Proof record

The one mission

One autonomous mission has completed end to end in a private repository, with no human between the objective and the merge commit.

• An OpenAI model turned the objective into a bounded assignment naming exactly which files could change.
• An Alibaba model wrote the code inside a network-isolated container holding no credentials.
• Tests ran in a second container with no network at all.
• An Anthropic model, from a different vendor than the one that wrote the code, reviewed the diff and returned a verdict.
• A deterministic gate with no model involved checked every piece of evidence against policy.
• Protected branch checks passed and native auto-merge landed it.

One. Not a product, not a service, not a track record. One.

The failure history

• Models scored their own work and were wrong every single time. When an independent scorer checked the same work, it disagreed on every round. Not most rounds. Every one.
• A swallowed error became a confident claim. Something failed to look at a directory, the failure was discarded, and "I was not able to check" silently became "it is not there". That single pattern produced four false claims before it was found.
• Three models approved a specification written from prose. The one model that actually read the code reversed them. Every round. Agreement between models is not evidence, it is correlated guessing.
• A change passed validation, passed an independent cross-vendor review, passed a deterministic gate, and was still refused at the last step because it touched one file outside its allowlist.

That last one is the argument for all of this. Independent review is necessary and it is not sufficient. The thing that caught it was not a smarter reviewer. It was a dumb check at the boundary that acts, which did not care how confident anyone was.

The honest residual

I cannot verify that the system that verifies things works.

I can see that it produced records, that the records are internally consistent, and that it refused things it was supposed to refuse. An independent review found a real gap in it after everything else had passed, which is evidence the checking works and also evidence it was incomplete.

But I am the wrong person to certify any of it, and there is no version of this where I become the right person. It is checks all the way down and at the bottom there is someone who cannot read the code. Anyone telling you otherwise about their own setup is either an engineer, in which case ask them who checks their work, or they have not thought about it.

I would rather write that sentence than have someone else write it about me.

This site checks itself

This site's rendered copy has a guard. It fails the build on claims this site is not allowed to make, including a false count of completed autonomous work.

That does not make the site important. It is just the same rule in miniature: if a claim matters, a machine should check it before I publish it.