proven 223MB LoRA exact-hash) + model registry PLANE + torrent-seed/archive the zoo; Authorization header not url-token (rotate the leaked key); feeds ER-8 checkpoints + ER-5 LoRA training
F945730310engineercompanionREADYcompanion GAMES port: strip poker (5-card draw + holdem, old svc-poker parity) + truth-or-dare on the /gen surface - game state planes, body-zone wardrobe/occlusion integration (svc-wardrobe zones/layers -> CompanionWardrobe), reaction lines from her persona/state, per-round photo via the gen path; adult + non-adult modes; needs ER-8 for reliable staged exposure
F946640410engineer-videocompanionREADYcompanion sends VIDEO CLIPS: short motion clips of Elara in chat (img2vid/animate path on the engine or video-lane codec pipeline) + clip ingest to gallery + inline playback in elara_ui; composes the video lane (vcv codecs, F618 SIMD) toward VR/metahuman motion
F947830310refereecompanionREADYLOSSLESS ROUND-TRIP RULER (the vecteezy test): reference image -> spec -> regen -> MEASURED similarity per channel (semantic checklist, pose/structure IoU, print/texture, color histogram, identity) ladder L0 prompt-only (baseline PROVEN 2026-07-21: semantics ~85pct, geometry ~50pct, identity 0) L1 +genome/scene structured spec L2 +pose skeleton ControlNet L3 +img2img structure lock L4 +identity LoRA/IP-Adapter = near-lossless; medical/biotwin use rides rule-26 fence
F94891675engineer-ososDONEF103d-1 never-brick FENCE gate for UEFI-on-metal: mechanically prove the autonomous path is FILE+EMULATOR-only (zero persistent firmware/UEFI/CMOS/NVRAM/vBIOS write), metal-flash is operator-gated fail-safe removable-media-only; gate RED-flags any firmware-write path
F94982795engineer-ososDONEF103d-2 UEFI x86-64 PE/COFF EFI app: the NishiOS stem as BOOTX64.EFI (UEFI entry, BootServices console, clean ExitBootServices handoff) — the stem is today an rv64 flat binary for the sovereign emu, so this is the x86-64 EFI-image gap
F95082795engineer-ososDONEF103d-3 ESP image BUILDER (sovereign organ): emits a bootable image FILE = protective-MBR + GPT + FAT32 EFI System Partition + EFI/BOOT/BOOTX64.EFI; structural gate on GPT header CRC32 + FAT32 BPB + payload-at-path; produces a FILE only, zero device write (fence F948)
F95182795engineer-ososDONEF103d-4 EMULATOR boot-proof: boot the ESP image under OVMF+QEMU (external oracle) and/or the sovereign x86 emu, assert the stem serial marker + clean halt; the never-brick boot verification, NEVER on metal (fence F948)
F95271786operatorosREADYF103d-5 operator-gated METAL flash (THE guarded operator move, rule-26): write the emulator-verified image to REMOVABLE USB only, explicit operator confirm + checksum-verify, fail-safe/reversible, NEVER an internal-firmware/UEFI/NVRAM write; boots via UEFI removable-media path
TS06810113claude-fabletoken-shiftDONEctx-tax flush function: nx_seat brief CHECKPOINT verb + boot last_brief + warden lossless-flush protocol
F953720212researchercompanionREADYL0 RESEARCH PROGRAM: deep-read the 2026-07 harvest (l0_research_2026-07.md: FreeFlux RoPE-layer roles, RF-inversion family, FRAP dose-weighting, structured-captions, EDITOR prompt-inversion, RefVNLI/ProductConsistency identity eval, Complex-Edit/ByteMorph rulers) fetch-to-disk + replicate the load-bearing findings on OUR stack + benchmark where jul-2026 research stands vs our system
F954830311architectcompanionREADYCONVERGENCE ENGINE (InfiniGen-for-images): closed-loop reference->spec-invert->gen->per-channel-score->spec-adjust->iterate to 1:1; every iteration logs (word,position,dose,model,delta) into coefficient planes = system LEARNS first-shot fidelity; cross-model blast-radius tables make ER-8 select-by-intent data-driven; surgical per-attribute edit via spec-delta + RF-inversion structure lock graded everything-else-unchanged; design in l0_research_2026-07.md
CU01800014modelwrightcapupliftREADYpromptpack organ: compose capuplift- PT rows into per-model prompt scaffolds (claude/kimi/glm/local makers)
CU02900014modelwrightcapupliftDONEapply 5elem+XML+prefill+order-of-thinking to local maker prompts (per-maker) + re-measure n=14 ensemble target >85pct
CU03800014modelwrightcapupliftREADYOPD pilot: step-wise on-policy distillation teacher=frontier-LLM student=local gguf makers; ruler=nx_swebench_status
CU04700014modelwrightcapupliftREADYmaker refresh recon: Qwen3.6-27B-MTP + Devstral-24B Q4 gguf under symjudge (NB7 family, 16GB fits)
CU05600014conductorcapupliftBLOCKED CU01partner surface: serve promptpack scaffolds via MCP prompts/get so external harnesses (kimi/glm/claude) consume them natively
CU06500014librariancapupliftREADYfull Prompting101 transcript ingest (yt ysPbXH0LpIE) into corpus via a fetchable path; WebFetch cannot pull yt transcripts
SV01800014refereesovereigntyREADYsovledger rollup verdict: flag any external dep lacking named path+ruler as RED; wire into ecosystem_maturity_rollup
SV02700014architectsovereigntyREADYbase-vs-boundary audit of nishi browser + nishi os: MEASURE where 3rd-party formats sit in the BASE (historic burden) vs boundary (interop); file an exit rung per hit
F9550000149claude-mvm-0722READYMV-1 Bulk vault migration: nx_mvault_walk record+CID-dedup the real corpus (76004 recordings+books+movies then gens/images) into /vault seg-store; additive+resumable
F9560000149claude-mvm-0722READYMV-2 NXVI seek-index BACKFILL at scale: drive nx_galx_idx_backfill.elf over all 76004 recordings so keyframe seek is ready on FIRST open (galxidx now 1260/76004=1.7pct, lazy-built); proof=galxidx count->~76004, cold /segs ready:1
F957800014conductortoken-shiftDONETS06 deterministic delegation-editor verb on nx_dedup_migrate (known sj_-helper table -> server-side rewrite, no fs_write so DQ001 token-path deny is dodged); compose stage+edit+verify+mark = zero-Claude dedupq- drain via orchestrate/cron
F958800014modelwrighttoken-shiftDONETS07 grow the nx_dedup_migrate delegation table: add _put/_putn/_w/_wn helper shapes (put==cat, putn/wn==catn semantics) with per-shape guards -- putn-class REQUIRES the negative-branch marker (45-as-u8 write) since e.g. nx_cleanview ncv_putn lacks m<0 handling = divergent for negatives (measured); unlocks nx_token_counter/nx_cleanview/nx_dup_source_check class = the bulk of the 245-organ D001 census; walker+judge pipeline unchanged
F959700014conductortoken-shiftREADYTS08 zero-Claude standing drain: clockjobs/cron beat firing plan-ddq (nx_dedup_migrate queue 2) on cadence + F208 feeder files new dedupq rows from the 245-census -- the loop then eats D001 organ-dedup continuously with no session; clockjobs- plane contended so owner wires the row (never hand-TSV)
F960800014conductortoken-shiftDONETS09 wire the hardware-aware route into the maker pipeline: nx_autofix_ensemble/dispatcher + TS03 walker consume nx_hw_envelope route (maker+bench columns) instead of hardcoded model paths -- NishiOS runs the same loop phone->super by swapping hwtier- rows only
F961800014modelwrighttoken-shiftREADYTS10 bench-ruler run: (a) rule-2 load-smoke bench_qwen25_coder7b_q4km.gguf (4.68GB staged nx_stage) under nsv dense-i8 -- 7B dims may exceed 1.5B-tuned buffers, verify dont assume; (b) per-maker prompt adaptation (NB7 lesson); (c) official n=14 head-to-head: 7B-as-bench vs sovereign ensemble 85% under the SAME symjudge -- the external model is the RULER never the shipped maker (bench-only law); report both numbers
F962800014modelwrightnsbuilderREADYENGINE GROWTH (measured F961 finding): nsv i8-weight-cache build FAILS for 7B under the 15GB WSL VM (exit 1 mid-cache; 7B i8 cache ~7GB + i32 head ~2GB + 4.7GB file map > envelope) -- build QUANT-NATIVE inference (matmul directly on Q4_K blocks, no i8 explosion) = ~3x memory reduction, fits 7B-class makers/benches on edge-tier hosts; alternatively .wslconfig memory raise or NAS-hosted bench (35GB routes laptop-tier) are operator options
F963800014modelwrightnsbuilderREADYENGINE GROWTH 2 (measured F961): nsv garbles Qwen2.5-3B (token salad, smoke 0/2, judge held 0-UNSOUND revert-exact) though the file is STOCK format identical-family to the working 1.5B (metadata compared) -- the breaker is dim/config derivation (3B=2048-embed/36-layer/16-head vs 1.5B=1536/28/12): derive head_count/head_count_kv/block_count/embedding_length/ffn fully from gguf KV and size ALL buffers from them; tooth = 3B smoke emits compilable output; then the n=14 oracle sweep unblocks. Composes F962 quant-native for 7B-class
F964800014librarianminer-sotaDONEminer MS01 transcript-harvest organ: sovereign ingest laptop Claude transcripts->actlog frames + cron (kills MG03, subsumes TS02 backfill)
F965700014pmminer-sotaDONEminer MS02 outside-gauge: outside_permil per window from actlog + page, watch-it-fall vs MG00 baseline 653
F966700014researcherminer-sotaDONEminer MS03 sequential-pattern upgrade: n-gram -> closed sequential patterns (PrefixSpan-class), bench vs SPMF-class oracle per external-reimpls-bench-only law
F967700014researcherminer-sotaREADYminer MS04 process discovery+conformance: directly-follows graph -> inductive-miner-class model, conformance-check new sessions, PM4Py oracle judge-equivalence
F968800014refereeminer-sotaDONEminer MS05 eval harness: synthetic seeded logs w/ known-pattern oracle + neg-controls + insufficient-samples honesty (nx_triage eval precedent) = miner accuracy MEASURED
F969800014conductorminer-sotaDONEminer MS06 class-coverage: mine features(frontier-) bugs(debt-) capabilities(bigrock-) harnesses(gates) workflows(plan-) agents(seats) - per-class verbs + coverage gauge
F970600014researcherminer-sotaREADYminer MS07 research ingest: nx_dr_run on sequence/process-mining SOTA, drbench-graded, feeds MS03/MS04 algorithm choices
F971700014conductorminer-sotaDONEminer MS08 auto-propose: outside-residual class above threshold auto-files intake REQ (observed->mined->proposed, feeder/idempotent-marker pattern)
F972600014architectminer-sotaREADYLAP01 laptop host-runner design: sovereign cap-gated runner organ for laptop ops (schtasks/WSL/local builds) so MG04/MG07/MG10 become API calls; rule-26 fence from birth
F973800014librarianminer-sotaREADYminer MS01b PORTABLE harvester organ: NishiLang transcript-parser (nx_actlog harvest verb or nx_claude_harvest), NAS-built via /api/build, deployable to ANY host holding transcripts (gen-image engine pattern: compute-where-data-lives, API-orchestrated, per-host runner); PS bootstrap retires when live
F974800014modelwrighttoken-shiftREADYF961-completion: finish the 3B-oracle n=14 sweep -- 7 cells MEASURED (candidate2/3/4+max2_rf+absdiff_cf ALL GREEN FIRST-TRY symj-verified; sum100 MISS shared-with-everyone), 8 cells UNMEASURED after 4 host kills (WSL VM idle-death + load-correlated TERMs, crash-cluster adjacent -- run in a stable window or post-BIOS-flash); THE DECISIVE CELL = mut_absdiff_os (holdout; GREEN would prove 3-way union 13/14); use bench_official3b_q4km.gguf + af_bench3b_rem2 manifest + attached client per instance; then cascade union af_official_rerun+af_c15_adapt+af_bench3b_official*
F97591349companioncompanionBLOCKED F916companion STATE-COHERENCE ENGINE: one internal state store (wardrobe/arousal/mood/energy/location/arc/relational-dynamic) that chat persona + gen prompt composer + wardrobe logic ALL read-write per turn -- she dresses herself and STAYS dressed by construction; her choices follow her internal thought process (operator charter 2026-07-23). Composes F916 grammar planes + F917 card + F920 arcs; state plane = seg-store; deps F916,F917
F97681015finfinDONErename financial nx_debt twin to nx_debt_finance BEFORE any nx_debt rebuild eats seq240 seq121 brick risk
F977820211censusfinDONEfinance and legal compare domains q axes matrix vs TigerBeetle Stripe QuickBooks Bloomberg Clio Harvey NetDocuments - the missing SOTA yardstick
F978712311censusfinDONEadd finance as an ecomat maturity domain via nx_ecomat_put autoid adder - finance is absent from all 22 graded domains
F979710113refereeofficeREADYre-seed stale office-legal ecomat row to current 9-capability state and earn triangulation eats debt seq170
F980710113engineerofficeREADYnx_toolreg_reconcile to surface nx_docgen nx_billing nx_matter_intake as MCP discovery stubs - allowlist GREEN but tools-list under-reports
F98193145finfinDONEnx_ledger sovereign double-entry primitive accounts+transfers one invariant debits==credits integer minor-units immutable CID plane two-phase+linked chains gate proves money cannot be created or destroyed
F98272466finfinDONEnx_coa chart of accounts + period close and lock data-driven rule-11 no hardcoded account tree
F98372468finfinDONEnx_recon general three-way reconciliation primitive lifted from the proven nx_trust IOLTA implementation
F98462468finfinDONEnx_fx exact multi-currency no-float conversion with explicit rate provenance and rounding-remainder accounting
F98582686finfinDONEnx_stmt trial balance to P&L balance-sheet cash-flow DERIVED from the ledger never typed - a statement that cannot disagree with the books
F98683038researcherfinDONEnx_edgar SEC filings ingest data.sec.gov cursor-follow CID-stored - the R-MKT6 real-data rung banked 06-24
F98773368researcherfinDONEnx_xbrl XBRL and iXBRL fact parse over EDGAR filings feeding real per-company numbers into the valuation engines
F988810113finfinREADYexpose existing pure finance engines nx_fin_zscore Altman and nx_fin_mgmt_grade stewardship via register-mint-call - built and gated in June but ZERO nx_fin rows in the live allowlist
F98983477finfinDONEnx_iso20022 pain pacs camt message parse and emit - 2026 is the inflection Fedwire migrated Jul-2025 and Swift ends MT coexistence Nov-2026
F99063035researcherfinREADYnx_mktdata market data ingest plus OHLC bars on the seg-store with staleness verdicts and source provenance
F99163365finfinDONEnx_portfolio positions exposure and VaR integer-exact composing nx_ledger and nx_mktdata Kelly sizing already researched
F99253695finfinREADYnx_fix FIX order protocol PAPER-ONLY by construction - no live venue connection without explicit operator authorization
F99393038counselofficeDONEnx_cite_abstain FLAGSHIP grounded extraction that CITES or ABSTAINS fail-closed on the CID plane - 2026 benchmarks show 24pct of graded legal answers cite law that does not support the claim and every model fabricated at least one citation
F994830311counselofficeDONEnx_ediscovery EDRM legal-hold collection processing review production - composes the proven nx_box privilege log and Bates numbering
F99573368counselofficeDONEnx_citator citation graph and treatment signals Shepards KeyCite analog over the court corpus ties civic A-lane RECAP ingest
F99673368counselofficeDONEnx_clm contract lifecycle obligation extraction renewal and expiry auto-docketed - composes nx_docgen nx_docket nx_sign_ceremony
F99783475finfinREADYre-base nx_trust IOLTA and nx_billing onto nx_ledger - today both are bespoke double-entry; unifying proves the primitive and removes duplicate money logic
F99862795finfinBLOCKED F997LEDES e-billing export and matter costing on nx_ledger closing the fin-legal loop
F999710113wardenofficeREADYadd accept-loop socket timeouts to nx_legal_portal_daemon and the fin daemon - seq321 DoS starvation class proven live seq342 lists both as vulnerable
F1000610113supervisorfinREADYbring the fin daemon under the primary supervisor - one of 26 REG-ONLY daemons invisible to hostctl per seq216 two-fleet split
F1001910113analyst-sotaanalystDONEanalyst A1: analytics on REAL data AT SCALE -- shard past the seg_store 256/key scan cap and point ar_analyze at the live 150k-doc web index
F1002810113analyst-sotaanalystDONEanalyst A2: confidence INTERVALS on r + non-parametric (Spearman rank / permutation) + multiple-comparison correction
F1003810113analyst-sotaanalystDONEanalyst A3: AUTOMATED INSIGHT MINING -- point at a dataset with NO target, get ranked findings (segments, breaks, anomalies, Simpson reversals); compose the miner PrefixSpan lane not a reimpl
F1004710113analyst-sotaanalystDONEanalyst A4: MULTIVARIATE -- partial correlation / multiple regression so a driver can be reported CONTROLLING FOR the others
F1005710113analyst-sotaanalystDONEanalyst A5: CAUSAL layer + honest naming -- stop calling an associational argmax a key DRIVER; identification/adjustment/DAG (DoWhy-class)
F1006610113analyst-sotaanalystDONEanalyst A6: NL ANALYST INTERFACE -- plain-language question to structured nx_query to narrated result WITH grounding shown
F1007610113analyst-sotaanalystDONEanalyst A7: BI DASHBOARD -- query-driven data-bound interactive surface, quality judged by the nx_uiq family not by eyeball
F1008510113analyst-sotaanalystDONEanalyst A8: TRANSFORM-DAG (dbt-class) -- models + data tests + docs over seg_store, composing nx_store_lineage and nx_seg_schema
F1009700014refereeminer-sotaDONEminer MS06a BUGS miner: recurrence/clustering/aging over the 661-row debt- plane (census MC06 covered-but-UNMINED); done-when nx_debt gains a mine verb + gate + the census flips MC06 mined=1
F101091459companioncompanionBLOCKED F947SOTA-2026 MULTI-CHANNEL WARDROBE RUNTIME: she stays dressed via STRUCTURE-LOCK + IDENTITY channels, NOT prompt words. Measured basis F947 (English is a lossy codec: semantics 85pct, geometry 50pct, identity 0pct) + jul-2026 harvest (RF-inversion Taming/FireFlow/TightInversion/LORE = surgical edit w/ everything-else-unchanged; FluxAlreadyKnows = TRAINING-FREE identity, test before LoRA; ProductConsistency 2606.19103 = literally the clothing-persistence case; FreeFlux = which layers carry identity vs layout vs texture). Outfit change becomes a SURGICAL EDIT on a locked scene, never a re-roll. Grammar tables demote to L1 semantic prior + eval set. GPU-gated on BIOS flash.
