Implement controlled activation for local steering inference provider

[mdheller]

Goal

Implement a controlled local activation path for GPT-2 Small residual-stream SAE steering, using the already-registered gpt2-small.res-jb sourceset.

Current status

PR #38 landed the fail-closed preflight / real-path entrypoint tranche.

PR #39 landed the steering artifact receipt contract and pending fixture.

Issue #34 is not complete. It must remain open until a real status: applied local smoke record exists.

Landed by PR #38:

• requirements-steering.txt optional runtime dependency list
• agent-machine steer preflight --sourceset gpt2-small.res-jb
• agent-machine steer serve --sourceset gpt2-small.res-jb --host 127.0.0.1 --port 8080
• fail-closed not_configured behavior when dependencies/artifacts/receipts/admission are absent
• docs/steering-activation-path.md

Landed by PR #39:

• contracts/steering-artifact-receipt.schema.json
• pending gpt2-small.res-jb artifact receipt fixture
• validation wiring for SteeringArtifactReceipt
• docs/steering-artifact-receipts.md
• artifact receipt gate in docs/steering-activation-path.md

Still not landed:

• complete artifact receipt with exact repo, file path, resolved revision, local path, size, and SHA-256 digest for every model/tokenizer/SAE file
• storage receipt after artifact resolution
• policy admission / agent-registry grant records
• GPT-2 Small model loading
• blocks.6.hook_resid_pre SAE loading
• real activation injection
• real baseline vs steered completion
• local smoke record with status: applied

Scope

This issue is GPT-2 Small only.

Gemma sourcesets are explicitly out of scope for closure because Gemma model access depends on operator acceptance of Google/Gemma terms. Gemma remains registered but blocked until terms/access and artifact readiness are verified separately.

Implement the path needed for:

agent-machine steer serve --sourceset gpt2-small.res-jb
POST /steer

The served endpoint must remain compatible with Noetica's existing local steering client shape:

NEURONPEDIA_BASE_URL=http://localhost:8080
Noetica /api/steer -> Agent Machine /steer

Required implementation work

• Resolve the gpt2-small.res-jb sourceset from the SteeringSourceset registry.
• Fetch or locate GPT-2 Small model artifacts from openai-community/gpt2 with digest verification.
• Fetch or locate the residual-stream SAE artifact from jbloom/GPT2-Small-SAEs-Reformatted / SAELens release gpt2-small-res-jb / SAE id blocks.6.hook_resid_pre with digest verification.
• Emit a storage receipt after successful artifact resolution.
• Load GPT-2 Small.
• Load the blocks.6.hook_resid_pre SAE.
• Run a baseline completion.
• Run a steered completion by injecting the requested SAE feature vector at the specified layer and strength.
• Return a Noetica-compatible SteeringResult with status: applied, baseline, steered, diff_summary, feature_id, layer, and strength.
• Preserve the Issue Define Neuronpedia-compatible local /steer endpoint contract #32 stub path as a no-weights fallback.
• Emit or stage evidence sufficient for a local smoke record.

Acceptance criteria

• agent-machine steer serve --sourceset gpt2-small.res-jb starts the real local steering endpoint when required artifacts are present and verified.
• agent-machine steer serve-stub ... remains available and returns not_configured or noop only.
• If artifacts, storage receipts, grants, policy admission, or runtime dependencies are missing, the real serve path fails closed with a clear diagnostic and does not silently downgrade to fake applied.
• A local smoke record is committed showing:
  • sourceset: gpt2-small.res-jb
  • feature index: 10200
  • layer: 6-res-jb
  • strength: 5
  • prompt
  • baseline text
  • steered text
  • status: applied
  • evidence hash / receipt reference
• Noetica can set NEURONPEDIA_BASE_URL=http://localhost:8080, call /api/steer, and receive the applied response without code changes.
• Native validation / CI passes.

Non-goals

• Do not implement Gemma activation in this issue.
• Do not bypass policy/grant/storage/evidence gates.
• Do not store credentials or tokens in the repository.
• Do not claim production readiness.
• Do not remove the stub endpoint.

[mdheller]

Status update 2026-07-18: Significant progress has been merged — PRs #38/103 (fail-closed preflight, activation entrypoint), #39/107 (artifact receipt contract + pending fixture), #40/104 (artifact resolver), #41/105 (check utility), and #109/110 (sourcesets and steer endpoint). However the issue remains open because the following items are still not landed per the acceptance criteria: real artifact receipt with SHA-256 digest for every model/tokenizer/SAE file; storage receipt after artifact resolution; GPT-2 Small model loading; blocks.6.hook_resid_pre SAE loading; real activation injection; real baseline vs steered completion; local smoke record with status: applied. Do not close until a local smoke record with status: applied exists.

[mdheller]

Status update

The infrastructure is in place (PRs #38, #39, #103). What's still missing before this can close:

• Complete artifact receipt with exact repo/path/revision/sha256 for every file in gpt2-small.res-jb (model, tokenizer, SAE)
• Storage receipt after artifact resolution
• Policy admission + agent-registry grant records
• GPT-2 Small model load and blocks.6.hook_resid_pre SAE load
• Real baseline vs steered completion
• Local smoke record with status: applied

This requires running make validate-all on a machine with the artifacts downloaded. Keeping open until that smoke record exists. Gemma remains explicitly out of scope.

[mdheller]

Blocked on hardware: implementing a controlled activation inference provider requires a device with the relevant steering hardware present and accessible. No such device is available in the current build environment. Leaving open until hardware is procured. No code changes — this issue tracks a hardware dependency, not a software gap.