feat: add Composer Arena public preview

[seanphan]

Summary

• add a publication-candidate Composer Archive-to-Output Arena route under /bench/composer-arena/
• keep controlled-agent and end-to-end-system standings separate
• add confidence intervals, battle counts, deterministic success/failure, latency, token, cost, and task-slice fields
• add an anonymous A/B replay with swapped order, tie/both-bad choices, reveal-after-judgment, and browser-local votes excluded from rankings
• publish schema-gated synthetic fixtures, methodology, checksums, FineVideo provenance, and the exact real-data handoff contract

Safety and evidence boundary

• every current score/output is prominently marked synthetic and non-rankable
• the verified 24-cell FineVideo release is cited only as source/provenance evidence
• official ranking remains disabled until successful comparable runner records and governed human votes exist
• no customer/GMA data, credentials, signed URLs, hidden annotations, raw reasoning, deployment, or production-default change

Verification

• npm run check:bench
• npm run lint
• npm run build -- --webpack
• local production server: /bench/composer-arena/ returned HTTP 200
• Chrome visual smoke: controlled/end-to-end switch, A/B swap, local tie/reveal, and zero browser warnings/errors

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
av-web	Ready	Preview, Comment	Jul 19, 2026 3:59pm

[seanphan]

Founder decision — HOLD

Sean adopted the 2026-07-19 independent Composer review with amendments. This PR is intentionally returned to draft and must remain open, unmerged, and undeployed.

Reason:

• the v1 seven-family result is now classified as an internal reliability and integration gate, not an end-to-end intelligence benchmark;
• the unchanged mechanical control accepted 5/7 complete outputs, while the best model lane accepted 4/7;
• the pack contains no transcript, OCR, or semantic-event evidence, so it cannot establish that the tasks require agentic intelligence; and
• the most likely public reading would be “deterministic chronology beat the agents,” which does not advance the paid Composer diagnostic.

The branch, checks, hashes, and zero-rank page remain preserved. Do not add models, rerun cases, or revise the evidence on this PR.

Any later publication request requires a separately approved v2 construct-validity packet using only the same seven families, with rights-safe timestamp-bound semantic evidence, mechanical/plan-only/naive-scaffold/Composer-native controls, operator-grade acceptance, native tool compliance before any logged translation arm, and a reviewable before/after reliability result. That packet is design-only today; no execution is authorized.