What we still can't explain

Has the claim been proven? Not yet. The one test that would mean something — the same agent line, the same model, with the instructions versus without — has been run zero times. That is the honest state, not a placeholder.

What you can do in one minute instead: ask your agent for a scorecard — how many technical terms it put in plain English, how many decisions it made for you, how many claims it verified versus assumed.

paired trials run. It needs a model key, supplied by whoever runs it.

entries written from the one line that stumped it — five of the twenty-three signs came from it too.

things this project actually promises — measured by the trial, never assumed.

Why there is no coverage score on the wall anymore

There used to be one: the share of jargon in a sample that already had an entry. It was sealed on the day each line arrived, and it fell twelve points the day a real agent report landed — because of what it was describing. A dictionary we control, scored against a sample we choose, will always catch up. The number was a fact about our own filing, wearing the clothes of a fact about your project.

One line it could not explain

A builder sent over what their agent had just reported on their own project. This is the middle of it, exactly as it arrived:

npm test: 3794/3794 (5 files hit vitest worker-pool startup timeouts under load — a Mac-side flake; re-running those 3 files in isolation: 21/21 pass. Not a code failure.)

Nothing in it is dishonest and all of it may be correct. It is still four separate claims travelling under one green number — a pass in isolation is not a pass, and "a Mac-side flake" is a diagnosis presented as an observation. The reply that settles it is short: run the same command, the same way, twice more.

The four things this project promises

Every term defined in the same breath — no jargon left to look up later.
What was run kept apart from what is only expected — a prediction never wears a result's clothes.
Decisions made for you, named out loud — not left inside the narration.
Anything hard to undo, flagged as such — before it happens, not after.

What only you can measure

We collect nothing about your work, so we cannot tell you what changed for you. You can, in about a minute:

Give me a scorecard for this session: how many technical terms you put in plain English, how many decisions you made for me, how many claims you verified versus assumed, what checks you skipped, and anything you guessed.

What we have not proven

That the paired trial has ever been run, on any line, by anyone.
That the dictionary changes what an agent actually says, versus what it's told to say.
That people outside this project hit the same words we predicted they would.
That the entries themselves are all finished — a meaning without a "what to do" is half an entry.