Cassette Build Report 007 — A Specification Can Drift Without Anyone Lying

Scope note: This post covers how I kept Cassette’s research aligned with its original remit while the answers were being written. It is about process design, not a claim that every later requirement is final.

A project can drift without anyone making a false statement.

That was the problem I saw after the queue began to work. Each answer was individually sensible. Together, some of them described an application instead of a model system, a Mac installation instead of a drive-resident model, one physical machine instead of a general product, a paper instead of a working operation, or a short code sample instead of the smallest complete mechanism.

The words were not obviously wrong. The substitutions happened between answers.

I asked for an original-remit document using as much of my own language as possible. I needed an authority that later agents could consult without reconstructing the opening conversation from fragments. The request was not about preserving my phrasing for its own sake. It was about keeping the project from changing its completion condition while research accumulated.

Then I asked for a review inside each research answer. Before a point could close, it had to be compared with the original remit, earlier decisions, and the physical possibility of carrying out the combined instructions. If an answer introduced a new requirement, the answer had to state why the requirement was necessary. If it substituted an easier artifact, device, model, or service, it had to be rejected.

I asked for that correction in the research skill itself. I did not want a sentence in one document telling an agent to remember a sentence in another document. The recurring mistake had to be repaired at the source of the research process.

This changed what a good answer looked like. A research point needed a decision, a scope, a build instruction, an acceptance check, and a reopening condition. It also needed to say what it had not established. The review happened while the point was being authored, not after eighty answers had assembled into a conflicting specification.

The process gave me a way to talk to an agent about drift without reducing the problem to trust. I could point to the exact substitution: “This assumes a Mac-hosted model.” “This turns an example into a named target.” “This proves a paper mechanism, not the complete operation.” The agent then had to trace the change back to its source and either remove it or state the amendment.

That distinction matters to the field because AI systems are good at producing locally coherent work. A chain of locally coherent answers can still move the project away from the user’s purpose. The danger is not only hallucinated facts. It is accumulated interpretation.

The system now treats the remit as a living contract. Later research can clarify it. An explicit amendment can change it. A plausible answer cannot change it by implication.

This is one of the main things I am building with Cassette. The model should not only produce a technically competent answer. It should keep the answer attached to the problem that authorized it.