eval/

Package domain holds the evaluator's vocabulary: problems, attempts, what the browser observed, and the scorecards built from them.

Package resource declares what the evaluator needs from the outside world: where problems come from, the browser, and where results are kept.

Package service declares the evaluator's use cases: what the command line asks for.

Package probe decides what the UI probes found.

Package report turns attempts into scorecards and summaries.

Package run asks problems through the browser and records what happened.

Package score grades answers and builds the prompt.

Package uicheck judges one attempt's observations against the passive UI checks U-01 to U-08.

Package config loads and validates the evaluator's settings.

Package devtools drives the Web UI through Chrome DevTools MCP: it starts the server, calls its tools, and reads what they print.

Package gaia provides the evaluator's problems: GAIA from local parquet files (downloaded by fetch.go), and a synthetic set for the self-test.

Package safety holds the checks that keep data and secrets out of places they must not go.

Package store keeps a run's files on disk: its results one row at a time, its description, what the probes found, its summaries and its evidence.

Package synthetic holds the evaluator's self-test set: thirteen invented problems (none of it is GAIA content), the answers the scripted agent gives to them, and the results a correct evaluator must find.