F1011700014conductorminer-sotaDONEminer MS06b WORKFLOWS miner: mine plan- plane structure + planrun- step outcomes (census MC08 = 18 plans, ZERO readers); surfaces which steps fail/repeat = the workflow-quality signal
F1012700014refereeminer-sotaDONEminer MS06c HARNESS-OUTCOME journal + miner: census MC09 shows ZERO evidence for gates as a class (953 unjudgeable, no outcome history); emit gate verdicts to a journal then mine flake/erosion; composes the D001 anchor campaign
F1013810113analyst-sotaanalystDONEanalyst A2b: CONFIDENCE INTERVALS on r (Fisher-z) -- the third part of F1002, split out unbuilt
F1014800014modelwrightnsbuilderDONEPROMOTE 3B TO PRIMARY MAKER (measured 92pct union): re-run the official ensemble with bench_official3b_q4km as maker#1 (clean prompt) + 1.5B as maker#2 -- our nx_15b IS itself a stock qwen2.5-1.5b so this is a TIER UPGRADE not an external dependency; the sovereign moat is the HARNESS (symjudge+normalizer+gramdec+cascade+ensemble), weights are a hwtier- row. Done-when: official ledger shows 3-maker union >=92pct with 3B primary + nx_swebench_status republished + hwtier maker column flipped
F101573477finfinREADYF997 step 2 the actual re-base: make nx_trust_lib and nx_billing thin wrappers over led_ funcs keeping signatures so nx_trust_gate 7of7 is the regression oracle. BLOCKED ON DATA MIGRATION - re-basing changes the on-disk key layout and existing IOLTA rows are a compliance ledger where history is sacred rule 13; needs a read-old-write-new migration + byte-verify before any cutover. Equivalence PROVEN GREEN 10of10 by nx_ledger_trust_equiv_gate
F1016600014researcherminer-sotaDONEminer MS06e FEATURES miner: frontier- is COVERED by CPM (float/critical-path) but UNMINED (census MC05, 506 rows) - mine feature aging, lane churn, dep-bottleneck recurrence, done-rate per lane; done-when census flips MC05 mined=1
F1017500014conductorminer-sotaDONEminer MS06f LEASES miner: claims- is COVERED by walk (holder/TTL state) but UNMINED (census MC04, 221 rows) - mine contention hotspots, lease churn, crash-expiry rate per resource; done-when census flips MC04 mined=1
F1018913410companioncompanionBLOCKED F916MINE THE MEASURED CORPUS: the elder-ai data-vault holds ~28GB of SCORED generation data that dwarfs content.db (390MB grammar) and was nearly missed. MEASURED 2026-07-23 via nx_sqlite_rows: beauty.db/beauty_evaluations 568944 rows - quality.db/quality_gate_log 308683 + self_check_log 306266 - prompt_guard.db/conflict_log 301600 - prompt_data.db/render_tier_grade 102947 + ai_pathology_finding 101977 (the anti-body-horror defect corpus) + render_fact_event 19926 - scene_graph.db/graph_render_history 103167 = ~1.9M LABELED observations. This is the training+eval substrate F940 realism ruler / F942 word-impact / F954 convergence engine were designed around but never had. Plane the high-value tables, then use them as the eval set for the F1010 multi-channel work. NOTE fragment_combinations exists in quality.db but is EMPTY (0 rows) and combination_scores is absent from content.db - the v83 coefficients may be lost or live in tracking/feedback/reference_prompts (survey in flight).
F1019800014refereensbuilderDONESEV8 DEBT-RAIL IS LOSSY: nx_debt add dedup keys on OWNER not DESCRIPTION -- distinct findings are silently dropped while the message claims identical desc. MEASURED+REPRODUCED 2026-07-23: 3 adds refused; owner=supervisor twice (different text, sev7 then sev6) both matched existing_idx=688 (my earlier supervisor row); owner=referee with unrelated text matched existing_idx=605 (my earlier referee row). Not prefix-truncation (all differ in the first 32 chars), not sev. IMPACT: inverse of the tool purpose -- a lane filing its 2nd debt loses it AND believes it landed; defeats the DEC014 content-idempotence the message asserts; unknown how many findings were lost since. Cannot be filed via nx_debt itself (the bug blocks its own report) -- hence this frontier row. FIX: hash the DESCRIPTION as the key; gate teeth = two DIFFERENT descs same owner BOTH land + same desc twice yields exactly one row (the existing neg-control evidently only covered same-desc). Evidence + the finding this swallowed (seq688 refined root cause: reader-keep breaker IS engaging; real cause = leaked :8791 listen fd, F-210 class, gate on port-serving not process-name per seq648) are in project-nishi-dedup-migrate-judge-2026-07-23
F1020500014refereensbuilderREADYRETRACTION OF F1019 (read this WITH F1019, which is now closed): nx_debt add dedup is CORRECT and IS content-based -- there is NO finding loss. My error: I called it as (sev, desc, scope) but the signature is `add `, so the stored desc was the single word supervisor/referee and the dedup rightly matched MY OWN earlier row carrying that same one-word desc. VERIFIED by `nx_debt show 688` = epoch|7||open|supervisor. LESSONS: (1) READ THE SIGNATURE before accusing a tool -- I built a sev8 theory on 3 reproductions of my own misuse; (2) a REPRODUCIBLE wrong result is still wrong -- reproduction proves consistency, not correctness; (3) the louder the claim, the higher the proof bar, and mine was an inference chain with no read of the callee. WARNING debt rows 605/609/668/688 filed this session have scope/desc TRANSPOSED (long text in scope, owner word in desc) -- content is INTACT and readable via nx_debt show; do NOT re-file them. OPTIONAL real nicety: nx_debt add could REFUSE a desc that exactly equals a known owner/lane token, which would have caught this in one call
F1021500014modelwrighttoken-shiftREADYnx_hw_envelope: add a `verify [modelroot]` verb that fail-closes when a hwtier- row declares makers absent from the host. MEASURED NEED 2026-07-23: the NAS matched tier=laptop by RAM (35GB) while holding ZERO .gguf (glob *.gguf = 0), so `makers` emitted 3 phantom paths a consumer would only discover deep inside the maker loop. ROOT INSIGHT: a RAM tier states a CAPABILITY CLASS, it does NOT prove MODEL AVAILABILITY -- two different facts, and a row must never imply the second from the first. Interim cure already live: the nx_ensemble_routed.sh launcher now presence-checks every declared maker under MODEL_ROOT and exits 4 naming the missing file (neg-control proven), and the NAS tier row now says MODELS-NOT-STAGED-ON-THIS-HOST. The organ-side verb makes it structural rather than launcher-only, so any future consumer inherits the guard
F1022700014architectminer-sotaBLOCKED F973F973b HOST-DEPLOY MECHANISM (the remaining half of F973): nx_claude_harvest is now self-sufficient (delta hwm verb, gate 17/17) but its ELF only exists on the NAS - a second host cannot RUN it. Two concrete paths: (a) GET /api/artifact/ binary-fetch endpoint on mgmt (fetch+chmod+run, checksum-verified), or (b) build-on-host from synced source via the local toolchain. Done-when the laptop beat runs the ELF and nishi-ops PS parser is DELETED
F1023610113analyst-sotaanalystDONEanalyst A9: nx_analyze `mine` verb as JSON -- expose the no-target insight scan as structured findings for agents (today ai_mine emits human text; the report/sig/ci verbs are JSON)
F1024510113analyst-sotaanalystREADYanalyst A10: nx_analyze `report` should accept alpha + expose per-column PROFILE fields (min/max/mean/median/stddev/outliers) in the JSON, and a `profile` verb
F1025510113analyst-sotaanalystDONEanalyst A11: workflow recipe -- a saved Workflow that fans nx_analyze sig/ci over many column pairs then synthesises, proving the tool composes under orchestration at scale
F1026600014tools-apinsbuilderREADYseq722 FIX + ITS SAFETY GUARD (register drops the schema row): make ma_do_register write the tool_schemas.conf row in the SAME owner-gated handler that writes the allowlist row -- the schema= string is already accepted and currently dropped. ★★CRITICAL GUARD so the fix does not cause harm: schema ABSENCE is a LEGITIMATE do-not-advertise signal for internal tools (the banked noschema-deliberate-not-debt policy: nx_vault/nx_security_census/nx_evidence_refresh/nx_dataplane_census are deliberately schema-less). So (a) register writes a schema row ONLY when the caller PROVIDES schema=, never a synthesized default; (b) the one-time backlog reconcile-sweep must NOT synthesize schemas for the ~80 skipped tools -- each needs a per-tool human/owner decision (advertise vs deliberately-private). Blanket-clearing would EXPOSE private organs = a real harm. Done-when: an organ registered WITH schema= appears in tools/list with its inputSchema after one register call (no hand-edit), AND an organ registered WITHOUT schema= stays un-advertised. Interim per-tool workaround PROVEN 2026-07-23 (nx_fs_write conf-append anchored on header comment + reconcile). WARNING mgmt crown jewel, just recovered from seq709 memory outage -- nx_orchestrate on a calm-window precondition
F1027600014refereensbuilderREADYWIRE STORE-LAW ADHERENCE INTO THE AUTONOMOUS LOOP (operator flagged flat-files twice 2026-07-23: index flood + clock TSV 'tsv isnt what we build on'). The census ALREADY EXISTS -- nx_dataplane_census emits DP|| rows (verified live: found nx_worker_dispatch->worker_dispatch.tsv, nx_genealogist_gapscan->genealogist_targets.tsv, nx_vsz_*, + the clock seq726). CHECK-BEFORE-BUILD: do NOT rebuild the scanner. THE GAP is that it is measurable-but-UNTRACKED: (a) prints to stdout only, no durable evidence log (writes no knowledge/status/*.log unlike favela_census); (b) not MCP-discoverable (one of the 80 skipped-no-schema, seq722); (c) its output is not turned into tracked/ranked debt. SCALABLE FIX = apply the measure->rank->file autonomous-loop pattern that ALREADY EXISTS (F208 nx_backlog_feeder = evidence->frontier committer): (1) census colocates a knowledge/status/dataplane_census.log with a VERDICT + total count (favela_census is the exemplar to copy); (2) a feeder turns each DP row into a store-law debt row keyed by organ (content-idempotent, so re-runs add nothing); (3) cron beat on cadence. Result: the sovereign-formats-first law becomes a MEASURED, ranked, self-tracking dimension instead of caught-by-eyeball. Each converted flat plane = a seg-store migration rung (clockstate- for the clock per seq726 is the first). WARNING many flagged organs are sibling-owned -- the feeder FILES debt, owners migrate; never unilaterally rewrite a crown-jewel rail
F102891459companioncompanionDONEF975 RUNG 1 DONE 2026-07-23: nx_elara_state LIVE MCP TOOL = the companion state-coherence engine. Persistent agent-writable state (outfit/location/mood/arousal/energy/activity/arc_stage/dynamic) on seg-store via nx_registry (last-write-wins O(log)); every set journals to a real-clock TIMELINE = the flowing-day substrate. Gate 7/7 GREEN (deterministic re-runs via run-unique session id) + neg-controls (unknown-field refused, JSON-escaped hostile values). PROVEN over the REAL /mcp transport: agent set outfit=black_cocktail_dress then a SEPARATE get call returned it unchanged = she stays dressed across turns BY CONSTRUCTION. Full ship-complete lifecycle: register+schema+reconcile+cap(union nonce 747902)+verify-by-behavior. Rung 2 = the dress verb (validate vs elaragram- wardrobe grammar + location exposure ceiling) + F920 idle day-arc advance.
F1029800014conductortoken-shiftREADYDEDUP AUTO-FEEDER = the scale unlock for the D001 dedup machine (operator: 'it needs to scale'). MEASURED THIS SESSION: the walker+judge+beat drain the dedupq- queue hands-off (15 organs migrated 2026-07-23), but the QUEUE IS HAND-SEEDED (2-3 rows at a time) -- 568 organs carry the _w shape alone, ~2400 dup helpers total, so hand-seeding is O(200+ manual puts) = NOT scalable. The drain scales; the INPUT does not. BUILD nx_dedup_feeder (reuse the F208 nx_backlog_feeder evidence->committer pattern + the dedup editor's OWN shape table de_suffix _puts/_werr/_vlen/_slen/_cat/_catn/_put/_putn/_w): scan _hdl_build for organs defining >=1 migratable helper shape that (a) is byte-identical to the sj_ canonical and (b) whose organ is NOT already a dedupq- row and NOT already composing nx_sovjson_lib -> emit a content-idempotent dedupq- row per organ. Then walker+beat drain continuously, zero hand-seeding. GUARDS (both load-bearing): (1) ONLY-DETERMINISTIC filter -- exclude tree-scanners/censuses/live-probes (the ROUND-24 scope boundary: the behavior-equiv judge can't verify state-dependent organs, they harness forever; the feeder must NOT queue what the judge can't judge, or pre-mark them needs-review); (2) the feeder is itself a TREE-SCANNER -- honor the dup_source_check-v3 OOM lesson: allocate path buffers ONCE per organ not per file, declare the allocation profile, bound the sweep. SUB-RUNG: walker harness-count auto-park -- after N harness errors on a row, auto-mark needs-review instead of leaving queued forever (F208 bounded-retry pattern; nx_fwdconst_scan hit harness 2x this session and I marked it by hand). Done-when: feeder run seeds K deterministic-organ rows content-idempotently, a rerun adds 0, and the beat drains them with zero hand-seeding
F1030520212mgmtmgmtDONEF881 magic-ratchet exec port COMPLETED 2026-07-24 by the seq732 lane (coordination recovery, not a claim-jump): a sibling added the CALL md_exec_ratchet(nm,rpath) at nx_mgmt_api ma_do_build:1111 without the DEFINITION, leaving buildroot mgmt_api NON-COMPILING (undefined-fn) and blocking ALL mgmt builds. Defined md_exec_ratchet in nx_mgmt_data.nx (fork nx_magicratchet.elf , exit3=REFUSE, fail-open) mirroring md_exec_hostctl_capture2; nx_magicratchet.elf was already staged so F881 now ENFORCES live (deployed in the same never-brick mgmtapi train as the seq732 register-newline fix). If you were mid-design on md_exec_ratchet semantics, verify mine matches intent; it is minimal + fail-open.
F103181459companioncompanionDONEF975 RUNG 2 DONE 2026-07-24: nx_elara_state `dress [location]` verb LIVE = the wardrobe-VALIDATION primitive. She can ONLY wear an outfit that EXISTS in the elaragram- grammar plane (phantom outfit_key REFUSED, state untouched); on a real outfit it commits outfit(+location) and returns nsfw_level/location_tags + a location_appropriate verdict (sexy_witch @home=true, @beach=false). SCALABLE: reads the manifest + maps each segment .docs ONCE munmap'd (79ms full scan across 93 segs) after the naive 905x ss_get CRASHED via a seg-store mmap-leak (filed seq750, the real root). Gate 9/9 GREEN (T7 dress-real + T8 phantom-refused-state-intact). Promoted, .prev banked. Rung 2b = O(log) outfit_key->rowid index + wire location_exposure_ceiling (currently checks the outfit's own location_tags). NXR1 DRY lib filed seq751.
