Nishi Development Record

Nishi vs the field — every Nishi cell is measured against real organ source at emit time; each gap names the watch contract that will close it.

The development record itself -- every tool call, measurement, refusal and resume point captured, mined and proven -- versus a field that versions outcomes (git), operations (jj), or delivery exhaust (DORA platforms). pmdash selects the work; devmgmt records how it was done and turns that record into intelligence. Rows scored from the Aug-2026 four-lane research pass; grounds in devmgmt.refs.

Layer 1 · Executive

Where we are. Shipped and measured: the transcript parser lane (ch_frames, ch_delta, ch_ingest, ch_gauge_json over 869 files and 2.39 GB), action-journal workflow mining (al_mine_json), trajectory scanning (tj_core), workstream lifecycle (wm_mine_json), DORA from sovereign planes (d_band_deploy), gate-outcome mining (hm_mine_json), sovereign VCS push (sgp_walk, drift-AHEAD -- never promote), proof-carrying change (cd_show_lost), mutation-proven gates (gb_try), the drift census (ps_judge), the one-call ship loop with its durable journal (os_jrnl), content CAS on the shared tree (fsw_expect_val) and seat metering (st_boot). Open: the record is laptop-local and unmineable from the estate, and crash-resume mapping, lost-set mining and crash-write adjudication were each done BY HAND on 2026-08-19 -- twice for the resume map. SHIPPED SINCE, 2026-08-20: DM2 st_resume_map, and DM5 ps_owner_scan with the owner-signal axis (DECLARED, IN-FLIGHT-SUSPECT, UNDECLARED, SRC-UNRESOLVED) plus a declaration surface a seat writes when it builds without promoting, gate-proven 23 teeth and bite-proven in both directions against the pre-change binary. Corrected in the same pass on a sibling seat evidence: the drift census compares against the LAST BUILD ANYBODY RAN, so a BEHIND delta measured on a reference older than its source can be exactly inverted -- REF-FRESH, REF-FOSSIL and REF-UNKNOWN now qualify every BEHIND row, promote_considerable requires UNDECLARED and REF-FRESH together, and the unqualified promote-the-fix label is gone. SHIPPED 2026-08-20 second pass, in the ranker's order: DM3 rm_lostset, DM7 dc_adopt_delta, DM6 d_ship_join and DM8 os_doc_stage, each with an end-to-end gate that forks the DEPLOYED subject (nx_lostset_gate 26 teeth, nx_adoptdelta_gate 28, nx_shipjoin_gate 20, nx_docstage_gate 29), every one bite-proven in BOTH directions -- a mutant that fires on nothing and a mutant that fires on everything -- and every subject restored byte-identical afterwards. DM3 re-derives the 18-row 2026-08-19 lost set from the captured record with zero hand-mining: all 18 exec counts reproduce EXACTLY and 17 of 18 last-seen timestamps match to the minute (the 18th differs by 34 seconds because this miner stamps the RESULT record, where exec_meta actually lives, rather than the call), and the single row still LOST is nx_cc_equiv_staged, precisely the row the hand-mine flagged as unsafe to re-register. DM6 matches an INDEPENDENT hand count on every field over one week of organ_ship.jrnl: ships 40, builds 57, per_week_x100 13267, matched 40, unmatched 0, mean 43s, min 0, max 280. DM8 measured the four contract symbols this session shipped as DOCUMENTED with file and line named, and three internal helpers as UNDOCUMENTED, zero false DOCUMENTED. HORIZON, stated because it is part of the claim: DM3's evidence half is laptop-local by necessity -- exec_meta exists only in the client session transcripts (899 files, 2.36 GB, 869,819 records scanned in 26s), and the NAS capture frames carry the tool NAME but a CONSTANT verb and outcome column, so they cannot answer this question. Transcripts are also not immutable: the hand-mine saw 1,139 files where 899 remain.

Where we need to go. The development record itself -- every tool call, measurement, refusal and resume point -- is captured to the estate, mined by standing organs, and drives the ranker. The field versions outcomes [git-object-model], operations [jj-oplog], or delivery exhaust [devlake-dora]; nobody researched ingests session records as primary evidence, does crash-to-multi-seat resume maps, or coordinates concurrent agent writes. Exceed by making the record the first-class object: capture (DM1), resume (DM2), reconcile (DM3, DM4), guard (DM5), and measure delivery from receipts (DM6).

The unit. 1 u = one measured session-leg (organ + gate + live flip), the pmdash/charsim calibration.

