Disciplines · Audits

Eve creative release-gate extension preparation — 2026-09

Gap audits and as-built reviews.

0sections2 minread

On this page

Task 12.10 now extends the existing Task 12.7 family-release authority at its DCC family row. The creative artifact is subordinate evidence: it returns refusals to the family evaluator and cannot declare a release pass. Task 12.8 continues to recompute and verify the Task 12.7 report, so the charter gate automatically consumes the same creative decision without a second authority.

The content-hashed registration derives four exact family rows from the Task 12.9 benchmark: reference motion, original motion, reconstruction, and walkthrough. Each row binds the required case classes, criterion order, thresholds, and strata, and names its result artifact. The retained registration uses null result references because no live results exist. Its benchmark reference is current, but that benchmark itself remains blocked pending Task 17.7 ratification; both conditions become explicit DCC refusals.

When result evidence is supplied, the extension checks:

  • exact benchmark, release, family, case-class, criterion, threshold, and stratum identity, with the preregistered independent sample minimum applied in every stratum;
  • non-empty, content-hashed source, native project, delivered output, native readback, host, runtime, route, attempt, score, review, and decision evidence;
  • a fresh native-tool open-and-readback probe. Compressed native containers are accepted when the probe evidence is compression-aware; a generic byte or text scan cannot create a false unopenable-project failure;
  • fresh verified host and runtime pins, plus the exact approved registry route and observed runtime receipt. A test provider or contract fixture cannot be presented as a live route;
  • verified-complete job state, every attempt in contiguous order, every failed attempt retained, and provider, render, storage, and streaming cost legs reconciled to both attempt and job totals;
  • every target and floor computed separately, every required stratum passing, and no weakened or omitted quality threshold;
  • required case-level human approval and family-level independent review by the benchmark's named reviewer role, distinct from the execution owner; and
  • current visual-judge calibration over every case class when a visual judge is used. An unused judge must claim no calibration, and an uncalibrated score cannot be substituted for required humans.

The focused contract suite passes a complete four-family fixture, including a compressed native container opened through a native compression-aware probe. It observes red for 23 boundaries: family omission, stale or blocked benchmark, wrong release, postdated registration, missing readback, generic compressed-file probing, unopenable projects, test-provider evidence presented as live, stale host/runtime pins, partial jobs, hidden failures, unreconciled labor, unpriced render work, threshold drift, missing or undersized strata, missing or self-reviewed human decisions, uncalibrated or incomplete visual scoring, and unresolved limitations. The real CLI also requires the retained four-family candidate to exit blocked.

Inspect that retained creative decision directly with:

bash
node tools/eve-everywhere/verify-creative-release-candidate.mjs

The command exits nonzero while blocked. The authoritative composed decision is still the existing family gate:

bash
node tools/eve-everywhere/evaluate-family-release-gate.mjs \
  --manifest=docs/audits/eve-sota-family-release-gate/2026-09-13.candidate.json \
  --output=docs/audits/eve-sota-family-release-gate/2026-09-13.report.json

The retained production report remains blocked at 0 of 15 families with 106 refusals. The added refusal is the DCC creative extension: Task 12.9 is blocked and all four creative result artifacts are missing. No capstone, native project, human label, live provider request, host receipt, quality score, or release pass was manufactured. Task 12.10, Phase 12, G7, and G14 remain open.