F103281459companioncompanionDONEF975 RUNG 3 DONE 2026-07-24: nx_elara_state `context ` verb LIVE = the INTEGRATION KEYSTONE. Joins her live state (outfit/location/mood/activity/dynamic/energy/arousal) with the elaragram- grammar's RICH outfit description into a persona-ready block in ONE call: {structured fields + a narrative string ready to inject into a chat/gen prompt}. LIVE-PROVEN over real /mcp (status=200): narrative='You are wearing a tiny black sheer babydoll dress with striped stockings and a pointed witch hat, in the home, feeling playful while dancing. Your relational dynamic is dominant.' This is 'what she SAYS == what she WEARS == where she IS' made a single queryable call -- the bridge F917 persona-compile + the gen composer consume. Gate 10/10 GREEN (T9 = outfit_key->name+description join). Reuses the scalable es_find_outfit (79ms). NEXT: wire F917 nx_card_compile to inject this context so chat reflects her dressed state.
F1033300014researcherminer-sotaEXTERNALF966b OPTIONAL external SPMF/PM4Py head-to-head - RECLASSIFIED as a NON-GOAL for the current miner: nx_actlog does CONTIGUOUS gap-bounded tool-transition mining (closed + confidence, construction-oracle benchmarked 15/15); general PrefixSpan/PM4Py do arbitrary-length non-contiguous itemset-sequence mining = a DIFFERENT problem, so a direct head-to-head is a category error not a gap. Only build a general PrefixSpan organ IF an actual use-case needs non-contiguous subsequences; then bench it vs SPMF. Until then the miner is SOTA-complete for its scope.
F103471459companioncompanionDONEF975 RUNG 4 (aliveness rung 1) DONE 2026-07-24: nx_elara_state `tick ` verb LIVE = she FLOWS THROUGH THE DAY. Advances her AMBIENT state (arc_stage=time-of-day period + energy) to the REAL CLOCK -- grounded in the actual hour (CDT via named ES_TZ_OFFSET_SEC, not an invented decay curve): morning/rested, afternoon/energetic, evening/relaxed, night/sleepy. Meant to run on a cron so she keeps EXISTING between interactions; does NOT touch user/agent-driven fields (outfit/location/mood/activity/dynamic) -- intentional overrides win via last-write-wins. LIVE-PROVEN over real /mcp (status=200): a fresh session ticked to arc_stage=night energy=sleepy matching the real late-night hour. Gate 11/11 GREEN (T10 = arc_stage lands on one of the 4 periods). The state engine now EVOLVES, not just persists = the continuity law satisfied at rung 1. NEXT (rung 2): apply the planed activity_state_effects (293 rows) + a daily-routine arc so activity/mood also drift; wire tick onto a clockjob.
F1035812311companioncompanionDONEF925 COMPANION SOTA RULER DONE 2026-07-24: nx_companion_sota LIVE MCP tool = MEASURED + LIAR-KILLED head-to-head (Nishi vs Replika/Character.ai/SillyTavern) over 9 capability axes. Doctrine-faithful: every Nishi cell is LIVE-PROBED by forking nx_elara_state (set/get, dress-phantom-refused, dress+context grammar join, tick, dynamic, context) -- a claimed HAVE that fails its probe is DOWNGRADED to GAP and verdict=RED; competitor cells = documented capability presence. ★THE LIAR-KILLER WORKED ON ITSELF: first run went RED (verified 5/6) because the rich_grammar probe checked the dress verdict for the description, which only surfaces in `context` -- fixed the probe (not a fake HAVE), now 6/6 verified verdict=GREEN. ★MEASURED RESULT: Nishi LEADS on wardrobe_coherence + measured_learning_corpus (NO incumbent has these), sovereign_private, flowing_day_continuity, rich_grammar; competitive on persistent_state/relational/persona; HONESTLY BEHIND (PARTIAL) on multimodal_gen = the GPU-flash-gated axis. Full lifecycle: register (seq732 fix HELD in prod, row clean no-glue) + schema + reconcile + cap(nonce 748010) + PROVEN over /mcp status=200 (nested fork daemon->sota->elara_state). Doubles as a live regression gate on the whole companion stack.
F1036710113finfinDONEnx_piotroski F-score 9-point fundamental-strength SHIPPED 12/12 - integer-exact, composes nx_xbrl (missing fact -> INCOMPLETE not fabricated); pairs Altman-Z distress + mgmt-grade stewardship = the fundamental-analysis triad
F103771459companioncompanionDONEF975 RUNG 5 (lived day) DONE 2026-07-24: nx_elara_state `tick` now advances a full DAILY ROUTINE with CONFLICT-RESOLUTION, not just time-of-day. es_daystate maps the real hour to arc_stage+energy+routine activity+mood (morning=coffee/bright, afternoon=busy/content, evening=unwinding/relaxed, night=sleepy/drowsy). ★THE KEY: activity+mood DRIFT with her day ONLY IF the agent has not set them since the last tick -- 'influenced by our interactions': an intentional override STICKS until the next natural drift (arc_stage+energy are pure-clock, always advance). Uses es_field_epoch (the stored per-field write timestamp). ★FIXED a same-second precision flaw live: seconds-granularity let drift clobber an override set in the same wall-clock second as a tick -> switched field stamps to MILLISECOND (sys_now_realtime_ms) so set-right-after-tick is distinguishable; override now sticks even same-second (demoed: set activity=dancing -> tick -> still dancing). Gate 12/12 GREEN (T11 = agent override survives tick). ★nx_companion_sota re-run GREEN 6/6 = live regression gate confirmed nothing broke. Promoted, .prev-rung4 banked. NEXT: apply the 293 planed activity_state_effects for richer drift; wire tick onto a clockjob.
F1038710113infrastructuresearchDONEnx_planescan DONE 2026-07-24 = the SOVEREIGN SCALABLE PLANE-ENUMERATION primitive (storage foundation, first-byte-up). GAP: nx_sqlite_rows planes carry no ids-index -> fetchable only by exact key; the naive ss_get loop LEAKS ~O(segments) mmaps/call -> ~87k maps on a 93-seg plane -> vm.max_map_count CRASH (seq750). nx_planescan = leak-free: read manifest, map EACH segment .docs ONCE, walk, munmap. ps_count + ps_find_field. PROVEN AT SCALE: counted batch_history 200264 records across 93 segments in <1s leak-free (the exact op that crashed). Gate 4/4 self-contained + neg-control. Canonical home of hardened ps_nxr1_field (seq751 partial). MCP lifecycle: register+schema+cap 748021+PROVEN /mcp status=200 count=200264. NEXT: refactor es_find_outfit onto it; ss_get mmap-leak seq750 stays crown-jewel.
F1039810113finfinDONEnx_finhealth composite health verdict SHIPPED 12/12 - the recombination capstone: composes Altman-Z distress + Piotroski-F strength + balance-sheet identity into ONE verdict; fail-closed (missing fact->INCOMPLETE, unbalanced books->FLAGGED override good scores). The atlas thesis in production over the real-data pipeline
F1040712311companioncompanionDONECOMPANION LONG-TERM MEMORY DONE 2026-07-24: nx_elara_memory LIVE MCP tool = she REMEMBERS things about YOU and RECALLS the relevant ones (Replika's flagship capability, now sovereign). remember [kind] -> NXR1 record on seg-store plane elaramem-; recall [max] -> keyword-overlap ranked (WORD-BOUNDARY match + recency), scalable via the leak-free map-each-segment-once scan. LIVE-PROVEN over /mcp status=200: 'what do you know about my coffee' -> 'you take your coffee black with no sugar'. ★CAUGHT + FIXED a real quality bug live: substring matching had 'do' match inside 'adopted' -> mis-ranked recall; fixed to WORD-BOUNDARY matching (gate T4). Gate 5/5 GREEN. Full MCP lifecycle: register(row verified post-race seq772)+schema+cap 748030+proven. ★nx_companion_sota UPDATED to 7 probed axes incl long_term_memory (nishi HAVE keyword; HONEST that Replika ships semantic embeddings = our rung 2), re-verified 7/7 GREEN. NEXT: semantic recall via a CPU embed path (rung 2) = close the one remaining companion gap besides GPU gen.
F1041710113finfinDONEnx_beneish M-score forensic earnings-manipulation detector SHIPPED 12/12 - integer-exact 8-index Beneish model (flagged Enron), fail-closed to INCOMPLETE on missing data (never a fraud accusation on partial facts). Completes the company-evaluation quartet: Altman distress + Piotroski strength + mgmt stewardship + Beneish fraud
F1042710113finfinDONEnx_dcf discounted-cash-flow intrinsic valuation SHIPPED 16/16 - integer-exact PV + Gordon terminal + margin-of-safety verdict (UNDER/FAIR/OVERVALUED); FAIL-CLOSED on the growth>=rate trap (refuses the absurd valuation analysts produce). The operator R-MKT4 projections/intrinsic-value-vs-price rung
F1043710113finfinDONEnx_bond fixed-income valuation SHIPPED 17/17 - integer-exact bond price + yield-to-maturity by INTEGER BISECTION (no float Newton-Raphson drift) + Macaulay/modified duration; FAIL-CLOSED refuses an unreachable yield (BOND_NO_CONVERGE) instead of inventing one. The fixed-income companion to the equity DCF - together they value both sides of a portfolio
F1044710113finfinDONEnx_taxlot capital-gains tax-lot + wash-sale SHIPPED 24/24 - exact realized gain/holding-period/FIFO + IRC 1091 WASH-SALE disallowance: a loss with a replacement buy within +/-30 days is DISALLOWED and DEFERRED into replacement basis (preserved not destroyed, allowed+disallowed conserves the loss); FIFO overdraw-safe. Fail-closed on a real IRS compliance trap consumer tools get wrong
F1045710113legallegalDONEnx_sol statute-of-limitations deadline engine SHIPPED 29/29 - exact proleptic-Gregorian deadline math (leap-year + Feb-29->Feb-28 clamp correct), tolling, discovery rule; FAIL-CLOSED returns UNDETERMINED-FOR-REVIEW on an unestablished accrual instead of a confident wrong date. Missing a limitations deadline is the #1 legal malpractice claim - the legal analog of the finance fail-closed-on-unknown work
F1046710113finfinDONEnx_amort loan amortization SHIPPED 17/17 - level payment by INTEGER BISECTION (not float pow) + CONSERVATION-exact penny allocation (principal reductions sum to principal exactly, final payment clears to zero, no cent lost/invented on the last payment = the naive-amortizer failure mode); fail-closed on nonsense input. Composes with the financing-shield true-cost ranking
F1047710113legallegalDONEnx_redact redaction-completeness verifier SHIPPED 17/17 - the content half of e-discovery production (nx_ediscovery=allow-list): producible only if reviewed AND redaction-clean. Deterministic byte-exact SSN detection with precise boundaries (no false positives on wrong grouping/embedded digits) + privilege-marker scan; FAIL-CLOSED hard block not a flag-for-review, so PII/privilege cannot leak by omission
F1048710113legallegalDONEnx_damages prejudgment interest SHIPPED 16/16 - statutory prejudgment interest on an EXACT day count that COMPOSES nx_sol's Gregorian date engine (DRY, not re-implemented): simple + annually-compounded, integer minor units; FAIL-CLOSED on a judgment-before-accrual interval or negative inputs. Day-count + compounding errors are exactly where damages calculators go wrong. Legal-finance bridge organ
F1049810113finfinDONEnx_finance_svc agentic finance SERVICE FACADE SHIPPED + PROVEN LIVE - one MCP tool over the integer-exact finance stack (dcf/bond/taxlot/amort), verb-dispatched JSON, versioned envelope, structured 400/404/422, agentic describe verb (tools/list analog). FAIL-CLOSED refusals survive the facade as 422 REFUSED (verified live: dcf.terminal growth>=rate AND bond.ytm unreachable both returned 422). The MCP/API/agentic/workflow-ready/scalable layer over the gate-proven libs
F1050810113legallegalDONEnx_legal_svc agentic legal SERVICE FACADE SHIPPED + PROVEN LIVE - one MCP tool over the exact legal-calc stack (sol + damages), verb-dispatched JSON, versioned envelope, structured 400/404/422, agentic describe. WORKFLOW-COMPOSABLE proven live: sol.deadline emits a daynum that sol.assess consumes (OPEN/TIME-BARRED + days_remaining); damages.days feeds damages.simple. FAIL-CLOSED survives as 422 (judgment-before-accrual verified 422 live). Completes the agentic MCP layer for BOTH finance and legal lanes
F105122661616166207535900014next: retrieval few-shots from fixcorpus into the episode prompt (replace hardcoded min2/abs1), then callee-signature context assembly, then scale mine to 100+ across _hdl_buildmeasured: mine summary counts + selftest VERDICT=GREENBLOCKED 2026-07-24swe-harness
F10527110113engineer-gamesgame-adultBLOCKED F913Adult-game emitter PROGRAM card (roadmap 07-24): compose reference-title caps (Breeders/MGI/CM2/CarnalInstinct/DIK-VN/VaM/CoC/MMO-class) from certified parts; humanoid-only; adult-invariant fail-closed at compile; private surfaces only; roadmap=memory project-nishi-adult-game-emitter-roadmap-2026-07-24
F1053610113engineer-gamesgame-adultDONEPhase 0 RULER: extend nx_gamebench w/ adult reference set (10 titles in_corpus:0) + new axes (wardrobe-state, genetics-breeding, paired-anim, still-consistency, llm-npc, vr-runtime, customizer, relationship-sim, transformation, intimacy-scene, voice, adult-gating); fix seq359 var-len cap vectors; recomputed gap_queue = the build order
F105461459engineer-gamesgame-adultDONEnx_wardrobe_state part+gate: clothing/exposure/transformation state machine w/ coherence invariants (never self-contradicts across save/load/scene; adult-invariant fail-closed) = F975 game-side twin; M1 unblock
F1055520210engineer-gamesgame-adultDONEM1: CoC-class text RPG emitted from certified parts (story_vm+plotgen+rpgstats+companion+gamesave+wardrobe_state), browser-published on a PRIVATE gated surface + CDP-playtested win AND loss
F1056812311companioncompanionBLOCKED F925SEMANTIC-SALIENCE MEMORY (Nomi Identity Core parity) = the VALIDATED 2026 memory SOTA, nx_elara_memory rung 2. Grounded via live industry validation 2026-07-24: the memory LEADER is Nomi (NOT Replika/Character.ai) whose Identity Core does SEMANTIC recall + SALIENCE (actively decides which details matter) + PATTERN extraction (throughlines not just facts), recalling 8 months back UNPROMPTED. Nishi nx_elara_memory today = KEYWORD word-boundary recall only = behind. Rung 2 = (a) semantic recall via nx_vec_embed (sovereign distributional co-occurrence embeddings, exists) + cosine over the memory plane, (b) a SALIENCE score at remember-time (importance/emotional-weight so trivia decays and load-bearing facts persist), (c) pattern/throughline extraction. Non-GPU (distributional embed is CPU). Composes the leak-free nx_planescan for scalable recall.
F1057913410companioncompanionBLOCKED F922REAL-TIME AUDIO-DRIVEN VIDEO AVATAR = the MOVED 2026 synthetic-human frontier (validated vs research 2026-07-24). Our gen program (F1010) is STILL-IMAGE-centric, but the industry+research SOTA has MOVED to real-time audio-driven VIDEO avatars: HeyGen Avatar V (identity consistency at the MODEL level, not post-processing), Live Avatar 2512.04677 (identity+color stability on long-form, perpetual real-time synthesis on diffusion backbones), EchoMimicV3 2507.03905 (1.3B, identity+clothing+body-motion), Avatar Forcing (real-time interactive head avatar via diffusion forcing, low-latency user audio+motion), MIDAS 2508.19320. Open research problem industry-wide: LONG-HORIZON IDENTITY DRIFT (current metrics inadequate). ⇒ the synthetic-human vision (talk/dance/live) needs a VIDEO-avatar track feeding off the F1010 still-image identity work + the F975 state engine for what she does; addresses drift with the F947 identity ruler. GPU-gated on the BIOS .109 flash (rides F921/F922).