24 of 24 capabilities measured|0 of them measured exceeds|0 open|coverage 1000/1000|adoption 17 full / 7 partial

Layer 2 · Roadmap

Do this next — computed by the ranker, never chosen by a seat

Order from nx_compare_rank (nx_dr_ocm: (deficit + cost-of-delay + option + enables) x sponsor x self-sufficiency x momentum / cost). FINISH rows are rungs whose symbol is present but whose organ is short of full adoption: the cheapest closures on this board, listed before any new work. Stamp: # asof=1787883566 domain=devmgmt target_version=- rungs=8 done=8 open=0 finish=2 ranker=nx_dr_ocm

#	Stage	Rung	Priority	Derivation
F	FINISH	Lost-set miner (DM3) rm_lostset	REGISTERED-DARK	callable, authorised, no MCP invocation on record (a direct fork logs the runner, so this is not proof it never ran); no execution surface runs it either (clock, cron, daemon, roster, actlog and surfaced forks checked); only referenced by nx_lostset_gate -- a mention, not a run
F	FINISH	DORA on ship receipts (DM6) d_ship_join	REGISTERED-DARK	callable, authorised, no MCP invocation on record (a direct fork logs the runner, so this is not proof it never ran); no execution surface runs it either (clock, cron, daemon, roster, actlog and surfaced forks checked); only referenced by nx_shipjoin_gate -- a mention, not a run

Critical path — contract, done-rule, executor, cost

Rung	Closes with	Definition of done (pre-declared)	Executor	Est.
Session-to-NAS capture beat (DM1)	ch_beat	A clock row delta-harvests every laptop transcript (ch_delta) and lands frames in a NAS journal; the gauge runs NAS-side over the captured frames. The parser, delta and idempotent ingest already exist -- this rung is composition plus cadence, not a new parser	Organ	1 u
Crash-resume map (DM2) after DM1	st_resume_map	One call after a crash emits a per-seat lane, last-action, in-flight and resume-point map, validated against the hand-mined 2026-08-19 map (7 seats, both crashes)	Organ	1 u
Lost-set miner (DM3) after DM1	rm_lostset	Re-derives the 18-row 2026-08-19 lost set (knowledge/status/toolrows_lostset_20260819.txt) from the captured record with zero hand-mining, exec evidence reconciled against the registry	Organ	1 u
Crash-write adjudicator (DM4)	ap_landed	Given the 2026-08-19 in-flight write list, emits LANDED or NOT per row matching the hand adjudication with zero false LANDED -- wrong only toward NOT	Organ	0.5 u
Owner-scan on BEHIND (DM5)	ps_owner_scan	Every BEHIND worklist row carries an owner signal before any promote list publishes; a replay fixture of the nx_oo_extract_gate case is detected	Organ	0.5 u
DORA on ship receipts (DM6)	d_ship_join	Deploy-frequency and lead-time computed from organ_ship.jrnl match a hand count over one week of journal	Organ	1 u

Milestones

Milestone	Rungs	Cumulative
D0 · Capture and resume	DM1,DM2	2 u
D1 · NAS-mineable full record	DM3,DM4	3.5 u

Rung	Closes with	Definition of done (pre-declared)	Executor	Est.
Adoption-delta emitter (DM7) after DM5	dc_adopt_delta	Reads the daily artifactdrift.log and nx_catalog output and emits the per-class adoption delta, COMPOSING census outputs already written to disk rather than running a new census. Extends nx_deltaclass_lib, whose ledger reader already separates DELIVERED from SCOPE-MOVED from THRASH, so this is a second reader of an old log and adds no new collector and no new load [@devlake-dora]. ACCEPT: two consecutive beats emit deltas whose arithmetic reconciles against the underlying logs, and any row whose reference is REF-FOSSIL is excluded from the promotable delta rather than counted in it	Organ	1 u
Documentation-adoption stage (DM8) after DM5	os_doc_stage	A shipped symbol must appear in a compare matrix, plan or watch row, or carry a memory pointer. REPORT-ONLY first: a pre-flight that can refuse will be disabled, while one that only informs survives [@claude-code-sessions]. WATCH CONTRACT ONLY -- nx_organ_ship is another lane source and this rung licenses no edit to it. ACCEPT: the stage reports DOCUMENTED or UNDOCUMENTED per ship with ZERO false DOCUMENTED, measured against the symbols a session actually shipped	Organ	1 u

D2 · The record drives the intelligence loopDM5,DM6,DM7,DM87 u

Layer 3 · Engineering

How this is scored. Every Nishi mark is measured: the generator reads the real organ source on disk and requires the implementing symbol to exist (no self-grading). A watching tag names the organ and symbol contracted to close a gap — the mark flips itself on the next compare beat when that workstream ships, and the comparewatch- plane row flips with it. The flip is necessary, not sufficient: it proves the symbol exists, never that the capability is good. The bar is the rung's pre-declared done-rule, proven by its gate — a symbol shipped without the behaviour behind it is a defect, and the flip is exactly what makes that defect visible instead of quiet. Competitor marks record documented capability presence — presence, not depth or scale. Adoption is measured too: every measured row carries where its organ stands on the estate's ladder (source → built → promoted → registered → invoked; libraries by importer reach minus validation importers; gates by the execution surfaces that run them). A row is fully adopted only at the top of its ladder; anything short is tagged partial with the exact remedy, so a build nobody promoted can no longer read as shipped. Census stamps: importers asof 1787849099, gate census asof 1787855507 (unix seconds; -1 = census absent).

