regolo/brick-skill-tables

Brick public skill tables

Per-model skill vectors consumed by the Brick router. Each <model>.json maps a model id to a 6-dimensional capability vector in [0,1]. brick init reads this folder to seed skill_router.models[].skill_vector for known model ids, so a user never has to re-run inference for a model someone already measured.

This folder ships with the CLI (templates/ is published) and mirrors the Hugging Face dataset regolo/brick-skill-tables, which grows as users contribute measurements for new models (brick skills extract --publish, opt-in).

Capability order (canonical, never reorder)

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

Record schema

{
  "model": "claude-haiku-4-5",
  "provider": "anthropic",
  "capabilities": ["coding", "...", "world_knowledge"],
  "skill_vector": [0.73, 0.65, 0.70, 0.81, 0.50, 0.77],
  "source": "benchmark",          // benchmark | measured | heuristic
  "confidence": ["medium", "low", "medium", "high", "medium", "medium"],
  "imputed_capabilities": [],     // coords filled from the model mean (NA in public benchmarks)
  "support": null,                // {correct,total} per capability — only for source=measured
  "subset_hash": null,            // frozen probe-set hash — only for source=measured
  "date": "2026-06-11",
  "notes": "..."
}

source and trust order

1. measured — produced by brick skills extract running the frozen probe set against the model and grading verifiable categories (math exact-match, code tests, MMLU-Pro letter). Highest trust; carries support and subset_hash.
2. benchmark — derived from published AI-lab benchmarks (SWE-bench, AIME, GPQA, MMLU-Pro, tau-bench, …) normalized to [0,1]. Cold-start prior for frontier closed models. A measured record for the same model overrides it.
3. heuristic — interpolated fallback for an unknown id; marked explicitly.

measured and benchmark vectors are not on the same scale (different question mixes, different grading). Treat benchmark as a prior; prefer measured when both exist for a model.

NA imputation

Public benchmarks systematically omit some capabilities (creative writing is rarely benchmarked; Google does not publish instruction-following or agentic numbers for most Gemini tiers). For a missing coordinate we impute the mean of the model's known coordinates, list the coordinate in imputed_capabilities, and set its confidence to low. This keeps the vector usable by the router while flagging the estimate as soft.

Contributing

Run brick skills extract <model> --publish (opt-in consent prompt) to measure a new model on the frozen probe set and upsert a measured record to regolo/brick-skill-tables. You may also open a PR adding a <model>.json here.