F1058522410engineer-gamesgame-adultDONEnx_coc_engine SHIPPED (M1 engine half): CoC-class text-RPG emitter composing plotgen+story_vm+rpgstats+wardrobe_state+gamesave; 6/6 self-proofs (winnable/LOSABLE-real-death/deterministic/save-transparent/ADULT-GATE-HOLDS-THROUGH-LOOP-mutation-proven/solvable); LIVE MCP tool cap nonce 1784925186, verified 6/6 over /mcp. REMAINING M1 half = browser-playable page + CDP playtest
F1059405756391848798718400014next: min-similarity threshold in fsl_shots (weak shots must not displace the proven pair); callee-signature context (lets non-leaf organs heal); mine 100+ via mineadd; germination ground-state seed = the from-zero self-heal sporemeasured: gen0=31% gen1=50% over identical manifest + flywheel selftest GREENBLOCKED 2026-07-24swe-harness
F1060522410engineer-gamesgame-adultREADYM1 VISUALS + CLEAR FLOORS shipped: rendered character/scene, state-reactive SFW outfit tiers, real-time feedback; measured nx_game_critic structure 543 BASIC (Floor 1 cleared), anti-Goodhart held (dither capped TOY at coherence gate, +texture disclosed not headlined). NEXT rung = character-render-realtime = REAL geometry/shading detail (not texture) to earn RICH/SOTA honestly
F106168380233233320822300014fix gate-coverage measurement (cross-dir+plan gates); eat 733 mmap-in-loop backlog; add reproducible-build ruler + SBOM/provenance + symjudge-property-density (seL4 cousin)measured composite=807permille+selftest6/6; INFINIGEN cousin nx_infinibug selftest 4/4 (seeded procedural function-world, difficulty dial); bugforge mines it source-agnosticallyBLOCKED 2026-07-24worldclass-assurance
F1062000014dhDONEa
F106393038engineergame-adultDONEvsbench: head-to-head ruler vs 10 adult ref titles (titles_reference_run 0->N)
F1064820212engineergame-adultREADYvsbench: capture T-EXT ground-truth evidence per ref title + anti-fabrication gate
F1065820212engineergame-adultREADYvsbench: ADVANTAGE-weighted queue (where can we EXCEED) beside parity gap_queue
F1066830311engineergame-adultDONEnx_genome genetics-breeding part = flagship husbandry RPG rung C1
F106793368engineergraphicsREADYVISUAL-LANE CONTRACT for vsbench: emit MINE permil + capture a THEIRS artifact for visual-scene-quality (worklist rank1) and character-render-realtime (rank3)
F106892359engineergame-adultREADYvsbench: convert visual-lane captures into TIER-3 EXT cells so beat_index leaves UNGROUNDED (first real head-to-head number)
F1069730311engineergame-adultREADYRE-EARN the 4 demoted gamebench HAVEs with real evidence (walk t0/t3 screenshots, procgen terrain has NO producer at all, bunny needs bun_zipper on NAS)
F1070820212engineerrigorREADYHARDEN artifact teeth ecosystem-wide: existence-only checks are foolable by a blank render; require non-vacuity + GREEN-gate receipt
F1071800014engineer-graphicsmhDONEbeyond-metahuman P0: nx_mhbench ruler LIVE (19 axes, gap_queue, gate 5/5 over /mcp, /world/metahuman GREEN)
F1072900014engineer-graphicsmhREADYbeyond-metahuman P1 FACE: head oracle (seq805) + true feature geometry + per-part density (head judge 227 = worst region)
F1073900014engineer-graphicsmhREADYbeyond-metahuman P2 SOFT-TISSUE FORM: anatomically-shaped bases judged per-region/multi-view + seq798 landmark-warp HF re-run (seq793/795/823)
F1074900014engineer-graphicsmhREADYbeyond-metahuman P3 VILLAGE: correlated anthropometry sampler + red-teamable aesthetic scorer + population loop w/ family latents (seq804)
F1075800014engineer-graphicsmhREADYbeyond-metahuman P4 HAIR+SKIN+WARDROBE: B-spline strand grooms w/ softdyn physics + pore displacement at body-scale + cloth render
F1076800014engineer-graphicsmhREADYbeyond-metahuman P5 RIG PARITY: skin the PRODUCT body to the rig by rule + bind softdyn tissue + expression bases as facial rig (determinism = the exceed)
F1077600014engineer-graphicsmhREADYbeyond-metahuman P6 LIGHT TRANSPORT + REALTIME: SSAO/soft-shadows/DOF/GI then LOD+realtime -- AFTER relief lands (GX-39 law: light reveals only what geometry offers)
F1078603311claude-fablegenDONEnx_refbench reference-retention ruler: ship-complete + published
F1079802212unassignedgenBLOCKED F944Acquire ControlNet + PhotoMaker weights to unblock L2 geometry and L4 identity channels (code is compiled and plumbed, zero weights on disk)
F1080703311unassignedgenBLOCKED F954Convergence loop: drive reference-conditioned generation to a RETENTION TARGET using nx_refbench as the judge, instead of a guessed denoise
F1081900014engineer-graphicsmhDONEbeyond-metahuman P1a: nx_headcrop head-scale INSTRUMENT live (gate 7/7 MCP) + baseline MEASURED detail 79 (was flattered 227) + published w/ side-by-side
F1082900014engineer-graphicsmhDONEbeyond-metahuman P1b: NOSE+LIPS+BROW+CHIN as ruled GEOMETRY (face is painted not sculpted, seq857) -- must move head-scale detail 79
F1083900014engineer-graphicsmhEXTERNALbeyond-metahuman P1c: LIPS+BROW-RIDGE+CHIN+EARS+orbital depth as ruled PARTS (nose proved the pattern, part cap now 48) -- must move head 288
F1084900014engineer-graphicsmhDONEbeyond-metahuman P1d PREREQ (seq859): landmark-scored HEAD judge + retire mat-5 painted bands + ROTATION ABOUT Z for horizontal features -- all 3 before retrying lips/brow/chin
F108580559unassignedgenBLOCKED F1010Edit-compliance judge: measure whether the REQUESTED edit landed, optimise jointly with retention (retention alone converges on change-nothing - MEASURED)
F108670559unassignedgenBLOCKED F1010Masked inpaint edit channel: engine already accepts a mask param; garment-region mask is the edit mechanism retention-lock lacks
F1087900014engineer-graphicsmhDONEbeyond-metahuman P1d-1 DONE: nx_facemark landmark PLACEMENT judge LIVE (7/7 gate MCP); ranks rejected face 0 vs shipped 318 where variance judge inverted (357 vs 288)
F1088900014engineer-graphicsmhDONEbeyond-metahuman seq864 EATEN: nx_facemark now ANATOMICAL (2->19 landmarks, noise-floor-derived threshold 12, gate 7->8 GREEN); RETRACTED my own 318-vs-0 as a low-res artifact -- at anatomical resolution both faces tie 84/83
F108970559unassignedgenBLOCKED F1086Garment segmentation to replace the rectangle box in nx_recolor (box seam is visible; skin-vs-fabric segmentation is the real mask)
F1090600014unassignedgovernBLOCKED seq872Atlas discovery capability-KIND axis: score cross-KIND pairs (emitter x ruler) instead of proxying adjacency via plane-mentions
F1091900014engineer-graphicsmhDONEbeyond-metahuman P1e: OFF-MIDLINE sampling in nx_facemark (seq873) -- lateral profiles so brow/cheekbone/jaw are measurable; then lips+chin as Z-rotated parts (midline judge IS valid for those)
F109260559unassignedgenBLOCKED F1089Multi-hue + texture garment segmentation (current mask is single-hue + saturation floor; patterned or dark fabric will defeat it)
F1093900014engineer-graphicsmhDONEbeyond-metahuman P1f: LIPS + CHIN as Z-rotated parts, parent-derived attachment, retire their painted band -- midline lane 69 is the number to beat (it is the WORST lane and unchanged by the brow)
F1094090951build order is the ruler gap_queue not taste; rungs D2 shared gallery, E3 shared contacts, E4 portability, F1 gifts all rank 9 and 3 of 4 are SHARING = build the share plane onceBLOCKED Tsuite
F1095090952derived from the measurement not taste: 3 of the ruler's top-4 gap_queue entries (D2 E3 F1) are sharing features, so building them separately would be three copies of the same who-may-see-what rule (rule 15)BLOCKED Tsuite
F1096090952ruler D2 rank 9: Google shared albums + Apple Shared Album both cited, ours ABSENTBLOCKED Tsuite
F1097090952ruler E3/E4 rank 9 each: Google Workspace + Apple cited for both, ours ABSENT; E2 rank 6BLOCKED Tsuite
F1098080862ruler F1 rank 9, F2 rank 3BLOCKED Tsuite
F1099080862ruler C1 rank 6 C2 rank 6 C6 rank 6; debt seq885 = /calendar currently 200s with the generic fallbackBLOCKED Tsuite
F1100900014engineer-graphicsmhREADYbeyond-metahuman P1g: PERFECT THE LIP UNIT (seq888) -- width, distinct upper/lower masses, philtrum, vermilion edge, commissure tuck, height vs oracle landmark; then CHIN as the next unit
F1101800014operatorgenBLOCKED seq887Z-Image ref channel: extend positional encoding for concatenated ref_latents (z_image.hpp:448 assert) - unblocks native identity conditioning with NO download; supersedes F944
F1102900014game-adultgame-platformDONEnx_input_abstract: ONE game loop driven by keyboard/gamepad/touch/VR-controller (input-abstraction-layer GAP; the enabler for desktop+VR+mobile without per-platform forks)
F1103900014game-adultgame-platformDONETap targets to 44px WCAG 2.5.5 across frontier/drift/m1_engine (universal MEASURED mobile defect: 30-43px on all three)
F1104900014game-adultgame-platformDONEm1_engine responsive canvas: fixed 442px forces a 1.221x shrink-to-fit zoom-out on a 390px device; controls sit 2.05 screens below the fold
F1105900014game-adultgame-platformDONEplatform-desktop-native: interactive windowed sovereign client (gap_queue RANK 1, 21 of 22 titles blocked; engines today play full loops but headless)
F1106900011game-adultgame-platformREADYvr-runtime rung 1: stereo pairs exist, headset runtime/compositor does not (4 titles: CM3D2, VR Hot, Virt-a-Mate, Operation Lovecraft)
F1107800014unassignedgenBLOCKED F1101Reference WEIGHT control for the Z-Image ref channel (it currently dominates: at denoise 1.0 with one ref it reproduces the reference and ignores the prompt)
F1108900012engineer-graphicsmhREADYbeyond-metahuman P2 KEYSTONE (seq900, operator): PROCEDURAL SKULL as the head BONE layer -- inside-out; vault/frontal+supraorbital/orbits/nasal aperture/zygomatic/maxilla/mandible+mental protuberance, scored vs the 171k BodyParts3D skull oracle; then muscle+fat; face features become PROPERTIES of bone not parts stuck on skin
F1109900014engineer-graphicsmhDONEbeyond-metahuman P2a DONE: nx_skullgen shipped+gated 6/6 (7 bones, 34 rings, 6 hollows, dimorphic, DETERMINISTIC) -- machinery for generating a skeleton from anatomical rules now exists
F1110900014engineer-graphicsmhREADYbeyond-metahuman P2b (seq908): SKULL ANATOMY v2 -- shrink vault to BRAINCASE ending above the orbits, attach every facial bone RELATIVE TO THE VAULT SURFACE (parent-derived, bone-to-bone), and make orbits/nasal aperture OPENINGS BETWEEN SHELLS (relief cannot cut a hole); score vs the 171k oracle
F1111930311engineer-videovideoREADYSEARCH-COUNT reduction: the MEASURED 80.6pct hot path (supersedes F937 SATD, refuted 0.0pct). Bit-exact rung first: prime vm_search cap b[0] with SAD at the vcv-10 median-predicted MV without recording argmin (ring-order strict-< argmin provably unchanged, far losers abort after 1 row). Then measure adaptive-R + pre-2000 patent-clean hierarchical (TSS1981/NTSS1994) with a BD sweep since those DO move the bitstream.
F1112900014game-adultgame-platformDONEWIRE nx_input_abstract into the 3 shipped games (m1_engine/frontier/drift) so the certified part becomes REAL platform reach -- the wired standard, not the part standard; dep F1102
F1113900014game-adultgame-platformDONEnx_input_abstract JS last-mile mirror + bit-exact cross-impl proof of the input checksum (the GX-11 pattern), so browser input provably equals the sovereign core
F1114300014engineer-media-READYMM-1 nx_mmbench self-emitting: regenerate /mmbench.html from the organ on a cron beat with an append-only evidence line, retiring the hand-published snapshot
F1115600014engineer-media-READYMM-2 metadata-scrape to SOTA (nx_mmbench gap_queue #1, demand 6/12 -- the largest un-migrated surface: port the old Python media-manager scrapers/metadata extraction to sovereign organs writing into the vault schema)
F1116500014engineer-media-READYMM-3 collections-series derivation (nx_mmbench gap_queue #2, demand 5/12 -- group vault items into series/collections; unblocks faceted-browse and library-fulltext-search)
F1117090952MEASURED by runtime/_hdl_build/nx_emptylit_probe.nx run-exit=0; debt seq907 sev8; found because it silently broke the suite ruler's evidence-root resolver while the source read as obviously correct; workaround = spell the local root './'BLOCKED Tcompiler
F1118500014engineer-media-READYMM-4 multi-user-profiles (nx_mmbench gap_queue #3, demand 5/12 -- per-user identity across gallery/vault/companion; prerequisite for watch-history and per-user companion memory)
F1119300014engineer-companion-READYMM-5 companion memory-consolidation dream-cycle (nx_mmbench gap_queue, demand 3/12 -- Replika/SillyTavern/ChatGPT all ship it; cron summarises elara_mem.log into durable declarative facts = long-term growth beyond a 32-turn window)
F1120900014engineer-graphicsmhDONEbeyond-metahuman P2c RECONCILE BEFORE BUILDING (seq926): read nx_faceanat + nx_anatstack + nx_faceprofile_gate FULLY, coordinate w/ the SDF face lane, port their FORENSIC TISSUE-DEPTH + facial-canon DATA into the mesh canon, retire the duplicated derivation -- BLOCKS skull v2
F1121700014unassignedgovernBLOCKED seq207Retire buildroot/runtime/runtime/ nested tree (16 files landed one level too deep from a pack containing runtime/; causes 6 of 33 divergent clobber hazards) - structural fix
F1122600014unassignedgovernBLOCKED seq207Coordinated sweep of the 21 runtime<->_hdl_build divergent dual-copies via nx_dedupe_source (proven soft-delete+backup); needs lane announcement first
F1123800014engineer-graphicsmhDONEbeyond-metahuman DISCOVERABILITY FIXED (seq941): /world/metahuman was unlinked for 9 rounds; inbound banners added to anatomy_gpu+graphics+human_roadmap, all verified 200 -> 200 on clean URLs
F1124700014unassignedgovernBLOCKED seq207Extend nx_dedupe_source tree scope to match nx_dup_source_check (detector discovers 13 trees, reconciler knows 2 - 12 hazards are structurally unfixable by the tool)
F1125400014unassignedgovernBLOCKED seq207nx_dup_source_check should EXCLUDE runtime/_retired from hazard triage (an archive is meant to diverge; 3 false hazards)
F1126090951divergence 1 KEPT as a facade tightening (claim-before-share; rebac would bootstrap a first grant on an ownerless object) - divergence 2 MIGRATED as DATA (new album otype + curator relation in rb_expand) not dropped - the GIFT half (ownership transfer + provenance via new rb_list_subjects_history) was the only genuinely new capability. debt seq930 EATEN. LAW: a dedup search belongs in DESIGNING a primitive, not in cleaning up after oneBLOCKED Dsuite
F1127600014unassignedgovernBLOCKED seq207nx_dup_source_check: classify hazards by BUILD-RESOLVER PROBE-PATH membership (hazard = dup across two probed trees); bin/ and _retired/ are never probed so their twins are litter not hazards - would take honest divergent 6->0
F1128900014engineer-graphicsmhDONEbeyond-metahuman P2d: apply the REST of face_tissue.dat -- nose/lips/chin/cheekbone/temple to canon y + tissue-depth-derived projection (brow+eyes done, headline 92->138); then F1110 skull v2 uses the SAME table for bone = skin minus tissue
F112988086researcherpapersREADYAS-4 INDEPENDENT grader for manuscripts: RT-004 has NO admissible grade because nx_paperbench and nx_paper_audit share an author (R6 leakage). Needs a grading path authored by a different party/lane to feed pp_verdict dual-grade arm. THE blocker for any published quality claim.
F113066068librarianpapersDONEAS-5 eat the nx_papers_index HARDCODE: it lists 3 papers as baked HTML (stale 2026-06-17, violates rule 11 data-driven) and points at auth-gated /wiki/ which is NOT publicly citable. Compile the index from the papers registry and publish a PUBLIC research index carrying RT-004.
F113177077refereepapersDONEAS-6 regression proof INTO THE BEAT: the 4 paper-lane gates (paperbench 16/16, forge 9/9, panel 9/9, audit 9/9) are green but run only when a human invokes them. Wire into the zero-Claude beat with an append-only evidence line, per the law that a beat with no evidence line is a claim not a control.
