OpenAI Build Week Final v0.2.0 - Blind GPT-5.6 Audit & Safety Preflight

Final competition snapshot for Codex Control Tower.

Public demo: https://youtu.be/EvtguYLSNkg (YouTube displays 2:59)

Highlights:

• Real gpt-5.6-sol through pinned Codex CLI 0.144.3 in an empty, ephemeral, read-only, no-tool workspace.
• Expected answers are withheld; zero tool events were accepted; GPT-5.6 cannot overwrite locked local facts.
• The controlled audit caught missing durable rejected-payment audit proof and raised HUMAN_REVIEW_REQUIRED. That FAIL is the product working; the model invocation and CI both completed successfully.
• Destructive Action Preflight records $HOME/.. as <USER_HOME_PARENT> -> BLOCKED / NOT_RUN / executed:false; no delete command is run.
• Static judge exhibit: https://zyganali-glitch.github.io/codex-control-tower/
• Screenshot and caption pack: DEVPOST_SCREENSHOTS.md

InvoiceFlow Mini is fictional. The scans, two focused tests, hashes, provenance checks, and recorded GPT-5.6 run are real controlled executions. The old openai-build-week-final tag remains the untouched v0.1 baseline.

Reproduce:
npm install
npm run verify
npm run demo:safety
npm run demo:codex

OpenAI Build Week Final — Blind GPT-5.6 Semantic Audit

Final competition snapshot for Codex Control Tower.

• Real gpt-5.6-sol ran through the signed-in Codex/ChatGPT session.
• The audit ran in an empty ephemeral read-only workspace under a fail-closed no-tool contract.
• Recorded result: 3 SUPPORTS, 3 CONTRADICTS, 5 policy alignments, 1 semantic conflict, HUMAN REVIEW REQUIRED, and 1 locked NOT_RUN preserved.
• The controlled fixture verdict is honestly FAIL because durable rejected-payment audit proof and external gates are missing; the model invocation and product CI both completed successfully.
• GitHub Pages is the static no-install exhibit. The repository contains the exact prompt, lifecycle events, hashes, and reconciliation record.

Start with JUDGE_START_HERE.md or the Pages demo linked from the repository homepage.