Editor Coverage
THE ECOSYSTEM-AS-EDITOR LAW: the model is the author; the ecosystem editorial passes (gates · evidence-derived phase · adversarial verify · best-of-N plus judge · recombination) lift any seat good draft to award-winning — and every known model failure mode below must be covered STRUCTURALLY, with evidence, never by trust.
Covered
8
Partial
3
Gap
1
Structurally covered
666‰
| row | failure mode | covering capability (structural) | evidence | status | note |
|---|---|---|---|---|---|
| 1 | dithering-overask | nx_ws_cycle next = ONE derived action; scoped debt-gate refuses drift-work | knowledge/store/cycle- + ws_sync.jrnl | COVERED | loop R1/R2a proven live 07-18 |
| 2 | fake-done-asserted-progress | evidence-derived phase R2b: phase advances ONLY via machine artifacts, caller can point never assert | bench/nx_ws_cycle_gate 12/12 + rejected-DONE-over-debt flips | COVERED | the anti-fake-DONE tooth working live |
| 5 | stale-memory-assertion | live-query law: board/standup/nx_stale_check beat recalled state | /standup + query-board-first banked law | COVERED | progress is a live query not a memory |
| 6 | clobber-duplicate-work | announce+board law + claims/lease + nx_clobber_guard + nx_orchestrate wait-for-opening queue | ws_sync.jrnl + orchestrate.log + converge-stood-down precedent | COVERED | F748 anti-clobber program card composes these |
| 8 | hallucinated-context | rule-4 verified-facts + evidence layer + chain-of-evidence suite (DSSE/in-toto) | knowledge/status/*.log + coe_dsse.json valid envelope | PARTIAL | coe pillars F702-F704 open; every prompt fact must trace to a real artifact |
| 9 | sycophancy-false-positivity | liar-killer gates + honest UNMEASURED/ANALOG grammar + no-self-grade two-witness law | modelcard- plane grammar + ecomat triangulation | COVERED | fabrication has no code path |
| 10 | capability-overclaim | modelcard NISHI col MEASURED-only grammar + external-ruler harnesses | /modelcard live + F786-F791 filed | PARTIAL | harness climb = the measure rungs; edit-lift meter = F796 |
| 7 | silent-build-success-belief | fail-loud /api/build: seq132 truncate-stale-elf-before-compile + diag tail surfacing + promote ELF-magic refuses empty stage | BEHAVIORAL PROOF 07-20: good build 345B BUILT -> broken rebuild BUILD-FAILED w/ real nx_cc error surfaced (no silent stale BUILT) -> promote of 0-byte stage REFUSED 400 | COVERED | seq202 EATEN with live 3-probe evidence; the editor never lets a seat believe a failed build shipped |
| 11 | silent-build-wrong-target | NONE YET - caller-side discipline only (verify the build response echoes YOUR requested target before promoting) | seq236 filed 07-20: a concurrent /api/build returned staged=nx_printer_ctl for an nx_seat request (shared /tmp capture path cross-wire); a seat trusting the echo could promote the WRONG live organ | GAP | found BY this session's own tooling; structural root-fix (per-build unique capture path + fail-loud target-echo tooth) = seq236, queued for the F805 mgmt batch - NOT hidden |
| 4 | patch-cascade | nx_patch_guard: rule-3 mechanized - N consecutive patch-rows on a target since the last rewrite|gate row = REFUSE (threshold is a CONFIG row, not code) | gate 4/4 07-20: permit(1 patch)=rc0 / cascade(2)=REFUSE rc3 fail-closed / threshold-as-DATA (config 2to3 flips verdict) / rewrite-escape-hatch resets 2to0; plane knowledge/store/patchledger- | COVERED | the editor's last original GAP closed; enforcement-wiring into the ws_cycle debt-gate = filed rung F837 (today the tooth is provable but advisory) |
| 3 | silent-truncation | scale-law + nx_envelope_audit (F227 D): every tool DECLARES its envelope in output; the sweep MEASURES fleet compliance instead of asserting it, and the RELIABILITY GUARDS now declare their own bounds | MEASURED BEFORE/AFTER 07-20 knowledge/status/envelope_audit.log: baseline 225 tools declare_index 447 permille -> after the F846 guards-first pass 229 tools / 55 declaring / 54 candidates = 504 permille; the 4 guards (nx_deploy_ready nx_api_contract nx_cron_watch nx_debt_hygiene) flipped candidate->declaring, laptop SSOT + buildroot both synced; daily beat 06:50 | PARTIAL | 54 candidates remain (F846 continues under MIGRATE-ON-TOUCH). Sweep SELF-CORRECTED once already (SOURCE-MISSING over-claim -> NAME-PROBE MISS w/ the elf-vs-source divergence blind spot declared in its own output) |
| 12 | magic-number-bumping | nx_law_warden L001: mechanically detects inline numeric literals >=1024 on non-const non-comment lines, fleet-wide, hourly at :40 - rule-11 says the threshold belongs in a named const or a config row, never buried in code | MEASURED 07-20 knowledge/status/law_warden.log: 2348 literals across 600 scanned organs (capped=1, MORE exist); worst nx_archive_daemon.nx; remediation seq274 owner=census | COVERED | operator named the BEHAVIOR precisely: workstreams hit a bug and just give it a NEW NUMBER instead of root-fixing = the bump-the-constant sibling of patch-cascade. Detection is structural+hourly; the 2348 are the backlog. OWN-HOUSE-FIRST: nx_tooldiff (mine, shipped today) was itself a breach - rewritten rule-11-clean same round |
F795 LAW: a GAP row without a filed F-rung is a defect. Source plane knowledge/store/editcover- (hist- provenance) · envelope 256KB plane / 64 rows / 128KB out · proof lineage: autofix 71 pct resolve at 0.5-1.5B makers WITH best-of-N plus judge. See also standup · model card · frontier.