F1132500014engineer-media-READYMM-6 wire mx_extract into the media INGEST path (nx_media_extract is deployed+callable as nx_media_probe but the ingest pipeline still never calls it -- this is the difference between operator-reachable and product-wired)
F1133300014engineer-media-READYMM-7 deploy the 2 remaining DARK organs nx_feed_extract (rss-auto-acquire) + nx_archive_capture (site-scraper-adapters) -- both already built in-tree, both scored GAP only because nothing is deployed; deploy work not build work
F1134900014game-adultgame-platformDONEWire nx_input_abstract into /world/frontier and /world/drift (m1_engine done F1112); both still fail H8 playable-without-scrolling and the 44px minimum
F1135900014game-adultgame-platformREADYSuppress the global nx-ad-slot on /private/* and raise its 20px anchor to 44px -- it is the ONLY remaining sub-44px tap target on every page (seq892)
F1136900014engineer-graphicsmhREADYbeyond-metahuman GOAL (operator): MetaHuman quality is the FLOOR, reached bottom-up by building each layer intricately -- 4-stage industry pipeline: 1 skeletal generation (bone lengths+joint constraints from height/proportion params) 2 MUSCLE SIM (volume-preserving meshes BOUND TO BONES, authentic bulging) 3 FASCIA+FAT soft-body layer bridging muscle to skin 4 SKIN DEFORMER envelope computing wrinkles/stress from inner-layer movement. Bench: Ziva VFX, Character Creator, MetaHuman Creator, Houdini
F1137900014engineer-graphicsmhDONEbeyond-metahuman STAGE 3 CHEAPEST BIG MOVE (seq973): BIND nx_softdyn (spring-damper soft-body, gate 7/7, stability enforced by construction, replay-deterministic) TO THE EMITTED MUSCLE+FAT LAYERS = fascia/fat tissue sim, our AdonisFX/Ziva-class primitive already built and never wired
F1138900011engineer-graphicsmhREADYbeyond-metahuman ANNY-TEMPLATE UNBLOCK (seq973/seq804/seq823): encode the correlated shape manifold from PUBLISHED WHO anthropometric statistics (scan-free, differentiable, continuous phenotype axes) -- the covariance we need is PUBLISHED not scanned, which retires the one-cadaver blocker on the village sampler
F1139500014supervisor-READYMM-8 GENERALISE the built-vs-live sweeper: nx_reader_liveness covers the reader lane and nx_mmbench covers media+companion, but NO ecosystem-wide organ answers 'which capabilities are built and unreachable' across all lanes. Compose the two (bin_ref serving-binary test + per-lane cap tables) into one sweeper feeding the RACI monitor activity (A=supervisor)
F1140500014unassignedgovernBLOCKED RK001nx_page_verify intermittently lands on the DSM :443 co-squat and reports 404/RED for pages that are live 200 - should use the deterministic connect-override like nx_https_get
F114170559unassignedgenBLOCKED F1092Multi-modal garment mask: accept several hue modes learned from the seed region (nx_color_v2 k-means palette already exists) - the rotate verb is inert until the mask stops being single-hue
F1142600014referee-READYMM-9 harden nx_page_verify (seq990): always apply the our-domain connect-override or FAIL LOUD; print the observed X-Served-By and REFUSE a verdict when the responder is not nishi-host/nishi-substrate; auto re-probe any our-domain RED with the explicit override before reporting. This tool manufactured a false sev-7 debt that reached a live page
F114370559unassignedgenBLOCKED F1141Multi-modal mask via CORE DILATION not bounding box (bbox admits background hues as garment tones -> grade 371 vs champion 620); dilate the trusted core to reach an adjacent second tone
F1144600014referee-READYMM-10 RECOVER the nx_page_verify auto-override source before any rebuild (seq1008 sev8): the live elf implements our_domains.conf auto-override, NO .nx source in the tree does, so the normal build+promote loop would silently regress every our-domain verification to a :443 coin flip. Reimplement per the documented contract (host TAB endpoint; fail-safe when conf absent; explicit argv[2] wins), gate it, then rebuild. BLOCKS F1142
F1145500014architect-READYMM-11 build a MISSING-SOURCE detector for the fleet (generalises seq1008): nx_dup_source_check finds duplicate sources but nothing detects a DEPLOYED binary whose source is absent. Diff live-elf string literals against the source tree per organ; any live string with no source origin = unreproducible binary = a rebuild landmine
F1146900014engineer-graphicsmhDONEbeyond-metahuman STAGE 1 PART A: nx_skelgen LIVE (gate 8/8) -- anthropometric bone lengths per-mille of stature (scale-free, dimorphic, rounded) + FAIL-CLOSED joint constraints; composes with existing nx_skeleton rather than duplicating it
F1147900014game-adultgame-vsbenchDONEScore the remaining 8 adult refs on visual-scene-quality with nx_frame_score (recipe proven: store screenshot -> png2hex 400x240 -> chunkup 60KB -> score on NAS); each cell = one more EXT row
F1148900014game-adultgame-vsbenchDONEnx_vsbench_gate: anti-fabrication (every EXT cell's artifacts stat on disk: ours_m1.nxfh/breeders1.nxfh pattern) + anti-vacuity (zero EXT => headline MUST be 0/UNGROUNDED) + the hardcoded-tally regression tooth (per-title rows must be COMPUTED from cells)
F1149900014engineer-graphicsmhDONEbeyond-metahuman STAGE 2 LANDED: nx_myo volume-preserving muscle bound to bone (gate 8/8) -- radius=r0*sqrt(L0/L), bulge follows from joint angle alone; measured 95pct bulge at full flex with volume error under 1.5pct across the whole range
F1150000014engineer-bioacoustics2026-07-25BLOCKED seq1007 reed relaxation-regime not flutter; seq1025 source too low-bandwidth for short tracts; no licensed reference corpusbioacoustics
F1151900014engineer-graphicsmhDONEbeyond-metahuman STAGE 4 LANDED: nx_derm skin-as-envelope (gate 8/8) -- position DERIVED from muscle+fat so a bulge raises skin by construction; wrinkle from SIGNED compression only (bulge side taut, flexor side creases 80->400 bounded). ALL 4 PIPELINE STAGES NOW HAVE PHYSICS
F1152900014engineer-graphicsmhDONEbeyond-metahuman INTEGRATION LANDED (seq1031): nx_flexbody applies myo bulge + derm envelope to the REAL emitted body -- 1189 permil girth at 90deg, 64896 verts swelled, bone 0, render +6126 px VISIBLE, gate 5/5. Defect filed: band needs a TAPER profile (muscle bellies taper to tendons)
F1153900014engineer-graphicsmhDONEbeyond-metahuman P3 NEXT: bind the swell to NAMED MUSCLES with real origins/insertions on the generated skeleton (the taper now makes overlapping muscles sum without seams); then joint positions from skelgen + rig-driven fascia
F1154900014engineer-graphicsmhREADYbeyond-metahuman NEXT (seq1032): wire nx_myoset's composed per-station profile INTO nx_flexbody so the body swells per named muscle instead of one band -- mechanical wiring, both organs gated
F1155900014engineer-graphicsmhDONEbeyond-metahuman F1154b (seq1033): per-LIMB axis for the radial swell (toes explode because they run forward, far from the BODY axis) + derive the profile normalisation from the muscle peak sum instead of the invented FB_PROFMAX constant
F1156800014engineer-graphicsmhDONEbeyond-metahuman NEXT (seq1034 residual): toe-tip axis conditioning -- thin structures give a poorly-conditioned local axis; fall back to the nearest well-populated band ABOVE so the limb axis CONTINUES rather than reverting to the body centreline
F1157700014engineer-graphicsmhREADYbeyond-metahuman (seq1037): TOE GEOMETRY defect is PRE-EXISTING in nx_body_gen/nx_body_proc emitRaySet foot digits -- proven by rendering the UNDEFORMED baseline; investigate at current radial/sub settings
F1158900014game-adultgame-procgenDONEworldpipe LIGHTING stage (sun direction + slope shading + per-biome albedo variation) -- the MEASURED palette lever (227 vs 1904); re-score on nx_frame_score to prove the climb
F1159900014game-adultgame-procgenDONEwire the staged world INTO a playable surface (frontier/drift map bg or a /world/explore page) so the OURS head-to-head artifact is a GAME frame containing it
F1160500014architect-READYMM-12 FLEET SWEEP with the new DEPLOYED-AHEAD verdict: nx_stale_check now discriminates deployed-ahead from source-ahead, and found a second unreproducible binary (nx_media_inventory, 32KB gap) on a 4-organ sample. Sweep the whole registered-tool set, rank by gap size, and recover source for each hit before any rebuild
F1161700014engineer-compiler-READYMM-13 UNBLOCK THE BUILD-BLOCKED HALF OF THE FLEET (seq1052): 50pct of sampled deployed organs have source that no longer compiles because the compiler gained safety checks (fwd-const SS_MF_MAGIC, sys_accept arity). Sweep all 435, group by compiler-error class, fix per class not per organ, ranked by load-bearing first. Until then those organs cannot be patched at all
F1162500014supervisor-READYMM-14 make REPRODUCIBILITY a standing published gauge (seq1052): CURRENT/total from nx_stale_check is a first-class fleet-health metric and nothing tracked it before 2026-07-26. Currently 175 permil on a 40/435 sample. Wire a periodic sweep + publish beside the other health numbers
F1163090952the DETECTOR is not the FIX -- nx_cc must still reject an unresolved identifier at compile time (seq1012 stays open). ALSO seq1051: nx_lang_lint already existed with L1-L4 incl the empty-string landmine and is wired into NO build path, so a documented landmine with a working detector still lands in new code -- wire it as a PRE step in nx_sov_build_run. L6 covers CONSTANT-STYLE names only; lowercase typos NOT covered, stated not hiddenBLOCKED Tcompiler
F116455059librarianpapersREADYAS-7 publish the RT-004 manuscript + its .conf source as served artifacts: the content-upload allowlist rejects sites/nishifamily/rt004.md ('target not allowlisted') and nx_ship/nx_treepack carries ONLY .nx so it cannot move them either. Needs an allowlisted content path (md_content_pfx/md_content_ext_ok) for research artifacts, or nx_fs_write staging + nx_site_publish per file. Until then the published paper cites a repo path it cannot serve.
F1165500014referee-READYMM-15 give nx_stale_check a distinct BUILD-ENV-FAIL verdict (seq1066): BUILDFAIL currently conflates 'source cannot compile' with 'the machine was too busy to compile it', so a batch sweep reports its own load as source defects. Separate the two, then re-run the fleet sweep serially with pacing on a quiet box to get the true rate
F1166900014engineer-gamegame-vnREADYVN-R1 chroma-key sprite gen: 5080 gen lane emits characters on a flat KEY background -> nx_vnsprite keys+composes REAL art (VNCCS stage-4 analog: sprite-ready via color-key instead of alpha PNG)
F1167800014engineer-gamegame-vnBLOCKED F1166VN-R2 expression sets off ONE anchor sheet (VNCCS pattern): img2img low-denoise emotion variants, same identity -> smile/frown/blush/surprise sprite set per character (deps F1166)
F1168800014engineer-gamegame-vnBLOCKED F1167VN-R3 character LoRA identity lock (elara F1010 lesson: prompts alone = 0 percent identity): per-character LoRA on the 5080 -> cross-scene consistency, the VNCCS stage-5 dataset scaffold feeds training (deps F1167)
F1169900014engineer-gamegame-vnBLOCKED F1167VN-R4 SAKURA-CLASS SCENE SHIPPED: bg CG + 2 gen-lane characters + expression swaps + textbox composed by nx_vnsprite, LIVE on /world/vn; MEASURED vs real Sakura screenshots via the groundtruth harness + nx_vn_critic, not self-graded (deps F1167)
F1170800014engineer-gamegame-vnBLOCKED F1168VN-R5 BADIK-CLASS LOOK: two candidate paths, MEASURE both -- (a) badik-look LoRA on the gen lane (community proves the Daz-render aesthetic is diffusion-reachable), (b) the sovereign Daz analog: posed NXA/rigmesh characters through the GX-39-lit nx_anat_sov path; ruler = vsbench Being-a-DIK ref axes + groundtruth vs real BaDIK frames (deps F1168)
F1171900014engineer-gamegame-vnBLOCKED F1166VN-R6 nx_charsuite = HIT AND EXCEED VNCCS (operator 07-29): match its 6 stages (anchor sheet -> clone -> clothes -> expressions -> sprite cut -> LoRA scaffold) then EXCEED on 5 measured axes VNCCS lacks: (1) GATED+DETERMINISTIC per stage (fixed seed -> byte-identical set; VNCCS has zero gates), (2) identity MEASURED via facemark/refbench threshold not eyeballed, (3) STRAIGHT-TO-PLAYABLE (output composes into nx_vnsprite layers + story engine -> live scene URL; VNCCS stops at PNGs), (4) sovereign agent-invocable MCP organ, headless, evidence-logged per beat, (5) ruler-anchored vs REAL Sakura/BaDIK frames via groundtruth (deps F1166)
F1172820210engineer-browserbrowserREADYsovereign WebP L3: VP8 LOSSY still decoder (boolean coder + DCT + intra prediction + loop filter) -- photo thumbs content-negotiated as lossy webp DECLINE today (the 3rd magic-passing body on the OGRE census page). METHOD PROVEN on VP8L 2026-07-29: transcribe a stdlib-python ref from libwebp (_ops/vp8l_ref.py pattern), truth-anchor it vs a PNG twin of the same asset, then stage-bisect planes + per-code bit cursors until byte-exact; bank fixture+oracle as gate teeth (nx_webp_gate T5/T24 exemplars). Wire = the VP8 branch of nx_webp_decode.nx (container already routes it).
F1173720212engineer-browserbrowserREADYsovereign SVG rasterizer rung 1: path grammar (M/L/H/V/C/S/Q/T/A/Z) + nonzero/evenodd scanline fill + viewBox-to-layout transform -> RGB at the box size. Wikipedia pages skip ~5 .svg icons per article today (the census skip filter). KAT: an ambox SVG rasterized vs its own Wikimedia PNG thumb twin (fetch with Accept: image/png -- byte-oracle recipe proven 2026-07-29); tolerance = AA-aware band, structure exact. Entry: route .svg bodies in nx_img_bytes_to_rgb once rung 1 lands.
F1174622410engineer-browserbrowserBLOCKED F1172sovereign AVIF still decoder (codec ladder L5): AV1 intra keyframe -- OBU parse + symbol-adaptive multi-symbol arithmetic decoder + inverse transforms + CDEF/loop-restoration. Wikimedia/web thumbs increasingly negotiate AVIF; nx_img_harvest already recognises the URLs (seq719 half). Same truth-anchor method as VP8L: stdlib python ref first, byte-oracle KATs, stage bisect. BIG multi-session rung -- start AFTER VP8 lossy (F1172) since that shares the codec-gate scaffolding.
F1175620212engineer-browserbrowserREADYcodec truth-anchor harness AS AN ORGAN (nx_codec_gate_lib): generalize the proven vp8l/gif bisect method -- banked byte-oracle KATs (fixture + oracle.rgb, HARD RED on decline OR drift), real-world negative controls, stage-plane dumps, per-code bit-cursor stamps -- into one gate library every codec gate INHERITS (D001 pattern), so VP8/AVIF/SVG land with exactness teeth BY CONSTRUCTION instead of per-session scaffolding. Python reference decoders (_ops/vp8l_ref.py, _ops/gif_ref.py) stay benchmark-oracle-only per doctrine; the ORGAN side (compare, dump, verdict) is sovereign.
F1176720212engineer-browserbrowserREADYfont wrap-parity rung (the MEASURED half-lever): oracle ADVANCES + GLYPH RESCALE as ONE change. A/B 2026-07-29 (wikipedia, Chrome+Waterfox oracles agreeing): advances alone = f0 OCR 881->591/588 permil, scroll 491->475 -- pioneer glyph shapes at Liberation/Chrome advances lose legibility with no structural win. Build: scale glyph strokes horizontally to the oracle advance per char (or import Liberation OUTLINES, SIL-OFL -- web_assets/fonts/LiberationSans-Regular.ttf banked, table provenance byte-proven by _ops/font_provenance.sh), then decide via drift.ps1 OCR + TRUTHFUL scroll axes on the census pair. Toggle stays font_set_adv_oracle, default OFF until the pair wins.
F1177820212engineer-browserbrowserREADYCSS-on census lever: INLINE-BOUNDARY SPACE LOSS -- styled text runs drop the whitespace at text/link boundaries (wikipedia CSS-on renders 'benchmarkis', 'computer programa', 'applicablesoftware'; visible in _build/wiki_csson_top.png 2026-07-29). OCR word-matching pays full price: CSS-on wiki f0 = 705 permil vs 881 CSS-off while composition/height are already Chrome-like (page_h 7365/7665, sliver seq1132 class RESOLVED 0->705, measured+LOOKed). Suspect: the vec-measure/box-split walk around inline elements eats the boundary space (CSS-off path keeps it -- diff the two extract paths). Fix + rerun drift.ps1 census pair; this one defect likely closes most of the CSS-on OCR gap.
