A (chart, group) pair to mark complete. Used in IngestRequest.complete_groups.

One chart panel inside a benchmark. Each chart is fully self-contained: it owns its own groups, metrics, axes, direction, and ``better`` indi

Thin wrapper around ``httpx.Client`` for the live-bench REST API. Holds a single keep-alive HTTP/1.1 client for the lifetime of the CLI invoc

The body of POST /ingest. ``points`` carry their own ``client_ts`` set by the CLI when each point was produced (NOT when the POST was made —

A point as it crosses the wire from CLI → backend. ``client_ts`` is REQUIRED and set by the CLI at emit time with microsecond precision (``da

Raised when the backend returns a non-2xx response (after retries). ``status_code`` is the HTTP status, ``code`` is the backend's machine cod

A point after the backend has stored it. Adds ``seq`` and ``server_ts``.

One YAML file under ``benchmarks/``. Two supported shapes — exactly one of ``benchmark`` or ``benchmarks`` must be set: 1. **Single benc

Tracks emits per context size. Marks complete when both metrics have reported.

Tracks emits per task. Marks complete when all models have reported.

Tracks emits per context size. Marks complete when both metrics have reported.

Tracks emits per prompt. Marks a prompt complete when all 4 metrics have reported.

Tracks emits per task. Marks complete when all models have reported.

Tracks emits per prompt. Marks a prompt complete when all 4 metrics have reported.

Arbitrary key-value metadata for a (chart, group) pair. Used to attach qualitative data (prompt text, model responses) alongside numeric data

Spawns one long-lived mlx-vlm subprocess and reuses it across calls. The subprocess loads the model exactly once. From then on it stays warm:

Spawns one long-lived mlx-vlm subprocess and reuses it across calls. The subprocess loads the model exactly once. From then on it stays warm:

Spawns one long-lived mlx-vlm subprocess and reuses it across calls. The subprocess loads the model exactly once. From then on it stays warm:

Spawns one long-lived mlx-vlm subprocess and reuses it across calls. The subprocess loads the model exactly once. From then on it stays warm:

Captures every ``IngestRequest`` passed to ``ingest``.

Base model that serialises every ``datetime`` field as ISO-8601 UTC μs. Subclasses inherit the ``model_config`` and the wildcard ``@field_ser