Point and Call

Japanese train drivers point at the signal and say what it says out loud. Nothing is added: the signal was already there, already visible, already read. The gesture is the whole intervention, and it cuts errors dramatically. This is the study's result, in one habit.

Two operating contracts for a coding agent, matched to the same roughly 14 KB. Same length, same structure, both read by the agent. The only difference was whether the bytes addressed the task in front of it.

The generic contract left conformance at its 8 percent floor. The one carrying a single project-specific rule lifted it to 100 percent for Codex and 80 percent for Claude.

Why this is the cleanest result in the programme

Because it isolates one variable and nothing else moves. Length held constant. Structure held constant. Reading held constant, since the agent opens the document either way. The entire swing from 8 percent to roughly 100 comes down to whether the sentence was about this job.

That is a much sharper claim than "context helps", and it explains why two earlier studies disagreed about whether an operating contract does anything. They were both right. One was measuring volume, which is inert. The other happened to be measuring relevance, which is everything.

What it asks of you

The uncomfortable implication is that most of the work in writing an agent's environment is not writing. It is knowing which single arbitrary fact about your project the agent cannot possibly guess, and putting exactly that in front of it. The shape of an error body. Whether a model rejects unknown fields. The convention that could have gone either way and did not.

Padding the document does not help. The note is equally clear on the other side: relevance is necessary and not sufficient, and it says where it stops.

The two contracts, the per-agent numbers and the reconciliation of the earlier disagreement are in the research note.