PlateProof

A manual can look complete on screen and lose an illustration on paper. PlateProof extends MIT Flatpack with a Chromium print gate and an authored eight-plate reference. In the recorded Supervised workflow, seven plates were visible where eight were expected. A one-rule repair restored all eight on one A4 page, with retained PDF and CI evidence. The owner accepted that outcome outside the native flow. The public comparison is a reference build; the repair MR remains unmerged and undeployed.

PlateProof extends the MIT-licensed Flatpack, which turns narrated how-to video into illustrated instructions. The hackathon work adds a deterministic print-quality layer around the existing renderer: an authored eight-step reference, real Chromium print checks, machine-readable findings, and retained PDF, screenshot, layout, HTML, and JUnit artifacts. It checks visibility, image decoding, clipping, attribution, and page count. A missing browser, invalid input, or missing plate produces an explicit failure. The required reference checks run offline without a generation provider key or paid model calls.

The central demonstration is a native workflow with a bounded outcome. GitLab Duo session 9744278 diagnosed the failure, repaired it, verified the candidate, published the branch, and observed CI without an intermediate human checkpoint. It produced commit 4172ab98 on draft Showcase MR !53. Pipeline 2967288455 passed all six required jobs and retained the eight-plate, one-page result. On 20 October, the owner answered “Approve final print result” in Codex after the run. That was genuine outcome acceptance outside the native flow, not a GitLab approval-control click or permission to merge or deploy.

GitLab is essential to the workflow: Duo performs the bounded repair, merge requests retain the proposed change, and CI ties its checks and artifacts to the exact candidate revision. The evaluator, fixtures, and evidence are kept separate from the permitted repair surface. Unit checks, print verification, security checks, packaging, and artifact preview provide different kinds of evidence; a green result never grants production authority. The native candidate remains unmerged and undeployed, with production approval and monitoring outside the demonstrated outcome.

The public Vercel application lets judges inspect the reference manual, read its executed report, download the A4 PDF, and follow the evidence ledger. Its comparison page shows genuine screenshots from an earlier MR !2 failure and repair. Those images are explicitly historical, not screenshots from session 9744278. The deployed reference is source d9929621, backed by its own CI artifact job; it is not the unmerged native candidate. The newest local UI polish, including additional inspection interactions, is not claimed as publicly deployed.

The engineering challenge was keeping these identities clear while making the result easy to judge. Authored illustrations, rendered print evidence, agent execution, owner acceptance, and production release answer different questions. PlateProof preserves those distinctions instead of letting a single green badge stand for all of them. Structural checks do not certify that an illustrated action is understandable or safe, and physical printer output has not been independently verified. The demonstrated payoff is precise: a reproducible missing-step failure, a native repair, and an inspectable successful print artifact. The next step is broader manual coverage and a separately authorized production release.

What we added during the hackathon

PlateProof extends the existing MIT Flatpack project at https://github.com/himanshu748/flatpack, imported from baseline 90a65b7ff98bf7434897a7d47414c2e517d37b47 with history, license and third-party media notices preserved. Flatpack already supplied narrated-video-to-illustrated-manual generation and its print renderer. New hackathon work added eight original authored reference plates, validated fixture contracts, real Chromium print/PDF quality checks, retained JSON/JUnit/HTML/layout/image artifacts, controlled missing-plate regression coverage, GitLab CI/security/package jobs, a native Supervised repair workflow and a public read-only evidence viewer. The reference fixtures are not model-generated output. Current Vercel source d9929621d195a0341ad66a67701087e3b4fad13e is a separate published reference, not native candidate 4172ab98. The latest local seven-app UI refresh is not included in claims about public deployment. The demonstrated improvement is the executable quality-and-repair workflow around an existing manual generator; no new Livepeer generation or production deployment is implied.

The result is checkable on paper: eight plates on one A4 page, with the PDF and CI evidence retained. GitLab Duo carried out the print repair; Codex helped with engineering and evidence assembly, and Claude reviewed the work read-only. Deepgram narrates edited captures from the recorded workflow. Owner acceptance, the unmerged candidate and the public reference remain separate records.

Lifecycle stages demonstrated: verify, secure, package. Release and monitor are not claimed.

Built With

• chromium
• docker
• gitlab-ci/cd
• gitlab-duo-agent-platform
• javascript
• node.js
• pdf-lib
• playwright
• vercel