Registry: add ASA vector for gemma-4-12b-it-Q4_K_M (via atlas asa publish)

[itigges22]

Add ASA control vector for gemma-4-12b-it-Q4_K_M (auto-generated by atlas asa publish)

Summary

User-trained BiasBusters #4 ASA steering vector for gemma-4-12b-it-Q4_K_M,
uploaded to HuggingFace at https://huggingface.co/itigges22/atlas-asa-gemma4-12b.

Verification checklist (maintainer review per PC-061)

• HF link reachable: https://huggingface.co/itigges22/atlas-asa-gemma4-12b
• License is permissive for redistribution (apache-2.0)
• ast_edit_steering.gguf SHA256 matches: fbcfd8af85980c21fd0eb7d37c5627b027008934ad6c626b99eabf6e5587bb46
• Residual dim (3840-dim) matches the base model
• Trained at layer 36 (paper recommendation: ~75% of model depth)
• Spot-check: drop the .gguf at /models/ast_edit_steering.gguf and
  confirm llama-server boots without control_vector_load failed
• Behavior smoke-test: pre-vs-post task that benefits from the bias
  (whole-function rewrite via ast_edit over edit_file)

Suggested registry diff

Most ASA vectors are distributed alongside the Lens artifacts in the
same model entry. Add an asa_artifact_files field if the registry
doesn't already track ASA per-model (V3.1.2 work):

Model(
    name="gemma-4-12b-it-Q4_K_M",
    # ... existing fields ...
    lens_status="supported",
    # New (V3.1.2 forward-compat): ASA vector tracking
    asa_artifact_files=["ast_edit_steering.gguf"],
    asa_status="supported",
    license="apache-2.0",
),

Provenance

Trained locally via atlas asa build against gemma-4-12b-it-Q4_K_M. Algorithm:
mean-difference (positives−negatives) over per-token residuals at
layer 36, projected to a single direction. Same approach as the
Feb 2026 ASA paper (arxiv 2602.04935).

---

Stacked on #121 (this model's entry lands there). Merge #121 with branch deletion and GitHub retargets this PR to dev automatically — its diff shows only this change.

[itigges22]

Superseded by a combined-publish PR (atlas publish, PC-215) — retesting the one-PR flow.