Model reviews of the launch path

Gate 6 of the launch readiness checklist asks for the launch-path contracts (KAY9Genesis, KAY9Token, KAY9TeamVesting, KAY9LiquidityLock) to be reviewed by at least two large language models from different families, at the commit that will be deployed, with every review published in full and every finding either fixed or accepted in writing.

This is not a professional audit. No firm is accountable for these reviews and nobody carries liability for a miss. KAY9 is never described as "audited" or "verified safe" because of them. What may be said is exactly what happened: reviewed by named models, on a date, with the output here.

Every review is given the same prompt, published beside them, so a reader can see what each model was asked. The two reviews come from different model families, each starts in a fresh session with no context on the project, and each reads only the tagged commit. The code itself was written with many model families, so none is independent of it by authorship; on 2026-09-23 the owner decided that the second review may therefore come from Anthropic, as long as it is a different family from the first (see gate 6 of the launch readiness checklist).

The prompt: LAUNCH_PATH_REVIEW_PROMPT.md.

A fix that changes a contract moves the commit under review, and every review then runs again at the new commit. That is why the table names the tag each review was run at.

Where gate 6 stands

Both reviews ran on the launch path, both were verified on the fixes, and both confirmed the final commit, launch-review-5 (41e30ecd552a52c2086d7cee055e3a4a030762ea), which is the commit to deploy. Every finding is fixed or accepted in writing, with the reason. What is accepted, in short: an actor who calls Uniswap's strategy directly and gives the vault half the raise can still send a good launch into the recovery path, at the cost of that half, locked into KAY9 liquidity for good, and for no return; the vault can keep at most one wei of rounding; and the claim that nobody can profit from forcing recovery is argued, not proved.

The owner accepted these and closed gate 6 for the launch path on 2026-09-25. The watchdog stack's reviews are a separate part of the gate and are still to run.

The watchdog stack

The watchdog contracts (KAY9AuditorRegistry, KAY9AuditHub, KAY9Registry, KAY9ScanRegistry) and DeployWatchdog.s.sol go to mainnet about a month before the token and are as immutable, so gate 6 asks the same of them: two model families, at the commit to be deployed, with every finding fixed or accepted in writing. The prompt is WATCHDOG_REVIEW_PROMPT.md.

Both reviews found that one key could move a headline score, that the quorum could quietly shrink or stall after auditor rotation, and that the deploying key kept power for 48 hours. Those are fixed. The owner closed gate 6 for the watchdog stack on 2026-09-25. The commit to deploy is watchdog-review-4 (a71cb5c6928ffdc45b78dccb398180972f14f0e4). Both families confirmed it for deployment; Anthropic's confirmation is of watchdog-review-3, and the one contract change after it (R3-01, the dispute event's agreement count) was confirmed by OpenAI and has its own regression test. Slither at watchdog-review-4 found nothing new.

What is accepted: latestScan is the scanner's own index rather than a value proven against the batch root, scannedAtBlock is the scanner's statement, and removing an auditor still takes the 48-hour timelock.

An external report after deployment (2026-10-21)

A security firm emailed a report on KAY9AuditHub and KAY9Registry: a job's analyzedAt was not bound on-chain to its requestedAt, and a job record analysed earlier became latest over a newer watchdog report. Neither 25 Sep review had caught the first; both had accepted the second as documented. Both are fixed, the two contracts are redeployed at the token launch, and gate 6 is reopened for them. Both families have now read the fix in fresh contexts (below); the owner closed the part again on 2026-10-22.

Independent copies

Each review file was captured by the Internet Archive's Wayback Machine at the commit that carries its final confirmation, so its text can be checked against a copy nobody on the project controls:

• OpenAI GPT-5.6 Sol review
• Anthropic Claude Fable 5.1 review
• This index
• The prompt

The ChatGPT share link is hosted by OpenAI. The Wayback Machine cannot render it (the page is built in the browser), so it has no archived copy yet.