Capability matrix — measured against source

◉ leads / measured exceed● present◐ partial○ absent · click any capability for its evidence

Capability	Nishi	git	Jujutsu	Devin	Faros AI
DEV
full-fidelity session record harvested to sovereign frames Measured: `ch_frames` exists in `runtime/_hdl_build/nx_claude_harvest.nx`, verified at emit. Parses the Claude Code transcript JSONL (measured 2026-08-19: 869 files, 2.39 GB, laptop-local -- the honest limitation) into actlog frames; ch_delta is HWM-incremental and ch_ingest is marker-idempotent, portable by design. A git commit records exactly a tree, parents, identity and a message [git-object-model]; the transcript contract is declared internal and version-unstable, so the sovereign frame is the durable form [claude-code-sessions]. Devin keeps a full server-side activity record; Faros ingests assistant telemetry counters, never sessions. Adoption: `LIVE` — fully adopted (top of its ladder).	●	○	●	◉	●
nishi-vs-outside tooling gauge Measured: `ch_gauge_json` exists in `runtime/_hdl_build/nx_claude_harvest.nx`, verified at emit. outside_permil over harvested frames measures what fraction of real dev actions still ride non-sovereign tools -- dev-loop sovereignty as a number. OTel GenAI semconv standardizes the naming of agent and tool-call telemetry but ships no storage, replay or resume semantics [otel-genai-semconv]; Faros computes tool-adoption mix from telemetry, the nearest partial. Adoption: `LIVE` — fully adopted (top of its ladder).	●	○	○	○	●
empirical workflow mining over the action journal Measured: `al_mine_json` exists in `runtime/_hdl_build/nx_actlog.nx`, verified at emit. Recurring tool:verb bigrams and trigrams with support at or above minsup are the estate's de-facto workflows, MINED not declared; al_harden_json files hardening candidates and al_append is the frame writer. CodeSight process-mines GitHub PR-to-deployment exhaust -- the delivery workflow's shadow, never how the code came to exist [codesight-procmine]; Devin's session search is retrieval over the record, not pattern mining. Adoption: `LIVE` — fully adopted (top of its ladder).	●	○	○	●	○
trajectory anti-pattern scanning Measured: `tj_core` exists in `runtime/_hdl_build/nx_trajscan.nx`, verified at emit. Deterministic judge-free detectors (edit-without-verify, search loops, echo groups) over actlog frames. The field basis is research literature, not shipped product; no tool in the researched set scans agent trajectories for process defects. Adoption: `LIVE` — fully adopted (top of its ladder).	●	○	○	○	○
workstream lifecycle intelligence Measured: `wm_mine_json` exists in `runtime/_hdl_build/nx_ws_miner.nx`, verified at emit. Per-lane kickoffs, beats and dones with DONE ACTIVE STALE ABANDONED states and per-actor engagement from the ws_sync journal; nx_ws_sync resume lists in-flight lanes from the same record. Devin tracks session state and PR takeover; Faros-class platforms infer lifecycle from repo and issue exhaust. Adoption: `REGISTERED-DARK` — PARTIAL: callable, authorised, no MCP invocation on record (a direct fork logs the runner, so this is not proof it never ran); no execution surface runs it either (clock, cron, daemon, roster, actlog and surfaced forks checked). adoption REGISTERED-DARK	●	○	○	●	●
DORA delivery metrics without git-poll Measured: `d_band_deploy` exists in `runtime/_hdl_build/nx_dora.nx`, verified at emit. Derives the DORA keys from sovereign event planes (ws_sync journal plus the debt plane) where the field polls git, Jira and CI; nx_dora_gate stands beside it. DevLake computes DORA from exactly pull requests, deployments and incidents, the clock starting at the first commit [devlake-dora]; the DORA research itself is survey-based and structurally cannot see a session [dora-2025]. Adoption: `REGISTERED-DARK` — PARTIAL: callable, authorised, no MCP invocation on record (a direct fork logs the runner, so this is not proof it never ran); no execution surface runs it either (clock, cron, daemon, roster, actlog and surfaced forks checked); only referenced by nx_shipjoin_gate -- a mention, not a run. adoption REGISTERED-DARK	●	○	○	○	●
gate-outcome history mining -- flake and erosion Measured: `hm_mine_json` exists in `runtime/_hdl_build/nx_harnessmine.nx`, verified at emit. Mines the per-verdict frames nx_gate_verdict self-records; before it 953-plus gates ran with ZERO recorded outcomes (nx_minecov MC09), flake structurally invisible. The field ingests CI outcomes as delivery exhaust; none mines check-verdict history for flake and erosion. Adoption: `REGISTERED-DARK` — PARTIAL: callable, authorised, no MCP invocation on record (a direct fork logs the runner, so this is not proof it never ran); no execution surface runs it either (clock, cron, daemon, roster, actlog and surfaced forks checked). adoption REGISTERED-DARK	●	○	○	○	●
sovereign VCS push, zero-git Measured: `sgp_walk` exists in `runtime/_hdl_build/nx_sovgit_push.nx`, verified at emit. Tree walk to sha256 objects to packfile to receive-pack wire, no /usr/bin/git anywhere -- the substrate spoken natively [git-object-model]. git and jj are the incumbents on this row and score accordingly. CAUTION: the served binary is drift-census AHEAD class -- a promote from this campaign destroys capability. Adoption: `REGISTERED-DARK` — PARTIAL: callable, authorised, no MCP invocation on record (a direct fork logs the runner, so this is not proof it never ran); no execution surface runs it either (clock, cron, daemon, roster, actlog and surfaced forks checked). adoption REGISTERED-DARK	●	◉	◉	○	○
proof-carrying change -- the capability-loss ruler Measured: `cd_show_lost` exists in `runtime/nx_contentdiff.nx`, verified at emit. GREEN only at zero lost runs, every LOST named; the /api/promote guard shares the arithmetic (cdl_lost). The VCS frontier's mutation histories record graph surgery, never a measurement bound to a state; field CI diffs text. Nothing researched refuses a deploy on measured capability loss. Adoption: `LIVE` — fully adopted (top of its ladder).	●	○	○	○	○
