# Brick public skill tables

Public skill vectors consumed by the Brick router. The Hugging Face dataset
`regolo/brick-skill-tables` contains one CSV file, `skill_vectors.csv`, with one
row per model and six capability values in `[0,1]`. Brick uses these values as
cold-start priors, so users do not need to measure a model that is already listed.

The CLI also ships richer JSON copies under this folder for offline initialization.
The Hugging Face dataset is intentionally CSV-only; provenance and detailed
measurement metadata remain in the Brick repository and generated records.

## CSV schema

```csv
model,coding,creative_synthesis,instruction_following,math_reasoning,planning_agentic,world_knowledge
claude-haiku-4-5,0.60,0.60,0.68,0.62,0.58,0.70
qwen3.5-9b,0.58,0.45,0.72,0.75,0.62,0.76
```

The vector order is fixed and must never be changed:

```text
[coding, creative_synthesis, instruction_following, math_reasoning, planning_agentic, world_knowledge]
```

Values are normalized scores, not raw benchmark percentages. They describe
capability priors only; cost and routing preferences belong in Brick
configuration. `creative_synthesis` is generally low-confidence because there
is no reliable public cross-model evaluation for it.

The bundled JSON records may include provenance fields such as `source`,
`confidence`, `support`, and `sources`. The public CSV deliberately contains
only the model id and vector so it stays easy to inspect and consume.

## How Brick uses the table

At startup, Brick first checks its bundled cards. When a fresh card is requested,
the resolver can fetch `skill_vectors.csv` from Hugging Face, find the matching
model row, and cache the result locally. A measured profile from Brick can
supersede a benchmark prior locally.

## Contributing

### Add a model by pull request

1. Fork this dataset and create a branch.
2. Add one row to `skill_vectors.csv`.
3. Keep the exact header and capability order; use a stable model id and values
   between `0` and `1`.
4. Sort rows by `model` and open a pull request against `main`.
5. In the PR description, explain the sources or evaluation method used for the
   vector. Public benchmark links and measured Brick probe results are preferred.

Example row:

```csv
my-model,0.70,0.45,0.72,0.80,0.60,0.75
```

For a measured profile, the recommended local command is:

```bash
brick skills extract my-model --publish
```

This asks for consent and updates the matching CSV row. Manual PRs are welcome
when the model cannot be measured through Brick or when submitting a benchmark
prior for review.