# Run config

The run config names the candidate models for a sweep. It is optional:
with no config file, `quantproof run --pack <dir>` sweeps every model
already in the local Ollama store.

```yaml
# quantproof.yaml
backend: ollama      # "rapid-mlx" or "anthropic"; default ollama
candidates:          # evaluated first, pulled on demand if missing
  - gemma3:1b
  - qwen3:4b
use_local_models: true   # local backends: also sweep everything already available (default true)
```

That is the whole format. Unknown keys are an error so a typo cannot
silently change what runs.

## The rapid-mlx backend

`backend: rapid-mlx` runs the sweep against a local
[Rapid-MLX](https://github.com/raullenchai/Rapid-MLX) server
(OpenAI-compatible, MLX, Apple Silicon only) at
http://localhost:8000. The server holds exactly one resident model per
process, chosen when it starts:

```
rapid-mlx serve <model> --served-model-name my-model:tag
quantproof run --pack <dir> --config rapid.yaml   # backend: rapid-mlx
```

With no candidates list the sweep covers whatever the server is
serving. Models cannot be pulled or unloaded at runtime, so comparing
several MLX models means one server (and one quantproof run) per
model; the stored runs compare across reports. Fit is not predicted
(the server exposes no weight files); peak memory is measured from the
server's own Metal accounting, and the prompt cache is cleared before
every generation so latency measures fresh inference.

## The anthropic backend

`backend: anthropic` runs the sweep against Claude models over the
Anthropic API instead of local Ollama models. It needs an explicit
candidates list of model ids and the `ANTHROPIC_API_KEY` environment
variable:

```yaml
# api-sweep.yaml
backend: anthropic
candidates:
  - claude-haiku-4-5
  - claude-sonnet-4-5
```

```
export ANTHROPIC_API_KEY=sk-ant-...
quantproof run --pack <dir> --config api-sweep.yaml
```

List valid model ids with `quantproof models --backend anthropic`.
`use_local_models` does not apply (there is no local store to merge
in), and VRAM/fit columns render not-applicable because inference runs
on Anthropic hardware; see docs/methodology.md for exactly what is and
is not measured on this backend.

Notes:

- Explicit candidates run in declared order relative to each other, but
  the planner reorders the final set largest-first so out-of-memory
  surprises surface early.
- Ollama cloud models (`*:cloud`) are excluded automatically: nothing
  about them can be measured locally.
- Candidates predicted not to fit in free VRAM are skipped from the
  plan; pass `--force` to attempt them anyway. An attempt that runs out
  of memory is recorded as a result, not an error.
- `quantproof resume` reuses the run's original config exactly. If this
  file (or the task pack) changes on disk after planning, resume aborts
  rather than mixing two configurations in one result set.