Circuit discovery and feature graphs for mechanistic interpretability (Phase 177.3), in PyTorch, CPU-runnable.
Model-agnostic core#
graph.py — a CircuitGraph DAG of component nodes and candidate edges
(topological order, reverse-topological edge order, subgraphs). model.py — an
AdditiveDAG runner: a real differentiable computational graph where
value[c] = act(bias + Σ W[p,c]·src[p]) and an ablated edge feeds the child its
parent's corrupt-run value. With act=identity the network is linear, giving a
known ground-truth circuit the discovery algorithms are verified against.
Discovery#
- ACDC (
acdc.py, 177.3.1.1) — Conmy et al. Greedy reverse-topological edge ablation; prune an edge when removing it raises divergence-from-clean by ≤ τ. Recovers the minimal circuit. - EAP-IG (
eap_ig.py, 177.3.1.2) — Syed/Rager/Conmy. Integrated-gradient edge attribution along the corrupt→clean input path, scoring every edge from one backward-per-step.score(p→c) = ⟨Δvalue[p], W[p,c]ᵀ·∂metric/∂pre[c]⟩(IG). Dead / zero-signal edges score ~0; top-scored edges match ACDC's kept set. - Transcoders (
transcoder.py, 177.3.1.3) — a wide sparse ReLU replacement for one MLP predicting its output from its input residual, so every residual→feature→output path is a clean, additive feature-graph edge (f_i · W_dec[i]).
Visualisation (177.3.1.4)#
viz.py — a CircuitVisualisation that exports the discovered circuit (nodes by
layer, edges weighted by importance, per-prompt activation traces, downstream
token effects) to a JSON contract and renders it into a self-contained,
offline interactive HTML page (inline SVG + vanilla JS; hover a node for its
activation trace and top token effects). No external assets.
Tests#
python3 -m pytest tests/ -q — 17 tests: topological order + cycle detection,
dead-edge signal check, ACDC recovering the exact ground-truth circuit and
reproducing the clean metric, EAP-IG scoring dead edges ~0 / matching ACDC /
the output-edge completeness identity, transcoder MLP approximation (FVU < 0.15)
- additive feature contributions, and the visualisation JSON/HTML contract.
Not locally actionable#
177.3.2.1–177.3.2.4 (refusal / factual-recall / code-generation / deception
circuit investigations on real trained @nous/llm, @iris/code checkpoints)
require those trained weights and are left unchecked; the discovery tooling they
would use is complete here.