Skip to content

inference-local-steering.md

Latest commit

 

History

History
251 lines (182 loc) · 10.6 KB

inference-local-steering.md

File metadata and controls

251 lines (182 loc) · 10.6 KB
Raw

Local SAE Steering Inference Readiness

Status: inspection record for Issue #31. This document records current repository evidence for a Neuronpedia-compatible local steering path. It does not implement an inference server and does not claim runtime readiness. Status: inspection and registration record. This document records current repository evidence for a Neuronpedia-compatible local steering path. It does not implement an inference server and does not claim runtime readiness.

Summary verdict

Agent Machine is not currently ready to run local SAE steering for MODEL_SOURCESET=gpt2-small.res-jb.

The repository has a backend-neutral InferenceProvider contract and a probe-only llama.cpp provider example, but no current repository artifact registers:

  • MODEL_SOURCESET=gpt2-small.res-jb
  • GPT-2 Small SAE artifacts
  • Gemma 2B or Gemma 9B steering sourcesets
  • a Neuronpedia-compatible POST /steer endpoint
  • an activation command that starts a steering inference server

This is consistent with the repository status: provider activation is explicitly not implemented, and production readiness remains blocked by provider discovery and controlled provider activation.

Repository evidence inspected

Inspected on 2026-05-20 against main at commit c306852f26f4fbe62421a9b7801a59815d4daf9e. The repository now has:

  • a backend-neutral InferenceProvider contract
  • a probe-only llama.cpp provider example
  • a local /steer contract stub from Issue #32
  • registered steering sourceset records from Issue #33

The repository still does not have:

  • downloaded model artifacts
  • locked SAE artifact digests
  • storage receipts for model/SAE artifacts
  • a controlled activation path that starts a steering inference server
  • real activation injection into a model forward pass

This remains consistent with the repository status: provider activation is not implemented, and production readiness remains blocked by provider discovery and controlled provider activation.

Repository evidence inspected

Initial inspection occurred on 2026-05-20 against main at commit c306852f26f4fbe62421a9b7801a59815d4daf9e. Sourceset registration was added after Issue #32 closed.

Relevant repository evidence:

  • README.md describes Agent Machine as a bootstrap runtime-control substrate and marks provider activation as not implemented.
  • BOOTSTRAP_STATUS.md states activation remains dry-run control-plane evaluation and does not start a provider.
  • docs/index.md lists InferenceProvider as a contract family but keeps production blocked by provider discovery and controlled activation.
  • contracts/inference-provider.schema.json defines backend-neutral provider metadata and API surfaces such as OpenAI-compatible chat/completions/embeddings and native HTTP. It does not define a Neuronpedia-compatible /steer API surface.
  • examples/asahi-llama-cpp.inference-provider.json is status: probe-only, uses llama.cpp, and exposes an endpoint/health path for llama-server; it is not a steering server.
  • Repository search found no current references to MODEL_SOURCESET, gpt2-small, res-jb, Gemma, or /steer as implemented local steering surfaces.
  • contracts/inference-provider.schema.json defines backend-neutral provider metadata and API surfaces such as OpenAI-compatible chat/completions/embeddings and native HTTP.
  • examples/asahi-llama-cpp.inference-provider.json is status: probe-only, uses llama.cpp, and exposes an endpoint/health path for llama-server; it is not a steering server.
  • docs/local-steer-endpoint.md defines the local /steer contract and stub behavior.
  • contracts/steering-sourceset.schema.json defines the sourceset registry shape.
  • examples/steering-sourcesets/ registers GPT-2 Small and Gemma 2B candidates by reference only.

Current sourceset readiness

Sourceset Current repo status Ready today? Reason
gpt2-small.res-jb Not registered No No sourceset manifest, no GPT-2 Small SAE artifact reference, no /steer endpoint, and no activation path.
Gemma 2B Not registered No No Gemma sourceset manifest, no SAE artifact reference, no /steer endpoint, and no activation path.
Gemma 9B Not registered No No Gemma sourceset manifest, no SAE artifact reference, no /steer endpoint, and no activation path.
gpt2-small.res-jb Registered by reference No Sourceset record exists, but artifact digests, storage receipts, policy/grant admission, activation injection, and smoke proof are missing.
gemma-2-2b.pt-res Registered by reference, access-sensitive No Sourceset record exists, but model terms/access, artifact digests, storage receipts, policy/grant admission, activation injection, and smoke proof are missing.
Gemma 9B Not registered No Not admitted in Issue #33; register only after hardware/runtime profile and access posture are settled.

