Eve library

V6-002 — Template fallback frequency and disclosure evidence

Evidence

Revision 1

On this page
  • Owner: Codex implementation agent
  • Accountable product: V6
  • Priority: P0 live operational seam
  • Linked task: V6-002, measure template fallback frequency, label fallback output to users and evidence, and reduce the rate only through real provider reliability
  • Source commit: c100c9f70350c9014abe5b8b3ad269ceebf7fe2b
  • Evidence captured: 2026-07-24

Implemented boundary#

The cognition stack now produces a bounded v6.cognition-stack.template-fallback-frequency.1 report from the typed provenance on generation outputs. The report records the measurement window, sample count, template-fallback count, model-generation count, fallback rate in basis points, reason counts, and compact evidence tags. Empty windows report no rate instead of zero, and windows above 10,000 samples are rejected.

Only an output whose source is model lowers the fallback rate. Changing UI copy, evidence labels, or a reason string cannot turn a template into model generation. This preserves the task's provider-reliability condition at the measurement boundary.

The Tier-2 web fallback measures the current localized dialogue plus five Chronicle beats on every locale render. Because no native dialogue or Chronicle model is wired into that client, the current honest result is 6/6 template fallbacks (100.0%). The localization panel exposes that result to the user and the diagnostics surface repeats the count and rate. The existing provenance summary now says template fallback (no native model wired) rather than claiming native generation.

English, Spanish, and Yoruba browser coverage asserts the exact disclosure and window identity after locale transitions. The implementation does not set or claim a synthetic reduction target; the displayed rate can fall only when production generation supplies real source: model outputs.

Browser-runtime repair#

The browser gate exposed Node-only SHA-256 imports in the prompt-trust and Clio paths used by the Tier-2 client. Prompt trust now exports a synchronous, runtime-neutral SHA-256 implementation with standard empty, ASCII, and UTF-8 test vectors. Clio reuses the same implementation, and encoded-delimiter neutralization uses platform atob and TextDecoder rather than Node Buffer. The production web build no longer externalizes node:crypto.

The software-rendered desktop Playwright project requires more than its former 75-second ceiling for several established interactive scenarios. The bounded per-test timeout is now 120 seconds. All ten desktop scenarios completed below that ceiling. The combined two-project run completed desktop 10/10 and the first three mobile cases before its long-lived Vite server exited; a fresh complete mobile-project run passed 10/10. The server-lifetime event was ERR_CONNECTION_REFUSED, not an application assertion failure.

Verification#

Command / gate Result
Cognition-stack Vitest PASS — 2 files, 16 tests
Tier-2 web fallback Vitest PASS — 1 file, 12 tests
Prompt-trust Vitest PASS — 1 file, 9 tests
Clio-story Vitest PASS — 2 files, 16 tests
Direct cognition, prompt-trust, Clio, and web TypeScript checks PASS
Tier-2 direct production build PASS — 304 modules; no Node crypto externalization
Focused localization Playwright, desktop and mobile PASS — 2 tests
Full desktop Playwright project PASS — 10 tests
Full mobile Playwright project PASS — 10 tests
Strict ESLint over changed TypeScript PASS
Prettier and git diff --check PASS

Artifact hashes#

Artifact SHA-256
libs/v6/cognition-stack/src/index.ts 87f7922f0117db32ec0e8e81d00636fcd78cb3387e6ca800161b0a8e15b2fd6d
libs/v6/cognition-stack/src/index.spec.ts d09a9369029dab8743963cb891ca87e09b26ace788e830060a5ace3690916885
apps/v6/egbe-web-fallback/src/app.ts 1a49677d2c662ee84a82525c8e460cbf02fc6f23dd493a5671f943537d6ce9b3
apps/v6/egbe-web-fallback/src/app.spec.ts b68303997e1bf3ca519acc7c7ed10444daa8910321c7cf43d9993399d8b83ce1
apps/v6/egbe-web-fallback/e2e/egbe-web-fallback.spec.ts b5dd98b393fea1119a120b84aa5b42ccfd6f0abfe0fd3d5ae8a58d0fd182006e
apps/v6/egbe-web-fallback/playwright.config.ts 69d2e82cc73decd6ffef4d04bdb4fd51d9f7931e36795cfa79fd45efeb3d1502
libs/shared/prompt-trust/src/sha256.ts ca38c96bb5b7f0a5bd9c9a6fc371f278f4b82a16bf35f04da45d1f324019fa6c
libs/shared/prompt-trust/src/trust-zone.ts 8b4e4cd06ed8e1c6a7c87f635151a090fa9b9ef42a28be125a205aaa62a92d2f
libs/shared/prompt-trust/src/trust-zone.spec.ts 7cb24860677a17a2c17c7165d36eacbe8c829d7a8c5c110cc9fa07813037208c
libs/v6/clio-story/src/index.ts 17944a592fbe5a04200c337eb9c9f426bff59fbd99d34c7e6a0296c0ff192c69

Honest boundary#

This evidence closes the local measurement and disclosure requirement. It does not claim live provider traffic, a reduced fallback rate, provider reliability calibration, WebTransport/SFU deployment, vendor ASR/TTS, V1 promotion, OTLP export, Themis dispatch, in-run human co-authoring, mixed-lineage publication, or the remaining V6-002 continuity, consent, latency, failover, safety, and incident-response exercises.