F1178500014verify O(1) frontier append landedseg-storeBLOCKED 7o1-append-verify
WR193039refereewrite-rulerREADYRuler gap #1-3 (pop=378/378): multi-detector disagreement bound + judge-alpha floor + real-person depiction gate
WR2820212engineerwrite-rulerREADYLift the 3 write PARTIALs to HAVE: function-type 4-way registry, harm-taxonomy-as-DATA, refuse-class inventory
WR3730311engineerwrite-rulerREADYPersona layer (pop=279-342): identity 6-way + opening-register 4-way + three-beat greeting + hidden-attribute persona
WR482359censuswrite-rulerBLOCKED WR1WRITE domain ADMISSION: write.matrix + .gates + the 6 ADMISSION_STANDARD checks incl a 2nd method class (debt 1785866889)
GOV1920210wardenintent-govREADYMinor-coded-persona detector: the 18yo+school construction PASSES a raw age>=18 gate (3 ruler exemplars) -- highest-severity open axis
GOV2920210wardenintent-govREADYRefuse-class inventory as DATA: the 12 ruler-measured classes wired into rg_decision as config, never code
GOV3922410wardenintent-govBLOCKED GOV1Abliterated-seat burden: when the MODEL cannot refuse, OUR gate is the only gate -- route every VLM/LLM seat through rg_decision by construction
SEAT193038engineer-gpuseat-localREADYVRAM residency broker: extend the nx_resgov regime to VRAM; admission REFUSES rather than silently shrinking (the memfloor scar)
SEAT2810113engineer-gpuseat-localREADYVerify nx_gpu_broker actually SCHEDULES before adopting it (its sibling nx_gpu_ctl turned out to be a Vast rental plane, not compute)
SEAT382359modelwrightseat-localBLOCKED SEAT1Seat model MEASURED not shopped: Qwen3.6-27B-abliterated GGUF Q3 vs 35B-A3B MoE-offload on the 5080. NVFP4 REFUTED (20.4GiB weights + non-GGUF format)
WORLD193368modelwrightworld-runtimeBLOCKED SEAT1Characters as LoRA adapters over ONE resident spine (nx_lora_pool): N characters must not mean N models
WORLD2820210architectworld-runtimeREADYSimulation LOD: on-screen characters get inference, off-screen advance by deterministic state over nx_world_state carry-forward
WORLD3822410architectworld-runtimeBLOCKED WORLD2Modality latency tiering: text resident, TTS near-interactive, image queued, video batch -- nothing blocks the world tick
VR1610113researchergen-vrREADYREFUTED-AND-PARKED: real-time generative world model in VR is compute-blocked (SwiftVR 26fps 1080p mono on a 5090 vs 90fps stereo under 20ms). Revisit ONLY on a measured order-of-magnitude change
VR2830311engineergen-vrBLOCKED F1106HYBRID VR END-STATE (the path that works): conventional render + AI mind, as VaM2+Voxta independently converged on -- ours adds the governance layer nobody has. Rides F1106 vr-runtime
VR3720212engineer-graphicsgen-vrREADY360 equirect panorama generation: the ONE generated-VR capability feasible on this hardware today (stereoscopic environment from a single still)
PUB1720212librarianresearch-pubREADYRegister the 5 research-area pages through the publishing desk: pub- registry rows + nx_site_publish page bytes, so the hub is DERIVED not hand-curated
PUB293036librarianresearch-pubREADYBACKFILL the publishing registry: MEASURED 2026-08-04 registry_rows=1, registered_on_disk=0, orphans=180, debris=40, prev_artifacts=62. Reconcile every live page into pub- with owner/section/status
PUB392356wardenresearch-pubBLOCKED PUB2ANTI-RECURRENCE (the keystone): nx_site_publish must FAIL-CLOSED on an unregistered path, so a page cannot reach the docroot without a pub- row. Sprawl becomes impossible BY CONSTRUCTION, not by discipline
PUB482359librarianresearch-pubBLOCKED PUB2Every hub and index page DERIVED from the pub- registry only (nx_pub_desk board); retire hand-maintained index pages. A generated index must be a SUPERSET of the hand page before it replaces it
PUB583586architectresearch-pubBLOCKED PUB3VERSIONED TEMPLATES + CHILD ROLLOUT (the operator ask): page templates carry v1/v2/v3 and a registry-declared parent, so a template change PROPAGATES to every child instead of forking a new orphan
PUB6810113janitorresearch-pubREADYDEBRIS RETRACTION: 40 debris + 62 prev-artifacts are LIVE-SERVED from the docroot (.bak-predense, .bak-prebeir, 5x codecs.html.nxwNNNNN, sota.html.nxw3914, .prev-codelink). A staged artifact is a loaded gun aimed at the next deploy
PUB791568wardenresearch-pubBLOCKED PUB3STANDING ORDER 2026-08-04 (operator): NOTHING publishes to nishifamily/andelinwest/any docroot outside the house. nx_site_publish takes so ONE fail-closed gate covers every site by construction -- no per-site variant, no exemption. SSOT knowledge/compare/PUBLISHING_HOUSE_STANDARD.txt
RT00593586refereeresearch-pubBLOCKED PUB3PAPER RT-005 (queued behind PUB3, NOT hand-published): 'Intent-Respect as a Measurable Property of Generative Roleplay Systems'. Full apparatus: abstract, method, n=307 ruler results, negative controls incl meter-can-fail-to-fire, limitations, threats to validity, references, reproducibility via nx_dialogue_gate 24 checks
RT00683586researcherresearch-pubBLOCKED PUB3PAPER RT-006 (queued behind PUB3): 'A Population-Ranked Capability Ruler for Writing and Roleplay Systems' -- 28 axes from arXiv:2601.14324, coverage COMPUTED 482 permil, gap queue ranked by the ruler's own class counts, liar-kill gate 8 teeth. Cite-or-refuse via nx_paper_forge + evidence ledger
RT00783586architectresearch-pubBLOCKED PUB3PAPER RT-007 (queued behind PUB3): 'Multi-Character Simulation Under a Hard VRAM Ceiling' -- systems paper: residency as the binding constraint, characters as adapters not models, simulation LOD, modality latency tiering. Measured 362.1 GB/s Q4_K kernel, 16,002 MiB free, NVFP4 20.4GiB refutation
RT00872577researcherresearch-pubBLOCKED PUB3PAPER RT-008 (queued behind PUB3): 'Why Generative World Models Cannot Yet Enter VR' -- a FEASIBILITY/NEGATIVE result in the RT-004 tradition. SwiftVR 26fps 1080p mono on a 5090 vs 90fps stereo under 20ms; stereo unsolved; hybrid render+AI-mind is the converged architecture. Explicit reopen condition
PUB8920212researcherresearch-pubREADYMAKE THE PAPERS PIPELINE REAL before pushing RT005-008 through it: knowledge/papers/ is ABSENT and rt004_paper.conf scans 0/200k files, so nx_paper_forge cannot run and the live 'regenerate byte-identical' claim on /rt004 is unbacked. Bank manuscript .conf + evidence .tsv per paper, or retract the claim. Debt 1785869402
AL1900014engineer-gamegame-artloopREADYART-LOOP BEAT (recursive substrate): run the emit-score-look cycle on the clock plane -- nx_m2d_engine sweep per beat, each cycle's variant scores + sheet APPENDED to an artloop journal (never overwritten) so every cycle baselines the next; picks re-base the board OURS; ratchet gate refuses a cycle that regresses the shipped pick
AL2900014engineer-gamegame-artloopBLOCKED AL1PERCEPTUAL JUDGE for the pick step: an organ that ranks sweep variants the way the operator's eye does (caricature-control calibrated -- the composite critic is a structure floor, NEVER the judge); human override always; closes the loop so cycles run unattended
AL3900014engineer-gamegame-artloopBLOCKED AL2RECOMBINATION (the growth mechanism): winning material/world seeds BREED via the certified nx_genome gn_breed over parameter vectors; sweep the offspring each cycle; selection = judge + operator picks -- capability grows by selection pressure over seeded generators, the infinigen doctrine made recursive
AL4800014teamgame-artloopBLOCKED AL3GENERALIZE the loop shape to every emitter: one loop organ, per-lane subjects (worldpipe palettes, VN sprites post-F1166, mh face/body lanes, gen-lane art) -- emit variants, score, judge, pick, ship, journal; the same recursive setup lifts each lane toward sota
F1179900014engineer-graphicsmhBLOCKED knowledge/library/recursive_fit_loop.mdrecursive-fit-loop ADOPTION rung: body canon vs cadaver skin oracle -- extract nx_body_proc station params to a P-vector (prove byte-identity), tune w/ symmetric two-sided objective + fit-as-data (fit_src declared), body gapmap verb, clone nx_skullfit_beat onto the clock; recipe knowledge/library/recursive_fit_loop.md
F1180900014engineer-graphicsmhBLOCKED knowledge/status/skullfit_trend.jsonlskull REPRESENTATION GROWTH: fit curve flattened ~4.5mm = 16 primitives are the binding constraint (trend knowledge/status/skullfit_trend.jsonl); grow the primitive set (teeth row, orbital rims, mastoid, real arch bridges, skull base) as SEARCHABLE parameters adjudicated by the standing skullfitbeat loop -- never hand-placed
F1181900014engineer-graphicsprocgenBLOCKED knowledge/library/recursive_fit_loop.mdPROCGEN STRUCTURE GRAMMAR keystone (operator 2026-08-09: infinigen-style trees, houdini/zbrush/blender-class): generalize repeat-along-curve (ss_teeth, shipped in nx_skullsdf as instance 1) into a composable organ -- elements along parametric curves, RECURSION (structures carrying structures = branch-on-trunk, leaf-on-branch = L-system/space-colonization tree), size/angle profiles as tunable params; first deliverable = a TREE emitted from rules, judged per nx_procgen_board axes, learning via knowledge/library/recursive_fit_loop.md
F1182800014engineer-graphicsprocgenBLOCKED knowledge/store/nx_graphics_boardPROCGEN DISPLACEMENT/DETAIL FIELDS: all surface detail is currently attempted in control vertices which no SOTA system does (measured: 3.9x tris bought +1); build noise/feature-driven displacement over base surfaces (the geometry-node op class) + make the detail judge map-aware (nx_bodybench is structurally MAP-BLIND and would score the SOTA architecture WORSE) -- ruler first, then the op
F1183900014engineer-graphicsprocgenBLOCKED knowledge/library/recursive_fit_loop.mdFULL-CIRCLE INGEST->MATCH->EXCEED pipeline (operator 2026-08-09): (1) INGEST any real input at sota into normalized nxmesh/field oracles -- audit the ingest surface first (nx_gltf2mesh, obj/stl, photogrammetry via nx_photogrambench, image lanes; the jpeg restart-marker defect blocks most real photos); (2) REGISTRATION organ, the measured missing keystone: rigid align (t+R+s, integer ICP-class) BEFORE any gap measure -- twinbench read 13065mm of pure frame offset, gapmap bbox-centring only works when frames happen to agree; (3) MATCH: gapmap-class instrument + inverse fit (recursive_fit_loop) drives the controllable generator until emission matches the ingested input; (4) EXCEED: sample the fitted parameter space past the input (variation, beauty manifold, animation) -- infinigen-style controllable emission the input cannot do. Exemplar proven end-to-end on the skull lane today
F118493035engineer-graphicsbeachREADYbeach-realism F1: garment OBJECTS not paint -- conforming garment meshes generated from the body surface + garment table (thickness, hems, straps, own material) then XPBD drape + cloth-body collision (gm_xpbd_step, st_cloth_contact, dyna_cloth); gpe_garment_draw as shipped is paint
F1185920212engineer-graphicsbeachREADYbeach-realism F2: per-REGION skin atlas + generated areola/nipple/pores in the NPC asset (gpe_skin_microsurface asset half via nx_nxa_texbake region sets + an_add NXA1 relief) and page sampling by region; the whole-body 2048 atlas clamps every relief band away
F118695059engineer-graphicsbeachREADYbeach-realism F4: leave voxels -- a mesh terrain renderer over the staged worldpipe ground (GE14a/b wc_worldpipe_bind + bake), sand material + water shader, palms as frond geometry not cubes, then the splat tier (gs_avatar_genome, g4_avatar_rig) for the cast; the Infinigen-2 beach bar (white sand, palms, resort) is unreachable inside the voxel raster
F1187920212engineer-graphicsbeachBLOCKED F1108beach-realism F3: face GEOMETRY -- parse the FACE+MORF sections already carried by ref9d.nxa (nx_nxa_morf channels WIDT..JAW), upload deltas, weight per genome bits + mobmood in the MVS; the painted landmark face retires; then the procedural skull layer (F1108/F1110) becomes the head bone
F1188930311engineer-graphicsbeachBLOCKED F1138beach-realism F5: emittable beauty within a RANGE -- population marker-range canon from PUBLISHED anthropometry (F1138 WHO manifold), correlated identity sampling (bg_identity_space) so every genome is drawn inside the band, fa_averageness_axis mean face, Chara Maker sliders (cc_slider_ui) for hand adjustment, cjc_panel as referee; the operator refpop is aggregate-only, individual likeness stays behind rp_licence_admit
F1189930311engineer-graphicsbeachEXTERNALbeach-realism F5: emittable beauty within a RANGE -- population marker-range canon from PUBLISHED anthropometry (F1138 WHO manifold), correlated identity sampling (bg_identity_space) so every genome is drawn inside the band, fa_averageness_axis mean face, Chara Maker sliders (cc_slider_ui) for hand adjustment, cjc_panel as referee; the operator refpop is aggregate-only, individual likeness stays behind rp_licence_admit
F119095059engineer-graphicsbeachREADYbeach-realism F4: leave voxels -- a mesh terrain renderer over the staged worldpipe ground (GE14a/b wc_worldpipe_bind + bake), sand material + water shader, palms as frond geometry not cubes, then the splat tier (gs_avatar_genome, g4_avatar_rig) for the cast; the Infinigen-2 beach bar (white sand, palms, resort) is unreachable inside the voxel raster
F1191930311engineer-graphicsbeachBLOCKED F1138beach-realism F5: emittable beauty within a RANGE -- population marker-range canon from PUBLISHED anthropometry (F1138 WHO manifold), correlated identity sampling (bg_identity_space) so every genome is drawn inside the band, fa_averageness_axis mean face, Chara Maker sliders (cc_slider_ui) for hand adjustment, cjc_panel as referee; the operator refpop is aggregate-only, individual likeness stays behind rp_licence_admit
F119293365engineer-graphicsbeachBLOCKED F1184beach-realism F6: hair as STRANDS not a mop -- operator: the hair clips the body and moves as one mass; today one spring chain per ribbon cluster with no body collision. Bar (TressFX/UE Groom/augmented mass-spring 2024-25): guide-strand simulation with interpolated render strands, capsule body colliders (hair-body contact = softbody st_strand_contact), a wind FIELD with per-strand phase and drag, strand-level damping bands from measured hair; hr_strand_sys + gr_strand + st_strand_contact; page owner after the garment lane
F119393368engineer-graphicsbeachEXTERNALRETIRED 2026-08-31 AS A VERBATIM DUPLICATE OF F1192 (same text, same day, same lane) -- work it on F1192, not here. Original text follows. beach-realism F6: hair as STRANDS not a mop -- operator: the hair clips the body and moves as one mass; today one spring chain per ribbon cluster with no body collision. Bar (TressFX/UE Groom/augmented mass-spring 2024-25): guide-strand simulation with interpolated render strands, capsule body colliders (hair-body contact = softbody st_strand_contact), a wind FIELD with per-strand phase and drag, strand-level damping bands from measured hair; hr_strand_sys + gr_strand + st_strand_contact; page owner after the garment lane
F119494377engineer-graphicsbeachBLOCKED F1184beach-realism F7: garments GENERATED the Infinigen-2 way (operator: 'garments should be infinigen 2 generated; we should be growing all our capabilities') -- the offset-shell garments of F1184 are the crude rung; the bar is procedural sewing-pattern garments: a pattern DSL (GarmentCode-class: panels, darts, seams, parametric variety per garment class), panels placed around the body and STITCHED, then XPBD drape with cloth-body collision (gm_xpbd_step, st_cloth_contact, gmg_drape_xpbd), fabric classes from measured swatches, per-genome fit; sources mirrored + pinned (GarmentCode/GarmentCodeData, Sewformer, Infinigen-2 asset-generation idiom); accepted only by three-party agreement
F119593695engineer-graphicsbeachBLOCKED F1192beach-realism F8: page JS is INTEROP ONLY -- the emitter's authored JS logic ratchets to zero (operator 2026-08-30: 'i dont want js powering any of this it should be emitting all the way from the first byte up via nishi lang and nishi ecosystem'); nx_jslogic_census check read RED on 2026-08-31 (logic_unknown 228377 vs baseline 70555 -- every 08-30 page lane built into the JS strings); gameengine GE21 (wc_hair_tick: guide chains simulated in nx_wasm_craft.nx) + GE22 (gpe_wasm_hair_bind: emitter copies wasm buffers only); done when the census is GREEN and logic_unknown reads 0, never re-baselined
F119693695engineer-graphicsbeachEXTERNALRETIRED 2026-08-31 AS A VERBATIM DUPLICATE OF F1195 (same text, same day, same lane) -- two rows for one contract is how one edit gets adjudicated twice; work it on F1195. Original text follows. beach-realism F8: page JS is INTEROP ONLY -- the emitter's authored JS logic ratchets to zero (operator 2026-08-30: 'i dont want js powering any of this it should be emitting all the way from the first byte up via nishi lang and nishi ecosystem'); nx_jslogic_census check read RED on 2026-08-31 (logic_unknown 228377 vs baseline 70555 -- every 08-30 page lane built into the JS strings); gameengine GE21 (wc_hair_tick: guide chains simulated in nx_wasm_craft.nx) + GE22 (gpe_wasm_hair_bind: emitter copies wasm buffers only); done when the census is GREEN and logic_unknown reads 0, never re-baselined
