Acceptance review, October 21, 2026

This review checks the released Heval workflow and makes the remaining work explicit. Current public source is 6bd47be. No package was published, no production code was deployed and no shared worker was changed during this review.

Issue disposition

Public issues	Verified evidence	Remaining work
#9 / #16, onboarding	Earlier existing-account hosted Oracle; current setup/evaluation browser checks, reload, keyboard and narrow controls with simulated accounts	Fresh real Clerk signup and isolated worker pairing under that identity. The owner chose to continue without fresh-account acceptance. Keep open.
#11 / #17, report access	Earlier owner/anonymous hosted publication record; 93 current backend tests and 7 export tests; current browser export flow and actual local renderer	Real signed-in editor/viewer sessions, membership removal, history denial and authenticated hosted downloads. Keep open.
#12 / #18, reproducibility	Published commands executed from a clean clone of current public main; passing Oracle below; version/provenance and retention limits; release checklist linked below	Hosted account gaps remain separately tracked in #16/#17 and must remain unchecked in the release checklist.
#19, upgrade/rollback	Exact retained package hashes, four saved passing reports, raw rewards and preserved-file snapshots rechecked; prior installed smoke logs checked	No npm publication or active-worker upgrade is claimed.
#20, current references	Clerk/Harbor 0.24/Bun 1.4.2 pages and current automated UI captures; historical reports preserved	Simulated roles and no-token guidance are labeled, with no claim of full hosted acceptance.

Measured Oracle on current public source

The two shell blocks in the published guide executed on a separate Linux checkout at exact revision 6bd47be82952b2509a83766b04cd394ddc8fc46d. Versions were Node 22.23.3, Bun 1.4.2 and Harbor 0.24.0. The normalized result contains one trial, Oracle 1.0.0, reward 1 and passed 1. Agent time was 8.410 seconds; total trial time was 22.378 seconds. This is a task-verifier setup pass, not a model score.

Sanitized measured record contains the exact task, approval/content/execution digests, attempt limits, resource settings and normalized row. Model is unknown; Oracle uses no inference provider. Cost and token fields remain null. The profile's generic provider label is not observed inference provenance.

The task uses python:3.12-slim. Its executed immutable image digest was not recorded and remains unknown. The direct Harbor command applies task limits but does not enforce the connected supervisor's overall 300-second deadline. Raw artifacts remain privately on the isolated worker under operator control. No automatic retention period or backup policy was established. Model aliases, mutable images and harness installation dependencies limit later reproduction.

Rechecked installed-worker evidence

The October 20 record and structured artifact record describe registry CLI 0.2.0 with Harbor 0.23.0 and source-built CLI 0.2.0 with Harbor 0.24.0 at public dd69e2b. Exact tarball SHA-256 values still match. Each of the initial, upgrade, rollback and re-upgrade saved combined reports has one passing Oracle row and a corresponding raw verifier reward of 1.

Rollback preserved all 59 earlier state files and added 25; re-upgrade preserved all 84 and added 26. Rechecking the recorded before/after hashes found no changed or missing earlier files. The saved logs record 60 CLI tests, installed-artifact smoke and non-root setup smoke passing with Node 22.23.3. Matching health checks, mismatched-runtime guidance and rejection of Node 20 were recorded on October 20. Those transitions were not rerun today; today's observations are artifact and log verification, not new upgrade executions.

Browser and export verification

• scripts/first-smoke-ui-smoke.ts passed setup, real loopback provider UI, simulated pairing/dispatch/results, reload/navigation, skip-link keyboard access and narrow-screen controls.
• scripts/evaluation-ui-smoke.ts passed configuration, restore, inspection, combination, publication affordance and offline/unhealthy/incompatible guidance with simulated auth/Convex.
• bun run test:backend passed 93 tests. This includes owner/editor/viewer permissions, publication revision isolation, membership removal and export access regressions in the test backend.
• bun run test:exports passed 7 tests with mocked cloud transport/storage. The actual membership query and download handler are exercised by the authorization regression.
• bun run test:exports:renderer passed real bundled Node/Chromium PNG and all-question ZIP generation for White and Black styles.
• bun run test:exports:ui initially failed because it required model-logo markup absent from the public neutral style. With the assertion changed to the accessible rendered metric chart, it passed PNG enqueue, save/reload, authenticated mocked download and ZIP enqueue. This changes test expectations only.
• Anonymous production homepage, chart, login, report-library gate and unavailable-link guidance were inspected. A no-token unavailable-link check does not substitute for a revocation test.

Current screenshots and capture method separate live anonymous pages, synthetic account state and measured execution. A blocked collaborative preview prevented reusing its retained signed-in account. No auth bypass or replacement production identity was added.

Release procedure and open account checks

Use the existing coordinated release checklist. It separates default CI from manual Release checks, checks installed tarballs in Node, matches frontend/backend/auth versions, upgrades idle workers, and rehearses rollback while preserving pairing and run state. The hosted acceptance checklist still requires real accounts and downloads. Current source CI and the previous manual release runs are evidence for their named revisions, not proof of fresh-account acceptance.

Use Clerk setup, first evaluation, report access and private hosted exports for the implementation. The legacy Bun fixture deployment guide describes a different execution path.

Earlier delivered status reports remain unchanged. This later review records progress without rewriting their reported AMBER status or claiming the remaining acceptance gates are green.

Tracker update

Closed October 21 with evidence links: #12, #18, #19 and #20. #9/#16 and #11/#17 remain open for the account checks described above. The focused export assertion change is under review in PR #22, not merged into the tested baseline.

The GitHub credential used for closeout lacks read:project, so project-board fields could not be inspected or changed. Issue closure is verified; a corresponding board status is not claimed.