Concepts
Core objects
| Term |
Definition |
Where it appears |
| Claim |
Mechanistic hypothesis to verify |
hypothesis.jsonl |
| Bundle |
Full artifact package for evaluation |
bundle directory |
| Protocol |
Frozen acceptance policy and thresholds |
protocol.yaml |
| Raw cell |
Intervention/measurement record |
evaluation_result.json |
| Gate |
Deterministic evidence check |
evaluator outputs |
| Gate outcome |
pass, fail, not_evaluated |
claim_reports[*].gate_outcomes |
| Evidence tier |
Final claim classification |
claim_reports[*].evidence_tier |
| Lane |
Discovery source/workflow |
claim metadata |
Supported discovery lanes
| Lane |
Source style |
Typical output |
| A |
Circuit sweep workflows |
ranked component hypotheses |
| B1 |
OpenAI-style neuron explanation |
neuron claims |
| B2 |
Efficient prompt-tuned autointerp |
neuron/feature claims |
| B3 |
SAE feature pipelines |
sae_feature claims |
| C |
Behavioral/Petri-style workflows |
behavior-linked claims |
Supported component types
| Type |
Granularity |
head |
attention head role |
mlp |
MLP block role |
residual_stream |
residual stream contribution |
edge |
source-target component interaction |
mlp_neuron |
single neuron function |
sae_feature |
sparse latent feature function |
das_subspace |
causal subspace effect |
Required files
protocol.yamlhypothesis.jsonlevaluation_result.jsonmanifest.json
Why tri-state outcomes matter
not_evaluated distinguishes missing evidence from contradictory evidence, which prevents misleading boolean-only summaries and improves remediation planning.