Weight and artifact requirements

GPT-2 Small

Model weights:

  • Hugging Face model: openai-community/gpt2
  • Public model page: https://huggingface.co/openai-community/gpt2
  • Observed license on model page: MIT
  • Access posture observed from the public page: no account-gating was evident during inspection

Still required before readiness:

  • explicit Agent Machine sourceset record for gpt2-small.res-jb
  • exact GPT-2 Small model artifact reference and digest policy
  • exact SAE artifact source and digest policy
  • runtime storage/cache placement receipts
  • activation policy and agent-registry grant examples for the steering provider

Gemma 2B / Gemma 9B

SAE artifacts:

  • Hugging Face repo: jbloom/GPT2-Small-SAEs-Reformatted
  • SAELens release: gpt2-small-res-jb
  • SAE id: blocks.6.hook_resid_pre
  • Neuronpedia layer: 6-res-jb

Still required before readiness:

  • exact artifact digest lock
  • storage/cache receipt
  • policy admission
  • agent-registry grant
  • controlled activation injection
  • local smoke record

Gemma 2 2B

Model weights:

  • Hugging Face model: google/gemma-2-2b
  • Public model page: https://huggingface.co/google/gemma-2-2b
  • Hugging Face model: google/gemma-2-9b
  • Public model page: https://huggingface.co/google/gemma-2-9b
  • Observed license on model pages: Gemma license

Access posture:

  • Treat Gemma access as terms-sensitive until an operator verifies availability under the relevant Hugging Face account and accepts any required Google terms.
  • Do not claim Gemma readiness from public model-card existence alone.

Still required before readiness:

  • explicit Agent Machine sourceset record for each Gemma target admitted to scope
  • exact model artifact references and digest policies
  • exact SAE artifact sources and digest policies
  • hardware/runtime profile decision for Gemma model size
  • activation policy and grant examples
  • License/access: Gemma terms; operator must verify account access and accept applicable terms before runtime use.

SAE artifacts:

  • Hugging Face repo: google/gemma-scope-2b-pt-res
  • Artifact path: layer_20/width_16k/average_l0_71

Still required before readiness:

  • operator access/terms verification
  • exact artifact digest lock
  • storage/cache receipt
  • policy admission
  • agent-registry grant
  • controlled activation injection
  • local smoke record

Expected local endpoint shape

Target local endpoint for Noetica compatibility:

POST http://localhost:8080/steer

Expected request payload shape:

{
  "prompt": "Write one short sentence about Paris.",
  "model_id": "gpt2-small",
  "steering": {
    "feature_id": "10200",
    "layer": "6-res-jb",
    "strength": 5,
    "preset": "optional"
  }
}

Expected response payload shape compatible with Noetica SteeringResult:

{
  "status": "applied",
  "baseline": "baseline text",
  "steered": "steered text",
  "diff_summary": "short description of observed steering effect",
  "feature_id": "10200",
  "layer": "6-res-jb",
  "strength": 5
}

Allowed status values must align with Noetica steering semantics:

  • applied: a real steering backend applied the requested activation steering.
  • not_configured: required model/source/SAE/backend configuration is absent.
  • noop: the endpoint accepted the request shape but deliberately applied no runtime intervention.

This endpoint shape is a target contract. It is not implemented in Agent Machine at the time of this inspection. Issue #32 provides a stub endpoint that can return not_configured or noop. applied remains blocked on Issue #34.

Current Agent Machine commands

Current safe bootstrap commands from existing docs:

agent-machine version
agent-machine paths
agent-machine doctor --format json
agent-machine probe --format json
agent-machine render plan examples/local-podman-llama-cpp.agent-pod.json --pretty

Local endpoint contract stub from Issue #32:

agent-machine steer serve-stub --host 127.0.0.1 --port 8080 --status not_configured

Current validation command:

make validate

No current Agent Machine command was found that starts a Neuronpedia-compatible local /steer server.

Follow-up issues opened

  • Issue #32: define the Neuronpedia-compatible local /steer endpoint contract.
  • Issue #33: register local steering sourcesets for GPT-2 Small and Gemma.
  • Issue #34: implement controlled activation for a local steering inference provider after endpoint and sourceset records exist.

Readiness conclusion

Current status: not ready.

MODEL_SOURCESET=gpt2-small.res-jb is not loadable today from the repository state. The self-hosted M2b path requires at least:

  1. local /steer contract definition
  2. sourceset registration
  3. model and SAE artifact references with digest policy

Follow-up issues

  • Issue #32: local /steer endpoint contract — closed.
  • Issue #33: steering sourceset registration — current.
  • Issue #34: controlled activation for a local steering inference provider — still required.

Readiness conclusion

Current status: registered but not loadable.

MODEL_SOURCESET=gpt2-small.res-jb now has a registry record, but it is not loadable today from the repository state. The self-hosted M2b path still requires at least:

  1. artifact digest locks
  2. storage receipts
  3. policy/grant admission records
  4. controlled activation path
  5. evidence and grant wiring
  6. successful local smoke record

Until those are complete, Noetica should use NEURONPEDIA_BASE_URL=http://localhost:<port> only as a configurable endpoint target for future experiments, not as proof that Agent Machine can already perform local SAE steering. Until those are complete, Noetica should use NEURONPEDIA_BASE_URL=http://localhost:<port> only as a configurable endpoint target for contract and UI testing, not as proof that Agent Machine can already perform local SAE steering.