F1197810113engineer-compilercompilerREADYcompiler EMI / metamorphic variant testing -- the highest-value correctness instrument for a SOLE-IMPLEMENTATION language, because it needs no second compiler, no reference implementation and no specification: profile a program, prune the statements the profile shows were never executed, recompile the variant, require byte-equal output. The EMI project found 1602 bugs in GCC+LLVM (147 confirmed in eleven months). BUILD THE VARIANT GENERATOR AND COMPOSE THE INCUMBENT COMPARATOR -- measured incumbents, never re-implemented: nx_cc_equiv_gate eq_build already does build -> run -> raw wait4 status + stdout byte-equal over a 10-row corpus, nx_behaveprobe is the estate output comparator, and nx_shapeprobe is a SHAPE-axis generator that invents fresh programs carrying no equivalence relation. BRIEFED PREMISE CORRECTED BY MEASUREMENT: the incumbent gate varies the COMPILER (baseline vs challenger) and holds the source FIXED, the opposite axis to EMI, so it cannot catch a bug both binaries share; EMI holds the compiler fixed and varies the SOURCE and needs no challenger build at all. The missing half is exactly two things -- a profile-and-prune variant generator over real runtime modules, and the semantic-equivalence EXPECTATION that makes any output difference a miscompile
F1198800010engineer-compilercompilerREADYcompiler BOOTSTRAP FIXPOINT -- cc2 and cc3 must be BYTE-IDENTICAL. GCC treats a mismatch as a potentially serious bug. Cost is approximately ONE BUILD, NishiLang self-hosts, and the estate already applies this exact discipline elsewhere (nx_magic proves a refactor neutral by byte-identical rebuild). MEASURED GAP, not a guess: nx_cc_equiv_gate's self-host stage stops at gen2 compiles a probe module whose binary runs exit 0, and nx_bless_compiler bc_equiv_green byte-compares the CORPUS artifacts -- neither ever hashes gen2 against gen3, so the fixpoint is untested today. Deliverable: hash the two compiler binaries and refuse on any difference, wired into the bless path that already gates install, with the differing byte offset named rather than a bare RED. HONEST LIMIT RECORDED IN THE ROW: this licenses nx_cc IS A FIXPOINT and NEVER nx_cc IS CORRECT -- a miscompilation stable under recompilation survives it -- so a green here must never be published as a correctness claim
F1199800014engineer-compilercompilerEXTERNALcompiler BOOTSTRAP FIXPOINT -- cc2 and cc3 must be BYTE-IDENTICAL. GCC treats a mismatch as a potentially serious bug. Cost is approximately ONE BUILD, NishiLang self-hosts, and the estate already applies this exact discipline elsewhere (nx_magic proves a refactor neutral by byte-identical rebuild). MEASURED GAP, not a guess: nx_cc_equiv_gate's self-host stage stops at gen2 compiles a probe module whose binary runs exit 0, and nx_bless_compiler bc_equiv_green byte-compares the CORPUS artifacts -- neither ever hashes gen2 against gen3, so the fixpoint is untested today. Deliverable: hash the two compiler binaries and refuse on any difference, wired into the bless path that already gates install, naming the differing byte offset rather than emitting a bare RED. HONEST LIMIT RECORDED IN THE ROW: this licenses nx_cc IS A FIXPOINT and NEVER nx_cc IS CORRECT -- a miscompilation stable under recompilation survives it -- so a green here must never be published as a correctness claim
F1200720212engineer-compilercompilerREADYcompiler DIFFERENTIAL TESTING ACROSS OPTIMISATION CONFIGURATIONS -- the same source built every way, outputs must agree; no second implementation needed, which is exactly why it is affordable for a sole-implementation language. BRIEFED AS every -O setting AND CORRECTED BY MEASUREMENT: nx_cc HAS NO -O LEVELS. The pipeline is a single About-O1-class chain (const-fold SCCP mem2reg GVN CSE LICM DSE DCE, runtime/nx_opt.nx) with LICM correctness-gated, and the flags nx_cc does take are DECLARED MODES (--ownership --ptrprov --chkarith --optenforce --ir). So the real subject is the PASS SET plus the declared modes, and the FIRST deliverable is the pass-selection flag itself -- the differential is unrunnable until the pipeline can be varied at all. Then every configuration of the same source must produce byte-equal stdout AND equal wait4 status, composing nx_cc_equiv_gate eq_build as the runner rather than adding a second one. Highest-value configuration to test first is LICM armed vs disarmed, because the correctness gate on it is the estate's own admission that this pass has already shipped a silent miscompile once
F1201720212engineer-compilercompilerREADYcompiler DIAGNOSTIC QUALITY SCORE tracked per commit -- the field's genuine hole: fifty years and 300-plus references produced GUIDELINES, not a benchmark, and the 2026 benchmarks (ComBench, CCrepairBench) measure automated REPAIR of compiler errors, not message QUALITY. DExTer proved inside LLVM that a qualitative compiler axis becomes a tracked per-commit NUMBER the moment you drive the real consumer and score the transcript (approx 1.0 at -O0 vs 0.28 at -O2), so the method is precedented even though the metric is not. Score four things per diagnostic: does it name file/line/column, does it name the RULE violated, does it propose a FIX, and does it avoid CASCADE (is the first error reported the real one). COMPOSE THE INCUMBENT, do not build a second driver: nx_langdiag_gate already forks the LIVE _offc/nx_cc_sovereign.elf over a probe corpus and asserts the RIGHT refusal (exit code AND the rule's distinctive slug), so the corpus, the fork mechanics and the subject all exist -- what is absent is a SCORE rather than a pass/fail, and a number that moves per commit. MUST CARRY A NEGATIVE CONTROL: a deliberately bad message that MUST score low, or it is a detector that has only ever passed. Report the score with its denominator, and an unscoreable probe ABSTAINS rather than counting as zero
F1202640410engineer-compilercompilerBLOCKED F1198compiler DIVERSE DOUBLE-COMPILING (DDC) of the self-hosted compiler -- Wheeler's defence against the trusting-trust attack, and the strongest available technical claim on this board: self-hosting is CLAIMED by Rust, Zig, Go, Odin and Nim and PROVEN by none of them, and essentially no modern toolchain publishes a DDC result at all. NishiLang is unusually well-placed for it: self-hosted, emits its own ELFs, no libc, and a small OWNED nondeterminism surface, which is the variable that defeats DDC everywhere else. Wheeler's requirement is that the trusted second compiler be INDEPENDENT and CORRECT ENOUGH, NOT that it be good -- so a deliberately simple, slow, independently-written front end is a BOUNDED artifact rather than an open-ended research project, and that boundedness is the whole reason this is reachable here. EXPENSIVE BY CONSTRUCTION AND FILED AS THE STRONGEST AVAILABLE CLAIM, NOT AS NEXT WEEK'S WORK. Depends on the bootstrap-fixpoint rung landing first: DDC compares independently-produced compilers, so it is meaningless until the self-build is known to converge on its own
F1203720212engineer-compilercompilerREADYcompiler SPEC SURFACE COVERAGE -- the DENOMINATOR question the clause ratchet structurally cannot ask. BRIEFED AS AN UNMET CONFORMANCE GAP AND REFUTED BY MEASUREMENT: requirement-level spec-clause traceability is ALREADY SHIPPED here since 2026-08-25 -- knowledge/specs/nishilang.spec carries 33 clauses AS DATA, nx_langspec spec_conformance_run runs 35 cases each CITING the clause ids it exercises, traceability resolves MECHANICALLY rather than by assertion, and the published production run over the SHIPPED compiler reads 33 of 33 clauses exercised at 1000 permil against a ratchet raised from 800, unresolved citations 0. That already exceeds the open field, which publishes pass-COUNTS instead (test262 reports files passed, ACATS reports tests passed, and neither publishes which clauses have ZERO tests) and matches a COMMERCIAL product feature sold into ISO 26262 and DO-178C. THE RESIDUAL IS THE DENOMINATOR: coverage is two questions and only one is answered -- 1000 permil of 33 clauses says NOTHING about what fraction of the LANGUAGE those 33 clauses define, so any construct absent from the spec is invisible to the ratchet and reads as covered. Deliverable: measure spec surface against the language the compiler actually accepts (the nx_parse construct set and the nx_lex_kinds token set as the population), publish the UNCOVERED CONSTRUCTS BY NAME as a worklist because a count without a worklist is not actionable, and ratchet that second number too. THE STRUCTURAL LAW STANDS EVEN THOUGH ITS PREMISE DID NOT: you cannot claim CONFORMANCE without a specification to conform to, and a compiler that is its own specification can only ever claim CONSISTENCY and ROBUSTNESS, never CORRECTNESS -- which is precisely why the 2026-08-25 spec was the right rung, and why its SURFACE rather than its clause count is now the binding constraint
F120481895refereeautoBLOCKED F903compare PER-CELL EVIDENCE GRADES -- stamp every matrix cell with its evidence rung and REFUSE to render an ungraded cell. The field has evidence-grade vocabulary for PAPERS (ACM badging: Available, Functional, Reusable, Reproduced, Replicated) and disclosure rules for exactly ONE benchmark suite (SPEC), and NOTHING for a project's public self-comparison page: every language comparison table in the field renders a checkbox claim and a measured benchmark at IDENTICAL VISUAL WEIGHT, so a reader has no way to tell how much to believe any given cell. Grading them would make /compare the only comparison surface in the field that tells a reader how much to believe each claim. COMPOSE THE INCUMBENTS, never add a second ruler: nx_swcompare_evidence ALREADY opens the named organ source per claimed-present axis and captures the real source line plus occurrence count (per-axis GROUNDING, ungrounded is RED, with a bogus-symbol neg-control) and ALREADY forks each domain .gates for executed proof under a quorum from nx_evidence_verdict; nx_matrix_sym_lib is THE ONE READER of a row's symbol field. What is genuinely absent is the GRADE: grounding is BINARY and executed proof is DOMAIN-level, so a cell backed by a running gate and a cell backed only by a symbol existing in a file render identically today. It also composes the incumbent refs plane, which already pins the citation behind a claim. LANDS AS EMITTER CAPABILITY PLUS A DATA FIELD per the operator directive of 2026-08-31, NEVER as a side analysis; done when an ungraded cell CANNOT be rendered and the rung is visible to the reader on the published page
F1205820212researchercompareevREADYRESEARCH grounding cannot see wiring: a claim grounded on a zero-caller function is BUILT+UNWIRED arriving on a published board with a source line beside it as proof; join evidence grounding to nx_unwired (7132 sources 27678 funcs 398 unwired). No incumbent joins them
F1206810113conductorcompareevREADYre-stamp the fleet to profile v2: 92 of 95 stamps are pre-profile so the PM compare block correctly abstains on them. nx_evidence_beat runs maxwork=3 per beat; the gap profile is populated for only 3 boards today
F1207820212refereecompareevREADY42 of 96 compare boards carry no .gates map (54 measured, corpus_complete=1) so their capability marks rest on nothing executing. Mostly DECLARING gates that already run - engineering, no research, no operator decision
F1208710113engineercompareevREADYevidence profile machine-readable half: the HTML band ships on all 95 matrix boards, the api.json evidence object does not. Name the emitting organ first (nx_compare_unified writes coverage and band), do not assume nx_swcompare_matrix
F1209730311researchercompareevREADYRESEARCH symbol quality is unmeasured and load-bearing: one symbol can ground N unrelated axes, and smallos grounds a Linux-reference bar on sym=UNMEASURED x4522. Presence of one symbol cannot discriminate N claims; needs a metric that does not exist
F1210720212refereecompareevREADYRESEARCH the vacuous-gate class is reported-not-convicted on purpose: a gate with real teeth and a broken counter is indistinguishable from a gate with no teeth at all. Calibrate a fleet-wide bar before any verdict arithmetic changes
F1211810113operatorcompareevEXTERNALADJUDICATION the two-tree split: the stamp now REPORTS mroot and mbytes so divergence is visible on 95 pages, but flipping ct_compare_read PRIMARY changes 45 live verdicts. Coordinator or operator decision, never a side effect of a fix
F1212640410researchercomputationalREADYRESEARCH AUTOFORMALIZATION BEYOND PEANO ARITHMETIC: turn an informal natural-language mathematical claim into formal terms plus a kernel-checkable derivation, for GENERAL statements rather than closed-form integer arithmetic. The organ's own header names this the computational Frontier Radar number-one OPPORTUNITY (momentum 33, researcher-measured 2025/26) and records rungs 1-4 CLOSED -- but all four are arithmetic over a Peano encoding, so the closed work does not generalise and the leap has no known method. FILED AS RESEARCH SO IT IS NEVER SCHEDULED AS ORDINARY ENGINEERING, which is how a lane sinks. UNBLOCK CONDITION (concrete, not judgement): a held-out set of informal claims OUTSIDE Peano arithmetic carrying independently-authored formal targets, plus a pre-declared accept rule stating what fraction must kernel-check, written in source BEFORE any run per the pre-declare law. Until that corpus and that rule exist the axis cannot be honestly scored and must stay unscheduled
F1213830311researcherautodevREADYRESEARCH RETRIEVAL-BACKED FAULT LOCALISATION MEASURED ON HELD-OUT TASKS (autodev af_localize_index x0). The ingredients are GROUNDED and composed nowhere -- ci_query_top1 code index plus fc_alert_score failure classes plus gate stderr -- and composing them is easy engineering; what is unknown is a localisation that GENERALISES to unseen tasks, which is the live autonomous-software-engineering frontier. THE HELD-OUT REQUIREMENT IS WHAT MAKES IT RESEARCH RATHER THAN TUNING: hit@1 measured on the same tasks that selected the method is the vacuous-test defect raised to the level of the optimiser. ADJACENT AND DELIBERATELY NOT DUPLICATED: F1051 nx_bugforge mines instances and F1059 nx_fixflywheel does retrieval few-shots for REPAIR -- both act AFTER the fault is located, and nothing in the estate answers WHICH FILE AND WHICH LINE. UNBLOCK CONDITION: a held-out split of real estate defects whose true location comes from the FIX COMMIT and never from the run being scored, plus a pre-declared hit@1 bar and a negative control input that must NOT localise
F1214740410researcherautodevREADYRESEARCH HARNESS SELF-EVOLUTION SCORED WITHOUT OVERFITTING THE SPLIT THAT SELECTED IT (autodev ev_harness_search x0). The board already pre-declares the right guardrails -- proposals only, nothing auto-admits, matched budget against a test-time-scaling baseline -- so the GOVERNANCE half is settled and is not the gap. The open question is how to search harness space at all without the search laundering itself green on its own selection split: the vacuous-test problem raised to the optimiser, where the thing being fooled is the ACCEPT RULE rather than a tooth, and no published method survives a matched-budget comparison at this scale. UNBLOCK CONDITION: a THREE-way split -- search, select, and a final held-out written once and never re-read -- plus a matched-compute baseline given exactly the budget the search consumed, both declared in source before the first search step, so a win cannot be produced by spending more
F1215830311researchercharsimREADYRESEARCH LONG-HORIZON AGENT CONSISTENCY -- ONE CAUSE COLLAPSED FROM TWO BOARDS (charsim eb_longrun x0 and npcai dl_commitment x0 are the same unknown reported once per domain): how to detect personality drift, SAFETY EROSION, and broken narrative commitments over conversations far LONGER than any evaluation window. No accepted benchmark exists for either half, and commitment survival -- did the character keep what it said it would do over 20 to 100 turns -- is not a standard NLP metric. THIS IS THE ESTATE'S OWN SAMPLING-WINDOW LAW AT ITS WORST: anything slower than the window is invisible BY CONSTRUCTION, and here the window IS the whole evaluation, so a green means UNEXAMINED rather than healthy. ADJACENT AND DELIBERATELY NOT DUPLICATED: F1057 names long-horizon identity drift for VIDEO avatars (a visual subject, metrics declared inadequate) and F1056 is semantic-salience MEMORY -- neither measures behavioural consistency. UNBLOCK CONDITION: a conversation corpus longer than the evaluation window with commitments LABELLED at the turn they are made, a bar that REFUSES below threshold rather than reporting a score, and a negative control transcript that must fail it
F121685059researchergraphicsREADYRESEARCH IDENTITY-CONTROLLED GAUSSIAN SPLAT AVATAR GENERATED FROM A PARAMETRIC GENOME -- one cause on two boards (charsim gs_avatar_genome x0 and graphics gs_avatar_genome x0, which the graphics board itself labels THE LEAP). Every published splat pipeline is PHOTOGRAMMETRIC: splats are FIT from captures, and even the parametric-binding work that exists fits from a capture first and then binds to a face model. DRIVING a splat representation FROM a parameter vector, so that changing a gene changes the splat coherently while it stays riggable, has no known method. FILED BECAUSE THE ESTATE IS ALREADY SCHEDULING IT AS ENGINEERING: F1186 and F1190 both name gs_avatar_genome as a later step of a mesh-terrain rung, which prices an unsolved problem as a build item and is exactly the failure this row exists to prevent. UNBLOCK CONDITION: declare whether the target is GENERATE-from-genome or FIT-then-bind, because only the second is reachable with published method today; if GENERATE, the first deliverable is an identity metric over splats against which a capture-free sample can be scored at all
