Skip to published state

OpenAI Build Week / Developer Tools

public evidence

Canopus: Bounded Research for Codex

Give Codex a mission. Verify the work. Keep humans in authority.

Let agents do the work without letting them declare truth.

A verified 7,194-point Sidon witness

run_f68e4cfc-e5c7-4c73-86cb-d79807c47ec4
success
pass
Defer
0

Produced the exact reconstructed 7,194-point Sidon witness for {0,1}^24. GPT-5.6 found an exchange that improves the mechanically verified, unaccepted 7,193-point seed by one point.

The frozen verifier checked all 25,880,415 unordered componentwise pair sums for uniqueness. Clean-clone replay matched. Vela retained the evidence as a pending proposal without changing accepted scientific state.

Inspect the full evidence chain

Production is not verification. Verification is not acceptance.

Canopus gives Codex a preregistered, bounded mission. A separate verifier checks the artifact. Vela records a Receipt and routes the proposal. Defer means the evidence is retained while the protected human decision remains untouched.

  1. 01Canopusmission, custody, bounded worker
  2. 02Verifierindependent deterministic check
  3. 03VelaReceipt, policy route, human authority

A second implementation passed and its collision probe failed closed

auxiliary evidence

A separate JavaScript verifier encoded the binary vectors as exact base-3 integers, sorted all 25,880,415 pair sums, and found no duplicate. The same implementation rejected a deterministic mutation containing a known pair-sum collision.

This strengthens reproducibility. It is explicitly not a durable Vela verifier attachment or a human acceptance decision.

Inspect verifier, report, and CI path
sha256:ddbef08b366138005618c937d732f95361488db22104d86819f5d212b89bf254
sha256:39c3a08515ecd4ad7312a5a04f3d2ea5c0e0f73d00acff8bdfe47d550fab2937

Discovery and custody stayed separate from truth

clean-clone replay matched

What GPT-5.6 did

  • worked only inside a preregistered Sidon mission
  • removed baseline point 72 and added two new points
  • handed off one bounded witness artifact

What GPT-5.6 could not do

The model had no verifier authority, no Vela decision authority, and no human key. The separate verifier passed; signed policy routed the Receipt to defer.

sha256:3b529df4304b96890cdd7eed5f90fd862422b8b1e7b4da1cef0a478e6e7cdebe

The formal GPT-5.6 candidate did not pass Lean

stopped before landing

Before the Sidon success, GPT-5.6 produced one preregistered formal proof candidate. The frozen Lean 4.27.0 verifier found an unsolved inverse goal and sorryAx. Canopus created no Receipt, proposal, route, commit, or frontier mutation.

run_4c2ba5f5-04ac-44d5-adb6-8937eb2ea165
sha256:ef81cbf548d8a08e3811f0aa070b6ce0d58b52792f0bb56b5584dd806da4cb30
sha256:04e31b07889f94a1d205231942bf30fa6c3b27864520b6a42de58604da53e544

Release and deployment identity

b98c846d61d4c554e43388107bcce77c51307db5
sha256:cf2c2ed8d54f68b8adca94f25f7ff2adcb39501fb2e68fa0b5640403dba5266e
4289e05876f142e72af622672e190be26f6a6f1d
825657d7e87618c0aa6fc9af7e3182e05f324750
sha256:11a668d1cdd9caa18f6c3c78ac8c03431bd2ab9beb71f3c62c747d8a7b6571cd