Eve library

V8-005 flagship readiness boundary

Evidence

Revision 1

On this page

Date: 2026-07-27

Source commit: e8a324655f8c4eaf85782dab96635dfcf84782b0

Outcome#

V8-005 now has one retained, machine-readable admission package over the existing canonical evaluateV8ReferenceSlice decision. The package is blocked on all six ledger claims. It records neither a synthetic candidate digest nor an invented sample threshold.

The contract also closes two proof weaknesses in the original evaluator. A candidate must now retain exact-artifact executions for each canonical G1–G7 gate, and the best-authored comparison must be preregistered, held out, blinded, randomized, independently rated, adequately sampled, and clear a lower 95% preference bound above 0.50.

Required evidence#

scripts/v8/verify-flagship-evidence.ts requires one concurrent package with:

  • four successful non-fixture production provider roles and exact-artifact G1–G7 execution receipts;
  • the named UE5.5 Linux project, editor and cook commands, archive digest, placement, play, accessibility, and unique real-media digests;
  • exact-artifact zero-human and full takeover/edit executions with journal, review, edited-artifact, and mixed-lineage evidence;
  • a distinct best-authored held-out baseline and predeclared human comparison whose lower 95% candidate-preference bound exceeds 0.50;
  • provider-failure, journal-restart, reliability, canary, incident, rollback, and recertification drills plus a policy-sized reliability window; and
  • signed, unexpired GA authority after every provider, gate, engine, route, comparison, reliability, and recovery timestamp.

Audit mode (--report) emits the blocked report; strict mode exits nonzero. Malformed nested JSON fails closed rather than crashing the evidence consumer.

Retained audit#

The repository has no real V8 flagship candidate, provider execution package, UE5.5 cooked/play receipt, real-player authored comparison, complete operations drill, or post-evidence GA approval. The authorized study has not supplied a power analysis from which to set the minimum rating count, so that policy field remains null.

All six V8-005 checkboxes remain open. This contract changes no user-facing UI, so Playwright, mobile, and accessibility automation are inapplicable to this repository-only evidence-boundary change.

Verification#

Executed on dev-server, Linux 6.8.0-110-generic x86_64, Node v24.14.1, pnpm 10.25.0, on 2026-07-27 UTC:

  • pnpm exec tsc --noEmit -p libs/yemaya/case-eval/tsconfig.lib.json — pass;
  • node --import tsx --test scripts/v8/verify-flagship-evidence.test.ts — 5/5 pass, including malformed JSON, absent evidence, missing G7, weak comparison, and retained-report fixtures;
  • pnpm exec nx test yemaya-case-eval --runInBand — pass;
  • focused ESLint with zero warnings and pnpm exec tsc --noEmit -p scripts/v8/tsconfig.json — pass;
  • retained audit/report regeneration and byte comparison — pass;
  • strict retained-package execution — expected exit 1;
  • direct TypeScript emit for yemaya-case-eval — pass, 1,267 output files.

pnpm exec nx build yemaya-case-eval --parallel=1 was attempted after the required resource check. It was stopped after unrelated dependency targets failed: @oshun/tracing has five existing implicit-any diagnostics and @oshun/native-libs lacks declarations for pako and pngjs. The owning library's direct typecheck, full test target, and direct emit all pass; this evidence does not claim the broader dependency build is green.