gates proven non-vacuous by mutation Measured: `gb_try` exists in `runtime/_hdl_build/nx_gate_bite.nx`, verified at emit. Plants byte mutations and asserts the gate goes RED; taxonomy BITES, INCONCLUSIVE, UNCONTROLLED. Fleet measured 2026-08-15: 2,317 gate sources, 135 deployed, 30 invoked, 0 invocation gaps. No researched tool proves its own checks can fail. Adoption: `LIVE` — fully adopted (top of its ladder).	●	○	○	○	○
served-vs-built drift census by remedy Measured: `ps_judge` exists in `runtime/nx_artifactdrift.nx`, verified at emit. BEHIND, AHEAD and RESIZED-EQUAL partitions reconciled plus the INERT exec-bit axis (ps_execbit); per-row lost-run worklists and a daily append to knowledge/status/artifactdrift.log, so the trend exists and not just the level. The field's nearest concept is deploy tracking from CI exhaust; none compares served bytes against a rebuild of source. Adoption: `LIVE` — fully adopted (top of its ladder).	●	○	○	○	○
one-call ship loop with a durable journal Measured: `os_jrnl` exists in `runtime/nx_organ_ship.nx`, verified at emit. build, gate, contentdiff, behaveprobe, promote, offc-twin, register and verify as one organ; every stage appends knowledge/status/organ_ship.jrnl and the trace survived four client deaths. Copilot's coding agent links each commit to its session log, a read-only audit surface with no export and no resume [copilot-session-logs]; Devin lands PRs end-to-end without a verifiable stage journal. Adoption: `LIVE` — fully adopted (top of its ladder).	●	○	○	●	○
multi-seat compare-and-swap on the shared tree Measured: `fsw_expect_val` exists in `runtime/nx_fsops_write.nx`, verified at emit. Content CAS whose refusal hands the caller the expect token (tools-api name nx_fs_write); nx_atomic_publish is the whole-file all-or-nothing lane, grounded by the 2026-08-15 concurrent-writer clobber measurement. jj proves lock-free concurrent repo ops with op-log merge [jj-oplog]; Amp threads are durable shared objects with one writer each [amp-threads]; Devin sidesteps coordination by VM isolation. Nobody researched coordinates concurrent agent writes to one shared working tree. Adoption: `PROMOTED-UNREGISTERED` — PARTIAL: a real binary nobody can call over MCP: /api/tools/register it; no execution surface runs it either (clock, cron, daemon, roster, actlog and surfaced forks checked). adoption PROMOTED-UNREGISTERED	●	○	◉	●	○
seat token-engine boot and metering Measured: `st_boot` exists in `runtime/_hdl_build/nx_seat.nx`, verified at emit. One-call boot digest at or under 4096 B versus a 40K-token memory-file boot (12-session baseline: cache_read 920.8M vs output 6.1M tokens); st_meter_append is the per-session dev-cost ledger. DX meters utilization, impact and cost from tool telemetry and surveys, outside-in [dx-ai-framework]; the seat meters itself from inside the session. Adoption: `LIVE` — fully adopted (top of its ladder).	●	○	○	○	●
per-seat crash-resume map from transcript tails Measured: `st_resume_map` exists in `runtime/_hdl_build/nx_seat.nx`, verified at emit. Hand-done TWICE on 2026-08-19 (7 seats died near 11:26 and again near 21:09; the coordinator mined transcript tails, scratchpads and output tails for exact per-seat resume points). Langfuse session replay is viewing the record, not resuming the agent [langfuse-sessions]; Devin's session search is post-hoc analytics with no path back to a resumable state; jj op restore rewinds repo state only. Incumbent floor: nx_ws_sync resume reads the journal but cannot see last-action or in-flight writes; compose ch_frames over the tail. Adoption: `LIVE` — fully adopted (top of its ladder).	●	○	●	●	○
session-to-NAS capture beat Measured: `ch_beat` exists in `runtime/_hdl_build/nx_claude_harvest.nx`, verified at emit. Parser, delta (ch_delta high-water mark) and idempotent ingest already exist and are portable; what is missing is cadence plus a NAS landing, so 869 files and 2.39 GB of dev record are unmineable from the estate. Checkpoints live tool-locally outside repo history, two unjoined histories by platform default [claude-code-checkpoints]; Devin's sessions land server-side and searchable by activity class, the strongest field capture [devin-session-search]. Done-rule: a clock row delta-harvests every transcript and frames land in a NAS journal. Adoption: `LIVE` — fully adopted (top of its ladder).	●	○	○	◉	●
exec-proven lost-set miner Measured: `rm_lostset` exists in `runtime/_hdl_build/nx_registrymine.nx`, verified at emit. On 2026-08-19 an agent hand-mined 836,673 transcript lines and 32,122 MCP calls to find 18 exec-proven tools with no registry row (knowledge/status/toolrows_lostset_20260819.txt, debt 1787175603); the shape belongs in the registry miner as a standing verb. Faros reconciles AI-tool adoption from telemetry and repo exhaust, never exec-proven from the session record [faros-ai-impact]. Adoption: `REGISTERED-DARK` — PARTIAL: callable, authorised, no MCP invocation on record (a direct fork logs the runner, so this is not proof it never ran); no execution surface runs it either (clock, cron, daemon, roster, actlog and surfaced forks checked); only referenced by nx_lostset_gate -- a mention, not a run. adoption REGISTERED-DARK	●	○	○	○	●
concurrent-ownership detector on BEHIND rows Measured: `ps_owner_scan` exists in `runtime/nx_artifactdrift.nx`, verified at emit. BEHIND is necessary-not-sufficient for promote: a BEHIND row can be another seat's in-flight work (caught live on nx_oo_extract_gate; nx_treediverge.nx carries two scratch files beside it now, nx_artifactdrift.nx nineteen). GitButler hands agents a CLI for commits yet documents no record of which agent made which change [gitbutler-vbranches]; git names an author only once a commit exists, jj stamps username and hostname per operation. Done-rule: every BEHIND row carries an owner signal before any promote list publishes. Adoption: `LIVE` — fully adopted (top of its ladder).	●	●	●	○	○
