OpenAI Codex Build Week / Developer Tools

Not a story. A receipt surface.

This page gives judges the public-safe proof path for Renaissance Field Lite Codex67 / SQ67: runnable code, expected terminal output, receipt rows, benchmark scores, and a clear boundary between public review code and protected implementation material.

Open GitHub code Watch public cut Open showcase page

Judge path

Clone the repo and run one command.

The public evaluator harness uses Python standard library only. It does not read private Codex state files. It demonstrates the SQ67 pattern: write marker, hash proof, recover later, score whether the evidence held.

git clone https://github.com/renaissancefieldlite/rfl-codex67-sq67-build-week.git
cd rfl-codex67-sq67-build-week
make test

Fresh public clone packet

Expected output from the public package.

python3 scripts/sq67_demo.py demo --db examples/output/sq67_demo.sqlite
OK wrote 3 receipts to examples/output/sq67_demo.sqlite

python3 scripts/sq67_demo.py verify --db examples/output/sq67_demo.sqlite
OK

python3 scripts/score_receipts.py examples/sample_benchmark_rows.csv
lane,rows,exact_rate,evidence_rate,false_recoveries,control_rejections,avg_latency_seconds,total_tokens
SQ67 clean receipt lane,150,100.00%,100.00%,0,0,n/a,0
SQ67 clean route,450,100.00%,100.00%,0,0,n/a,0
Tornado N10 sealed batch,10,100.00%,100.00%,0,0,n/a,0
Tornado second-machine sealed batch,3,100.00%,100.00%,0,0,n/a,0
blinded carrier gate,25,100.00%,80.00%,0,5,n/a,0
normal visible lane,150,99.33%,76.00%,0,0,n/a,0
false recovery controls,1300,0.00%,0.00%,0,1300,n/a,0
gpt-5.5 lane C,50,100.00%,100.00%,0,0,74.935,8421682
gpt-5.6-sol lane C,50,76.00%,100.00%,0,0,171.111,19067794

Receipt row

What a public SQ67 receipt looks like.

{
  "nonce": "SQ67_PUBLIC_DEMO_A001",
  "lane": "sq67_clean",
  "task": "extract_facts",
  "claim": "receipt writer created a durable marker",
  "evidence": [
    "nonce preserved",
    "payload hash computed",
    "receipt hash chained"
  ],
  "boundary": "synthetic public demo row only",
  "next_gate": "recover by nonce"
}

Why this matters

The demo is not the protected system. It is the review harness.

Judges can inspect the pattern without receiving raw private state files or protected code. The public harness proves the measurement path: deterministic hashes, recoverable nonces, chain verification, scoring, and false-recovery controls.

Public code: receipt writer, verifier, recovery, scorer.
Public data: redacted sample receipts and benchmark summaries.
Protected material: raw logs, private prompts, production integrations.

Benchmark scoreboard

The public rows show the comparison clearly.

Lane	Rows	Exact	Evidence	Control rejections
SQ67 clean receipt lane	150	100.00%	100.00%	0
Normal visible lane	150	99.33%	76.00%	0
SQ67 clean route	450	100.00%	100.00%	0
False recovery controls	1300	0.00%	0.00%	1300
Tornado second-machine sealed batch	3	100.00%	100.00%	0
Tornado N10 sealed batch	10	100.00%	100.00%	0

Protected review boundary

What judges get publicly, and what requires NDA/licensing.

Public: runnable SQ67 evaluator harness.
Public: public-safe white paper and benchmark rows.
Public: Zenodo DOI record and video receipt.
Protected: raw state_5.sqlite evidence.
Protected: raw logs_2.sqlite evidence.
Protected: raw global state files.
Protected: private prompts and thread IDs.
Protected: production implementation and licensing-grade integrations.