[JenD86]

Point 1 of community contributions: the library now reads community/vectors.yaml (the repo's copy) instead of the hard-coded BUNDLED_VECTORS.

How a model's vectors are assembled at startup

1. Your own vectors from vectors/<model>/registry.yaml. These win on a name clash.
2. Index entries for the model. verified entries load by default. Unverified community entries load only with SteeringAgent(..., include_community=True); they're capped at ±8 unless a maintainer has calibrated them, and marked [Unverified community vector] in the model's prompt.
3. Auto-extraction, unchanged, if the model ends up with no vectors at all.

Safety of index files

• Each file is fetched at its pinned HF revision, then checked against its sha256 and the model's hidden_size/layer count before it's cached in vectors/<model>/.
• A cached file that already matches is reused with no network call.
• A local file that doesn't match is reported and never overwritten.

registry.yaml stays yours: index entries live in memory only, and save() never writes them. This avoids rewriting a hand-commented, tracked file on every run. A removed or demoted index entry simply stops loading.

Migration

• The hand-written assistant_axis entry is removed from the Qwen registry.yaml, so the index provides it, using the same committed file (sha matches, no download).
• Two warnings from the old hand-written entry ("Strong effect", "can get weird at high magnitude") move into the index entry.
• setup.sh fetches through the index instead of hard-coding the Butanium download.

Also: load_index/fetch_vector move into modulo/community.py, shared by the library, CI and gpu_verify.py. The docs (AGENTS, CONTRIBUTING, README) are updated.

Testing

• 66 CPU tests pass. The 8 new ones cover:
  • verified loading by default, community opt-in with ±8 and the unverified marker
  • reusing a cached file
  • not overwriting a mismatched local file
  • not caching a bad download
  • the model-dimension check
  • your own vectors winning, and index entries never being saved
  • an invalid index being ignored
  • the real repo index serving Qwen's committed file
• GPU (RTX 4090 RunPod, fresh clone): setup.sh fetched assistant_axis through the index and loaded the model; pytest 69 passed (CPU + GPU); a 3-turn run.sh --auto session loaded the vector from the index, steered at -8 and decayed as expected.

🤖 Generated with Claude Code