crash-write adjudicator Measured: `ap_landed` exists in `runtime/_hdl_build/nx_atomic_publish.nx`, verified at emit. At both 2026-08-19 crashes the in-flight async writes split roughly half landed and half not; six lanes were adjudicated by hand from artifacts. The jj op log records every repo mutation with a full view snapshot, the strongest landed-or-not substrate for repo state; Cursor checkpoints snapshot files with crash persistence undocumented [cursor-checkpoints]. Done-rule: given expected writes, read back size and hash, emit LANDED or NOT per row, wrong only toward NOT. Adoption: `LIVE` — fully adopted (top of its ladder).	●	○	●	○	○
DORA fed by ship receipts Measured: `d_ship_join` exists in `runtime/_hdl_build/nx_dora.nx`, verified at emit. nx_dora's constants name only the ws_sync journal and the debt plane while organ_ship.jrnl carries the actual promote and verify events; the join turns deploy-frequency and lead-time from workstream proxy into the real ship record. The field's delivery clock starts at the first commit and consumes only PR, deploy and incident exhaust; even the strongest measurement frameworks concede agentic measurement must still evolve. Adoption: `REGISTERED-DARK` — PARTIAL: callable, authorised, no MCP invocation on record (a direct fork logs the runner, so this is not proof it never ran); no execution surface runs it either (clock, cron, daemon, roster, actlog and surfaced forks checked); only referenced by nx_shipjoin_gate -- a mention, not a run. adoption REGISTERED-DARK	●	○	○	○	●
ADOPT
the whole adoption chain judged end to end with a weakest-link verdict Measured: `cat_one` exists in `runtime/nx_catalog.nx`, verified at emit. SOURCE then BUILT then STAGED then PROMOTED then REGISTERED then AUTHORISED then INVOKED, one call per name, and the published answer is the WEAKEST link rather than a headline -- so BUILT-UNPROMOTED, PROMOTED-UNREGISTERED and REGISTERED-DARK are three answers with three different remedies instead of one word. This is the adoption axis for a single CAPABILITY; growth.matrix gs_axis_adoption measures the estate-wide STOCK from source-level callers, and the two are deliberately kept apart rather than merged into a duplicate ruler. Faros computes adoption and acceptance rates from assistant telemetry and repo exhaust, which measures whether humans use a TOOL, never whether a capability this process built ever became callable [faros-ai-impact]. Copilot links each commit to its session log for audit, a read-only surface that says nothing about deployment state [copilot-session-logs]. Adoption: `LIVE` — fully adopted (top of its ladder).	●	○	○	○	●
an unpromoted build is DECLARED by its owner, never inferred by a reader Measured: `ps_decl_has` exists in `runtime/nx_artifactdrift.nx`, verified at emit. The standing law is that any build you do not promote must be said out loud with a reason. This gives that statement a place to live and a reader that acts on it, so a claimed target reads DECLARED on the promote worklist and an unclaimed one cannot hide among the claimed. Rows are anchored on field 0, so a target name appearing inside another row prose can never claim it -- a whole-buffer search would have reported rows as owned that nobody owned. jj stamps username and hostname on every operation, the closest field equivalent, but it records who DID something rather than who is HOLDING something right now [jj-oplog]. GitButler hands agents a CLI for commits and history surgery and documents no record of which agent made which change [gitbutler-vbranches]. Adoption: `LIVE` — fully adopted (top of its ladder).	●	○	●	○	○
the adoption delta as a per-class trend, never a level Measured: `dc_adopt_delta` exists in `runtime/_hdl_build/nx_deltaclass_lib.nx`, verified at emit. A level says 147 binaries are behind; only a delta says whether the promotable backlog is shrinking or merely changing hands. The incumbent to extend is the ledger reader that already separates DELIVERED from SCOPE-MOVED from THRASH, and both inputs are census outputs already on disk -- artifactdrift.log once per beat and nx_catalog once per name -- so this rung COMPOSES and never sweeps. DevLake computes delivery rates from exactly pull requests, deployments and incidents with the clock starting at the first commit, so a capability built and never deployed is invisible to it by construction [devlake-dora]. DX meters utilization, impact and cost from tool telemetry and surveys and its own authors concede agentic measurement must still evolve [dx-ai-framework]. Done-rule: two consecutive beats emit deltas whose arithmetic reconciles against the underlying logs. Adoption: `LIB-WIRED importers=2 nonval=1` — fully adopted (top of its ladder).	●	○	○	○	●
documentation as a shipping stage, reported per ship Measured: `os_doc_stage` exists in `runtime/nx_organ_ship.nx`, verified at emit. A capability that is built, promoted, registered and undocumented is one nobody will find, and the operator named documenting alongside deploying and wiring as one failure rather than three. The stage asks one mechanical question at ship time -- does this symbol appear in a compare matrix, plan or watch row, or carry a memory pointer -- and REPORTS rather than refuses, because a pre-flight that can refuse gets disabled while one that only informs survives. Claude Code sessions persist locally with the full tool history restored on resume and no link to any capability register [claude-code-sessions]. Amp threads are durable shared objects searchable by label, file, repo and author, which shares the CONVERSATION rather than registering the capability it produced [amp-threads]. Done-rule: DOCUMENTED or UNDOCUMENTED per ship with zero false DOCUMENTED. Adoption: `LIVE` — fully adopted (top of its ladder).	●	○	○	○	○