F1217830311researcherfailmodesREADYRESEARCH DETECTING A METASTABLE FAILURE STATE IN PRODUCTION -- a system whose SUSTAINING FEEDBACK LOOP HAS OUTLIVED ITS TRIGGER (failmodes fc_sustain_probe x0). The estate has LIVED this exact class: a wedge-kill supervisor whose every kill discarded a partially-completed startup and re-queued a fresh instance behind the same lock, converting a slow start into an UNRECOVERABLE one, while a direct probe of the same port returned 200 throughout. Distinguishing STILL RECOVERING from TRAPPED IN A SELF-SUSTAINING LOOP, from outside the system, is open -- the field describes the class and ships no detector. ADJACENT AND DELIBERATELY NOT DUPLICATED: F750 nx_govern_sweep detects respawn-without-backoff and F751 adds the circuit-breaker; both are the KNOWN remedy for ONE named instance, and neither can decide whether an arbitrary loop is self-sustaining. UNBLOCK CONDITION: a signal that separates the two states WITHOUT removing the load -- the candidate worth testing first is whether reducing offered load reduces the sustaining rate -- plus a bite fixture that reaches the metastable condition and is ASSERTED to have reached it before any verdict is read
F1218740410researcherfailmodesREADYRESEARCH SILENT-DATA-CORRUPTION SCREEN ON A SINGLE NODE -- a component returning a WRONG ANSWER rather than failing (failmodes fc_sdc_screen x0). Every published approach, including the Meta and Google fleet-scale mercurial-core work, depends on REDUNDANCY ACROSS MANY MACHINES to vote, so the method does not transfer to one box and cannot simply be adopted. The estate's only current answer is its own brute-force-oracle discipline -- keep the naive implementation and cross-check against it -- which is sound per computation and does not scale to a screen. DIRECTLY RELEVANT: an SDC is invisible to every availability proxy we run, the same blind spot as the capability-loss class. UNBLOCK CONDITION: name what plays the role of the SECOND VOTER on a single node -- the candidates are a differently-compiled twin of the same kernel, a recomputation at a different time under different thermal load, or an algebraic self-check the correct answer must satisfy -- and then PROVE the chosen voter fails INDEPENDENTLY of the subject, which is the property fleet redundancy supplies for free and a single node does not
F1219830311refereeinstrumentREADYRESEARCH NON-VACUITY PROOF FOR PERCEPTUAL GATES -- one unknown collapsed from three boards (instrument _ABSENT_:ev_llm_judge G-Eval class; mangagen, whose own .gates file states generated-image QUALITY has no gate here and needs a VLM judge panel with a pre-declared threshold rather than a unit gate; and the npcai idle-plausibility half of gb_npc_bench). TWO HALVES OF ONE METHOD, BOTH ABSENT: (a) how to BITE-PROVE A JUDGE whose output is a SCORE rather than an exit code, and publish its false-positive rate on real data before its findings are believed; (b) what the LEGAL MUTATION SPACE of a RENDERED subject even is, so a harness that rebuilds the subject knows a mutant is REACHABLE rather than cosmetic. MEASURED: mangagen ships four end-to-end gates that fork promoted elfs and read PNGs back through our own decoder, and a bite sweep has NEVER reached them (bite_nx_manga star matches=0, corpus_complete=1). ADJACENT AND DELIBERATELY NOT DUPLICATED: F905 validate-the-validator and RIGOR1 dynamic gate-rigor both mutate SOURCE TOKENS of code gates; F911 BUILDS a perceptual judge but does not prove one; F1070 hardens artifact EXISTENCE teeth. UNBLOCK CONDITION: a pixel-level or geometry-level mutation whose effect on the rendered subject is DECLARED IN ADVANCE, plus a control set of real subjects a correct judge must NOT flag, so the false-positive rate is measured against controls and never against planted fixtures
F1220830311researchertrafficsafetyREADYRESEARCH ROLLBACK TRIGGERED ON MEASURED CAPABILITY LOSS RATHER THAN ON AVAILABILITY PROXIES (trafficsafety ts_rollback_on_caploss x0). Every automated rollback in the field fires on error rate, latency, crash rate or CPU, and Google's own canary service states plainly that it is not trying to establish absolute health. A BUILD THAT RETURNS 200 TO EVERYTHING WHILE HAVING SILENTLY STOPPED DOING SOMETHING IT USED TO DO IS INVISIBLE TO ALL OF THEM. THE BLOCKER IS THE INSTRUMENT, NOT THE PLUMBING: our only capability-loss ruler nx_contentdiff compares STRING RUNS, and 140787 bytes were measured vanishing while it reported lost_from_live=3, all three binary padding fossils -- so a code-only delta is invisible to the guard BY CONSTRUCTION. ADJACENT AND ALREADY SOLVED NARROWLY, DELIBERATELY NOT DUPLICATED: F804 auto-rollback on a route-contract regression and F805 compiled-in shrink enforcement both work precisely because routes and tools are ENUMERABLE; the general case has no enumeration. UNBLOCK CONDITION: a capability inventory a binary can be interrogated against that is not a string proxy -- the candidate worth building first is the declared verb or entry-point set exercised under a recorded workload and compared before and after -- carrying a control that proves the measure does NOT fire on a byte-identical rebuild
F1221840410researchertrafficsafetyREADYRESEARCH SOCKET HANDOFF BETWEEN PEER PROCESSES THAT ALSO CONTEND ON AN EXCLUSIVE NON-SOCKET RESOURCE -- no published handoff model contemplates it. nginx, HAProxy and Envoy pass the descriptor, systemd inherits it, and ALL of them assume the only contended object is the listener. OUR PRODUCTION CASE IS THE PEER CASE, measured from daemons.reg: nx_daemon_supervisor and nx_hostctl FORK the daemons, so the socket owner can never be the new generation's parent and FD inheritance is structurally unavailable -- SCM_RIGHTS answers that half and is now exercised. THE UNPIONEERED HALF IS THE REGISTRY FLOCK: the incoming instance takes a process-lifetime lock the outgoing instance still holds, which is what produced the measured eight-minute sovereign-tool blackout where the new process slept in locks_lock_inode_wait and NEVER REACHED ITS LISTEN, making the handoff mechanism unreachable by construction. The invariant to preserve is that no instant has zero processes holding the listener, and no published model states that invariant for a SECOND exclusive resource. UNBLOCK CONDITION: decide whether the registry lock can be made HANDOFF-SHAPED and passed alongside the socket, or must be made NON-EXCLUSIVE before the handoff can begin -- the two answers need different mechanisms and the choice is not derivable from any published design
F1222940410wardensynthroomREADYRESEARCH PROVING A CONTINUOUS MORPH AND PROPORTION SPACE CANNOT EXPRESS A MINOR-APPEARING BODY (synthroom sr_agefloor x0, on a board that declares an adult-content surface with 14 grounded capability axes). THIS IS A PROVABLE-IMPOSSIBILITY CLAIM OVER A CONTINUOUS PARAMETER SPACE -- not a filter, not a classifier, not a threshold -- and no amount of sampling establishes it. THE ESTATE HAS NO METHOD FOR IT: our bite methodology proves a gate FIRES on planted bad input, which is the opposite quantifier, and nothing here can show that NO POINT IN A SPACE SATISFIES A PREDICATE. ADJACENT AND DELIBERATELY NOT DUPLICATED: GOV1 is the minor-coded-PERSONA detector over TEXT (the 18-plus-school construction that passes a raw age gate) -- a different subject and a detector rather than a proof. SAFETY-CRITICAL: this should gate the whole board, and until it is answered the axis must read UNPROVEN rather than closed, never GREEN, because an unproven safety claim on a shipped adult surface is the most expensive class of wrong. UNBLOCK CONDITION: constrain the space so the claim becomes decidable -- the two candidates worth evaluating are interval arithmetic over a bounded parameter box, and a monotone parameterisation where the extreme corner bounds the whole set -- and declare which was chosen in the axis itself
F1223830311researcherpmdashREADYRESEARCH ATTRIBUTING A RUN OF GENERIC TOOL FRAMES TO THE NAMED CAPABILITY BEING PERFORMED -- CLAUDE.md names this as the missing primitive and the pmdash ranker's absent terms depend on it. A hand-authored ritual currently appears as anonymous nx_fs_write frames attributed to the FILE WRITER rather than to the capability. nx_actlog al_mine_json already mines the n-grams; the N-GRAM-TO-CAPABILITY-NAME MAPPING IS THE HOLE, and it is research rather than engineering because the label set does not exist and cannot be bootstrapped without the very attribution being built. CONSEQUENCE MEASURED, NOT ASSERTED: seat-turns-removed-by-building-X is unmeasurable for every X not yet built -- exactly the class the frontier lives in -- so the board structurally cannot price its own most valuable work, and cost-per-episode spreads 1.63x against frequency spreading 15.1x, letting frequency dominate and score every unbuilt capability at zero. ADJACENT AND DELIBERATELY NOT DUPLICATED: F966 upgrades sequential patterns and F967 discovers a directly-follows process MODEL -- a process model is not a capability NAME -- while F969 counts classes without naming them. UNBLOCK CONDITION: a seed label set produced by a source INDEPENDENT of the miner, the candidate being the closing frontier or debt row a session actually closed (it names the capability in prose and is timestamped against the same frames), plus a held-out session the mapping has never seen
F1224940410refereecompareevREADYRESEARCH A REFEREE CLASS FOR A NON-PERCEPTUAL SUBJECT -- the binding constraint on the estate three-party completion law, measured across the whole fleet 2026-09-01. nx_accept referee writes a row ONLY as SUBJECT PNG TIER SEAT: it runs the cjc_panel percept lens on a capture and scores it against an admitted per-tier floor, and a lens that cannot see writes no row. So a REFEREE verdict is obtainable only for a subject that renders as an image. MEASURED, NOT ASSERTED: exactly 5 acceptance watch contracts exist, on 3 of 95 boards -- charsim ga_accept_face, graphics ga_accept_hair and ga_accept_skin, gameengine ga_accept_garments and ga_accept_locomotion (grep over buildroot knowledge compare matrix, corpus_complete=1) -- and all five are visual. The other 92 boards (deploy, loadgov, trafficsafety, search, compiler, every infrastructure domain) have no image and therefore cannot obtain a referee row by ANY path that exists today, so under the estate own law they can never reach COMPLETE for a reason that has nothing to do with whether the work is finished. THE ESTATE ALREADY STATES THE HOLE IN ITS OWN WORDS: the devmgmt acm_emit row says the mechanism own acceptance needs a non-visual referee class that the percept-only referee plane cannot yet express, stated rather than faked with a cell that could never flip. ADJACENT AND DELIBERATELY NOT DUPLICATED: F1219 asks how to BITE-PROVE a perceptual judge whose output is a score -- that is the COMPLEMENT, a method for validating the referee we already have; this row asks what a referee even IS when there is nothing to look at. A gate verdict is not the answer, because the gate is written by the same seat that wrote the subject, which is exactly the independence the three-party law exists to supply. UNBLOCK CONDITION: name a referee whose verdict is INDEPENDENT of the seat under judgement for a non-rendering subject -- the two candidates worth evaluating first are an adversarial re-derivation by a second seat that never saw the first implementation and must reach the same number from the raw artifact, and a pre-registered accept rule written into the plane BEFORE the work starts so the referee only checks arithmetic it did not choose -- plus a negative control subject a correct referee must REFUSE, or the class is a rubber stamp with extra steps.
F123193038engineermediaingestREADYSEMANTIC RETRIEVAL TIER (NX_IT_KIND_SEMANTIC) -- nx_imgsearch reserves a fifth tier slot and publishes its own gap as semantic-lookalike permille=-1 UNMEASURED, with the build order pre-declared in the organ: (1) a ruler from labelled same-concept pairs, (2) measure the EXISTING descriptors as the honest baseline, (3) only then a quantised learned image embedding. Step 1 decision core SHIPPED 2026-09-01 (nx_semruler_lib + nx_semruler_gate 20/20 bite-proven). THIS ROW IS STEP 3 AND IT IS THE FRONTIER HALF: a learned image encoder quantised through the estate nofloat int8 stack, so the engine can answer same-KIND-of-subject rather than same-PHOTOGRAPH. WARNING BANKED IN THE ORGAN ITSELF: nx_embedding.nx is the LLM token-ID lookup, NOT an image encoder -- there is no image embedding in the estate today. ACCEPT RULE, PRE-DECLARED BEFORE ANY RUN: the new tier must EXCEED the existing four-tier fusion on the same labelled pair set, measured through nx_semruler admissibility (near-duplicate pairs REFUSED so the copy tiers cannot score the semantic class), or it does not get wired.
F123283368engineermediaingestBLOCKED F1231ANTHROPOMETRIC RATIO RETRIEVAL -- retrieve by BODY PROPORTION rather than by photograph. The estate already carries nx_body_measurement_extract (silhouette + landmarks -> BodyDimensions in mm) and nx_garment_fit_match, and the extractor own header NAMES the missing brick: nx_body_landmark_detect, given a silhouette emit the landmark_y array automatically, currently caller-supplied and therefore not automatable. TWO FINDINGS THAT CHANGE THE SHAPE OF THIS WORK, both read from source 2026-09-01: (1) THE SCALE-REFERENCE PROBLEM DISSOLVES. The extractor wants a credit card or A4 sheet in frame to fix px-to-mm, which no scraped photo has -- but matching needs RATIOS, and a ratio cancels the px-to-mm factor entirely, so waist-hip, waist-height, shoulder-waist, leg-torso are all derivable TODAY from the landmarks it already emits. Single-image body shape is scale-ambiguous in principle, so the ratio formulation is FORCED rather than chosen. (2) THE UPSTREAM DEPENDENCY IS A PERSON MATTE, not more geometry: nx_image_gray plus a threshold will not survive a real background, so this needs a matting model hosted through the nofloat int8 stack. ACCEPT RULE: silhouette IoU against a hand-labelled set before any ratio is published, and the ratio vector must separate a held-out same-subject pair from a different-subject pair better than the existing four-tier fusion does, or it stays unwired.
F1233920212instrumentcompareevREADYAN ABSTAINING AXIS NEEDS AN AXIS-BLIND ALARM -- the estate law is abstain-never-acquit, and it is only HALF a mechanism: abstention is safe for the verdict and INVISIBLE to the operator, so a detector that goes blind produces no alarm anywhere, BY DESIGN. LIVE INSTANCE MEASURED 2026-09-01, and it is on the estate own headline honesty gate: nx_sota_status discovers domains from two compare trees but reads ROWS from one (45 .matrix in knowledge/compare vs 96 in buildroot), so roughly 51 of 101 boards publish cov=0 exceed=0 gaps=0 on POPULATED matrices. The tell is that all three counters are zero together -- mediaingest reported gaps=0 while its matrix carries four _ABSENT_ rows, which proves the row reader returned ZERO ROWS rather than that coverage is zero. CONSEQUENCE: the published moat (171 unique capabilities) and frontier gap count (338) are computed over 45 of 96 boards and understate by roughly half, on the one organ whose entire purpose is to refuse over-claiming, and it fails in the direction nobody audits because a silent zero reads as an unstarted board rather than an unread one. SECOND WITNESS THE SAME DAY: mediaingest now reads MECH-OK WITH cov=0 -- the classifier can see its .gates in the bare tree while the row reader cannot see its .matrix in buildroot, so one domain is simultaneously graded and uncounted. THE FRONTIER QUESTION IS NOT THIS ONE BUG: it is what a general axis-blind alarm looks like, given that the safest verdict is also the least informative one and every abstaining instrument in the estate has this shape. Filed as debt 1788268797 for the instance; this row is the CLASS.
F1234810113engineer-engineengineREADYWasm-vs-native bit equality is STRUCTURAL here: nx_wasm_craft has zero floating point (proven), and the wasm spec's only arithmetic nondeterminism is FP
F1235720212engineer-engineengineREADYGround truth emitted by the SHIPPING runtime: prove nx_world_snap gt reproduces every channel bit-for-bit from a seed
F1236710113engineer-engineengineREADYArt style is a spec word in the served engine (P_STYLE idx 49), doctrine adopted from nx_voxel_world after 7 gate-only importers
F1237830311engineer-engineengineREADYTHE REPRESENTATION AXIS IS ONE: three renderer BACKENDS bind one voxel wire format (vox_off to O_VOX). A gap, not a frontier -- rivals ship many representations