Compared To Discovery Tools

AutoMechInterp is not a replacement for discovery or intervention frameworks. It is a verifier layer that sits downstream of them.

Tool landscape

Tool / framework	Primary purpose	Typical role in pipeline
TransformerLens	activation inspection, patching, circuit analysis	generate candidate mechanistic hypotheses
nnsight	model internals inspection/intervention at scale	collect intervention traces and candidate explanations
pyvene	declarative intervention experiments	run causal tests and produce structured intervention outputs
SAELens	sparse autoencoder training/analysis	derive feature-centric hypotheses
Captum	attribution/explainability toolkit	produce attribution evidence for hypothesis generation

Layer boundaries

Discovery layer (tools above)

• prioritize hypothesis generation speed and exploratory coverage
• allow diverse intervention and analysis methods
• optimize for finding potentially interesting mechanisms

Verification layer (AutoMechInterp)

• enforce fixed acceptance policy
• evaluate claim evidence under controls/robustness/statistics/integrity
• optimize for reliability and comparability

Why strict separation helps

If discovery and acceptance are merged, standards often drift with method changes.

A separate verifier layer avoids this by keeping acceptance criteria stable while discovery methods continue to evolve.

Integration pattern

1. Run your preferred discovery/intervention tool.
2. Convert outputs into claim bundle artifacts.
3. Evaluate with AutoMechInterp.
4. Use gate diagnostics to design the next discovery experiments.

Mapping tool outputs to bundle artifacts

Bundle artifact	Upstream source examples
hypothesis.jsonl	ranked components, neuron explanations, feature candidates
evaluation_result.json	intervention cells, patching traces, causal test outputs
protocol.yaml	frozen evaluation policy chosen for this run
manifest.json	hash binding over finalized artifacts

Typical failure loop

• Discovery proposes promising claim.
• Verifier rejects due to method sensitivity or controls.
• Discovery reruns with targeted interventions and stronger controls.
• Bundle resubmitted and reevaluated.

This loop is exactly where verification benchmarks add value: they convert “interesting idea” into “evidence that survives scrutiny.”

Practical takeaway

Discovery tools answer “what might be happening?”

AutoMechInterp answers “is the current evidence strong enough under a fixed contract?”

Both are necessary for robust mechanistic-interpretability science.