Risk register

Risk	Likelihood x impact	Mitigation
The capture beat hammers the NAS with a 2.39 GB backfill	possible x medium	ch_delta is HWM-incremental so the beat lands deltas only; the first backfill runs as a named one-off job under the io-admission ruler, never on the beat.
A resume map read as authority replays a stale plan	possible x medium	DM2 emits evidence rows (lane, last artifact, in-flight writes, timestamps); the relaunching coordinator judges -- the map never auto-relaunches a seat.
A watch contract flips on symbol existence, so a stub reads green	possible x high	The done-rule proven against the 2026-08-19 hand-mined ground truth is the bar; the flip is the receipt, never the proof.

On these two registers. Rows are declared in the domain's plan file and carry the debt id, which is the join key back to the sovereign debt plane — that plane, not this page, is the authority on state. Reconciling them automatically (the regen reading the plane and refreshing these rows) is a named, owed rung; until it lands, treat an id here as a pointer to look up, not a status to trust.

Honest verdict. The coverage above is capability presence measured against source — not depth, scale, or polish, where mature rivals may lead. Exceeds are claimed only where a mechanism backs them. Every open gap is a watch contract: it names the organ and symbol that closes it, and this page flips the cell itself when that workstream ships.

Person · product · place — not yet measured for this domain

Every compare carries this layer. Declare knowledge/compare/devmgmt.ppp (rows surface|nishi or c1..c4|label|url|connect naming OUR live surface and each rival's front door), run nx_ppp_probe domain devmgmt, and this section fills itself on the next beat: the same ruler on both sides — privacy and CX (third-party hosts, tracker classes, cookies, security headers), design and longevity (design hygiene, computed WCAG contrast, render-blocking resources, unsized media, script weight, theme and motion queries), findability (landmarks, skip link, on-site search, breadcrumb, headings, internal links).

References

Beyond a link list. Every reference below resolves twice — the publisher's copy and, where banked, the estate's own non-rottable library mirror with a content pin — and carries its evidence class plus the exact claim on this page it grounds. Keyed marks like [key] in the matrix notes jump here. A dash means honestly absent, never assumed.

