Component · spec & evidence

Toy-Transformer Attribution Replay

Records which model features drove an answer, each tied to checkable evidence.

Contract

Does This takes the workflow of "tracing which internal features inside a model drove an answer" and turns it into inspectable local records. Each row links feature ids to a machine-readable graph of connections, records the before/after results of poking those features (the causal-intervention deltas), notes how far the explanation can be trusted (its faithfulness limit), and points to where the underlying evidence lives. The records show that every interpretability claim is backed by checkable evidence, and that they deliberately hold no model weights, no raw activations, no prompts, and no hidden reasoning — they carry only refs, digests, counts, and verdicts.

Scope limit It validates only the declared public circuit-attribution runtime-result record contract. It excludes model-transparency product claims, live model access, export of private weights/raw activations/proprietary prompts/hidden chain-of-thought, external model access, benchmark claims, or public sharing/launch.

Run

microcosm mechanistic-interpretability-circuit-attribution-replay run-attribution-bundle --input examples/mechanistic_interpretability_circuit_attribution_replay/exported_circuit_attribution_bundle --out receipts/runtime_shell/demo_project/organs/mechanistic_interpretability_circuit_attribution_replay

Evidence

ClassContract validatorevidence 4/5Real runtime result

research-workflowsforecastingprovider operations

Source

Source Design note · Source

Map Show on the map

Index Open compact index card

Paper module