AI evaluation becomes decision-ready when metrics, weights, minimums, slices, evidence, sample sufficiency, worst-slice behavior, and stable ranking remain visible instead of collapsing into one headline score.
Research, creation, decisions, audits and plans are publicly structured without claiming that every release gate is complete.
This page exists so the durable record does not depend on conversation history. Research, decisions, generated artifacts, audits, limitations, media metadata, plans and canonical source paths are retained here as part of the build itself. A listed repository path proves that an implementation record was retained; it does not independently support a real-world factual claim.
0 of 5 listed retained paths have a public reader link on this archive page. Unlinked paths remain named for traceability but are not represented as publicly inspectable evidence.
data/build-038-engine-v1.jsonVersioned metric, weight, minimum, required-slice, evidence, and stable-ranking rules over a synthetic fixture.
Changes weights and required slices, enforces minimums, reranks models, and exports locally.
Changes the visible model ranking as consequence, weights, and slices change.
Reusable metric/slice infrastructure for agent inspection, stop rules, and convergence.
Title, job, B-Web, and lineage are frozen canon; metrics, weights, slices, scores, thresholds, rankings, and copy are derived.
Evaluation remains tied to consequence, decision evidence, and release readiness.
A high average cannot bypass a required slice or per-metric minimum.
Results do not establish benchmark generalization, safety, model quality, or deployment authorization.
data/build-038-archive-v1.jsonRoutes, exact lineage, privacy boundary, and deferred gates.
data/build-038-privacy-support-v1.jsonSynthetic browser-local evaluation with no live model data.
data/build-038-freshness-performance-v1.jsonVersioned deterministic offline-after-load operation.
tests/build038.spec.tsComparable and missing-slice states, local export, 320px, and accessible B reranking.
Authenticated live browser, physical-device/manual accessibility, GitHub-hosted CI, domain evaluation expert review, and LinkedIn media certification remain separate.
039, 040, and 100 consume cap:038.
CURRENT VERIFIED MEDIA / 30-second film · 124.7 KB
FINAL PROPOSITION / Bound exactly what the score proves.
SHA-256 92bf272f3ccee414ed72bc845d784e8c5eed262ad6c257ae9b2178ffdb7aa2f7Reproducible from data/linkedin-film-specs-v1.json with scripts/render-linkedin-films.py. The website—not a chat attachment—is the canonical media source.
apps/web/app/100-builds/038/a/EvaluationStudio.tsxapps/web/app/100-builds/038/b/MetricSurface.tsxpackages/release/src/domain-ai-evaluation-engine.tsdata/build-038-engine-v1.jsondata/build-038-archive-v1.jsondocs/builds/038/SPEC.mddocs/builds/038/ARCHIVE.mdtests/build038.spec.tsThese paths preserve the implementation record. Only items with an explicit public reader link are publicly inspectable here. A file path, working interface or internal engine is not independent support for a real-world claim; synthetic propositions and design reasoning remain labeled within the retained record.