[git-object-model] Pro Git, Git Internals - Git Objects (official git-scm.com book): a commit records exactly a tree hash, parent hashes, author and committer identity with timestamps, and a free-text message -- nothing else; the development process collapses to snapshots plus prose. publisher · read in our library knowledge/fetched/cmp_devmgmt_git-object-model.html · pin h74c89abda3e3d044d865707d89d8b33714b34c74a3032937db53a54ba38f35f3 · accessed 2026-08-19 · vendor-docGrounds: DEV: full-fidelity session record harvested to sovereign frames
[jj-oplog] Jujutsu official docs, Operation log: every repo-mutating operation is recorded with a full view snapshot (refs, heads, per-workspace working-copy commits) plus username, hostname and time, enabling undo, op revert, op restore and lock-free concurrent operations -- who-ran-what-when on the repo only, no reason and no measurement attached to any state. publisher · read in our library knowledge/fetched/cmp_devmgmt_jj-oplog.html · pin h9c337e3458805334a24ca402cce2b8fcacf2a85f26b76f38981c2657a3d5b19b · accessed 2026-08-19 · vendor-docGrounds: DEV: multi-seat compare-and-swap on the shared tree
[devin-session-search] Cognition, Devin release notes (entries 2026-03-19 to 2026-07-22): Session Search analyzes past sessions with full search across shell, file, browser, git and MCP activity; Devin-manages-Devins delegates to parallel isolated VMs; PR takeover continues work across sessions -- post-hoc analytics with no documented path from the record back to an exact resumable work state. publisher · read in our library knowledge/fetched/cmp_devmgmt_devin-session-search.html · pin hb4c081723a73eb849e5d99504f2791c512d43c4cbd494dd7663545bf0a227d1b · accessed 2026-08-19 · vendor-docGrounds: DEV: session-to-NAS capture beat
[faros-ai-impact] Faros AI, AI Coding Assistant Impact and Productivity Metrics product page: ingests AI-assistant telemetry (Copilot, Claude, Cursor, Windsurf, Devin, Amazon Q) plus repos, Jira, CI and surveys; computes adoption, acceptance rates and percent-AI-generated by repository -- attribution stops at assistant telemetry and repo granularity, never the in-session process. publisher · read in our library knowledge/fetched/cmp_devmgmt_faros-ai-impact.html · pin h674d36e02fea9b26fffa4fda443203f5efc5a11a19f7f8aabe42c78878c49b00 · accessed 2026-08-19 · vendor-docGrounds: DEV: exec-proven lost-set miner
[claude-code-sessions] Anthropic, Manage sessions (Claude Code docs): sessions persist as local JSONL transcripts with the full history including tool calls restored on resume, plus branching and scripted access; the entry format is declared internal and version-unstable, and two terminals on one session interleave into one transcript -- no multi-seat write coordination, no cross-seat crash recovery. publisher · read in our library knowledge/fetched/cmp_devmgmt_claude-code-sessions.html · pin hc466e842c9a0c734163e77dc0019d9575cffb7163b337f844b14369066f1f5e3 · accessed 2026-08-19 · vendor-docGrounds: DEV: per-seat crash-resume map from transcript tails
[claude-code-checkpoints] Anthropic, Enabling Claude Code to work more autonomously (2025-09-29): checkpoints auto-save code state before each agent edit and can rewind code, conversation or both -- the first mainstream versioning of the intent trace; they exclude user edits and bash commands, live tool-locally outside repo history, and Anthropic directs users to run real VCS alongside. publisher · read in our library knowledge/fetched/cmp_devmgmt_claude-code-checkpoints.html · pin hfe0d69c45f12bfe98df5a8389f37ef21a0cdb83f858e92fbd23108a678e8a0be · accessed 2026-08-19 · vendor-docGrounds: DEV: session-to-NAS capture beat
[cursor-checkpoints] Cursor docs, Checkpoints: the agent automatically creates checkpoints before significant changes, capturing the state of all modified files, restorable from the chat timeline, stored locally and separate from git; persistence across restart or crash is undocumented, and there is no query, export or mining surface. publisher · read in our library knowledge/fetched/cmp_devmgmt_cursor-checkpoints.html · pin h3c18bc24b644b21c61a62690a7bcd1e363a25ec8049935e49b15f0f6c7dfaa4d · accessed 2026-08-19 · vendor-docGrounds: DEV: crash-write adjudicator
[copilot-session-logs] GitHub docs, Tracking GitHub Copilot sessions: session logs expose the coding agent's internal reasoning and the tools it used, and each commit message links to its session logs for audit -- a read-only surface with no documented resume of a crashed session, no API export and unstated retention. publisher · read in our library knowledge/fetched/cmp_devmgmt_copilot-session-logs.html · pin h26d8e81eaecd19d99676f1067b9b9a86a37e63a26b35ac5b49643724955e95f0 · accessed 2026-08-19 · vendor-docGrounds: DEV: one-call ship loop with a durable journal
