# Jev Bench

Jev Bench is a classifier workbench at https://jev-bench.cocode.dk/.
The results page at / declares native WebMCP tools through document.modelContext.registerTool.
Use its tools when WebMCP is available; the page also has an experiment selector and metric buttons, and an accuracy model selector.

- describe: describes the page, selected experiment, and available controls.
- list_runs: lists the published runs available in the experiment selector.
- get_results: reads the selected run's latency, accuracy counts, errors, and returned model names.
- get_accuracy: reads the selected model’s accuracy percentage, macro F1, per-class scores, and each incorrect classification.
- select_accuracy_model: chooses {"model_name":"..."} from the selected run and returns its displayed accuracy details.
- select_run: selects a published run using {"run_id":"..."} and returns the displayed results.
- set_metric: chooses {"metric":"median_ms"}, {"metric":"mean_ms"}, or {"metric":"p95_ms"} and returns updated results.

Read tools take {}. Responses always have {ok, error, data}; invalid inputs return ok:false and data:null.
The page uses local published data; its tools do not call classification services or need API credentials.
WebMCP is currently an experimental browser feature; the normal page works without it.

Repository and reproducible experiments: https://github.com/cocodedk/jev-bench
Published data: https://jev-bench.cocode.dk/data/runs.json