[amp-threads] Sourcegraph, Amp Owner's Manual: threads are durable server-side objects shared with workspace members by default, searchable by label, file, repo, author and date, with remote control across devices -- whole-conversation sharing for humans, one writer per thread, no mechanical reconstruction of interrupted work state and no coordination protocol between concurrent writers. publisher · read in our library knowledge/fetched/cmp_devmgmt_amp-threads.html · pin h946c2b24ccca92a4be495ff4d329799f7e07e6a7479808875684d62a61e9182b · accessed 2026-08-19 · vendor-docGrounds: DEV: multi-seat compare-and-swap on the shared tree
[langfuse-sessions] Langfuse docs, Sessions: a propagated sessionId groups observations and traces into a session with replay of the entire interaction plus session-level scores via UI, SDK and API -- replay is viewing the record, not resuming the agent; it observes an instrumented app and cannot restore or coordinate agent working state. publisher · read in our library knowledge/fetched/cmp_devmgmt_langfuse-sessions.html · pin he830faa9f9b0efaa150d3ee6fe507129e47586fb6711d3d5302fd11e38efd994 · accessed 2026-08-19 · vendor-docGrounds: DEV: per-seat crash-resume map from transcript tails
[otel-genai-semconv] OpenTelemetry, semantic-conventions-genai repository: vendor-neutral semantic conventions for GenAI spans, metrics and events covering GenAI clients, MCP and provider-specific conventions; in development with no declared stability -- standardizes telemetry shape only, with no storage, replay, resume or recovery semantics. publisher · read in our library knowledge/fetched/cmp_devmgmt_otel-genai-semconv.html · pin h101ed9c8ca53eb9cdde8d2f002b400266c45ec0e0a17c97a74c93473d16e2571 · accessed 2026-08-19 · published-standardGrounds: DEV: nishi-vs-outside tooling gauge
[dora-2025] DORA (Google Cloud research program), State of AI-assisted Software Development 2025: AI is an amplifier magnifying existing organizational strengths and weaknesses, returns come from the underlying organizational system; survey-based respondent self-report -- it measures perceptions and reported outcomes and structurally cannot see any actual development session. publisher · read in our library knowledge/fetched/cmp_devmgmt_dora-2025.html · pin h6bb53c243fc615849727e65b6a9aa33dbd6b0d5c4e63af5ce68137db0924b4d9 · accessed 2026-08-19 · published-paperGrounds: DEV: DORA delivery metrics without git-poll
[devlake-dora] Apache DevLake documentation, DORA metrics implementation: computes DORA from exactly three entities -- pull requests, deployments, incidents -- sourced from CI systems, issue trackers and repos, webhooks for the rest; lead time is measured commit-to-production, so everything before the first commit (IDE work, agent sessions, abandoned attempts) is invisible by construction. publisher · read in our library knowledge/fetched/cmp_devmgmt_devlake-dora.html · pin h253628287ee222cbc38c1216811d5cb00ac514bbf7ec7796c7effc79c925ea47 · accessed 2026-08-19 · vendor-docGrounds: DEV: DORA delivery metrics without git-poll
[dx-ai-framework] DX Research, Measuring AI code assistants and agents (AI Measurement Framework): three dimensions -- utilization, impact (AI-driven time savings plus Core 4), cost -- from AI-tool telemetry, git data and surveys; agents are treated as team extensions with contribution measured by emitted PRs, and the authors concede agentic measurement must still evolve. publisher · read in our library knowledge/fetched/cmp_devmgmt_dx-ai-framework.html · pin h778e2f333c133d411f930c36a9248dd4329f3118cf8de40acafa12c50701baa7 · accessed 2026-08-19 · vendor-docGrounds: DEV: seat token-engine boot and metering
[codesight-procmine] Dorado, Folgueira, Martin et al., CodeSight: process mining for software development workflows, arXiv 2510.25935 (Oct 2025): captures GitHub PR-to-deployment events as process-mining logs, mines workflow variants and trains an LSTM to predict remaining PR resolution time -- the event log begins at GitHub exhaust, so it mines the delivery workflow's shadow; how the code came to exist is outside the log. publisher · read in our library knowledge/fetched/cmp_devmgmt_codesight-procmine.html · pin hcab78693a99affba3d3555d9b98cd4237b81cf6cd9f026858adf390b27011bca · accessed 2026-08-19 · published-paperGrounds: DEV: empirical workflow mining over the action journal
[gitbutler-vbranches] GitButler docs, Virtual Branches: multiple branches applied to one working directory simultaneously with per-lane staging and hunk-level assignment of uncommitted changes, plus an oplog restoring prior workspace layouts; the agents doc gives agents a CLI for commits and history surgery but documents no session tracking and no record of which agent made which change. publisher · read in our library knowledge/fetched/cmp_devmgmt_gitbutler-vbranches.html · pin h805c988d5cf89d5316e28e7006caef196620e09483c2daf4b204fdaa4ce8a2fc · accessed 2026-08-19 · vendor-docGrounds: DEV: concurrent-ownership detector on BEHIND rows

generated by nx_swcompare_matrix (sovereign NishiLang organ) from knowledge/compare/devmgmt.matrix · every Nishi cell verified against organ source at emit time · watch cells re-measured on every compare beat · zero JS, zero trackers