Ground truth
The screenshot is the exact analysis frame the ground truth is computed on. Labels are imported (AtariARI / OCAtari) then verified causally on the bit-exact machine.
The exact ground truth for this ROM and how every interpretability method scores
against it — one of the 42 scored games. Every number is read from
site_data.json (games.boxing).
← back to the Ground Truth ROMs · Paper 2 audit · method catalogue
The screenshot is the exact analysis frame the ground truth is computed on. Labels are imported (AtariARI / OCAtari) then verified causally on the bit-exact machine.
Each XAI / mechanistic-interpretability method's faithfulness on this ROM, sorted by F (click a header to re-sort). The left border marks the tradition: green = causal / intervention, red = gradient / correlational, amber = other — the split the paper's headline turns on. F is all-regime; content-F and position-F are the two output regimes (n/a where this game has no record for that regime).
| Method | Family | F (all) | content-F | position-F | S | M |
|---|---|---|---|---|---|---|
| A2 · Single-unit lesions | causal / intervention | 1.000 | n/a | n/a | 0.489 | 1.000 |
| A8 · Whole-state recording | descriptive | 1.000 | n/a | n/a | 1.000 | 0.070 |
| Activation patching / causal mediation | causal / intervention | 1.000 | n/a | 1.000 | 1.000 | 1.000 |
| Causal scrubbing | causal / intervention | 1.000 | n/a | n/a | 1.000 | 1.000 |
| Interchange interventions / DAS | causal / intervention | 1.000 | n/a | n/a | 1.000 | 0.650 |
| Path patching / IOI circuit | causal / intervention | 1.000 | n/a | n/a | n/a | n/a |
| ACDC — automatic circuit discovery | causal / intervention | 0.867 | n/a | n/a | 0.727 | 1.000 |
| KernelSHAP / Shapley | gradient / correlational | 0.853 | 0.817 | 0.889 | 0.750 | 0.795 |
| Occlusion | causal / intervention | 0.847 | 0.817 | 0.877 | 0.792 | 1.000 |
| LIME | gradient / correlational | 0.844 | 0.817 | 0.870 | 0.667 | 0.795 |
| On-distribution counterfactual | causal / intervention | 0.809 | 0.817 | 0.801 | 0.750 | 1.000 |
| A3 · Tuning curves | gradient / correlational | 0.681 | n/a | n/a | 0.728 | 1.000 |
| RISE | gradient / correlational | 0.611 | 0.798 | 0.424 | 0.583 | 0.341 |
| Extremal / meaningful perturbation | causal / intervention | 0.595 | 0.582 | 0.609 | 0.750 | 1.000 |
| NMF/PCA dictionaries | dim_reduction | 0.502 | n/a | n/a | 0.830 | 1.000 |
| Expected Gradients | gradient / correlational | 0.487 | 0.817 | 0.156 | 0.542 | 1.000 |
| Grad×Input / DeepLIFT | gradient / correlational | 0.487 | 0.817 | 0.156 | 0.542 | 1.000 |
| Integrated Gradients | gradient / correlational | 0.487 | 0.817 | 0.156 | 0.542 | 1.000 |
| Vanilla gradient (saliency) | gradient / correlational | 0.487 | 0.817 | 0.156 | 0.542 | 1.000 |
| SmoothGrad | gradient / correlational | 0.487 | 0.817 | 0.156 | 0.542 | 1.000 |
| A4 · Pairwise correlations | gradient / correlational | 0.469 | n/a | n/a | 0.144 | 0.464 |
| Guided Backprop | gradient / correlational | 0.408 | 0.817 | 0.000 | 0.500 | 1.000 |
| A7 · Dimensionality reduction (NMF/PCA) | dim_reduction | 0.400 | n/a | n/a | n/a | n/a |
| Logit / tuned lens | causal / intervention | 0.291 | n/a | n/a | n/a | 0.081 |
| A5 · Local field potentials | gradient / correlational | 0.228 | n/a | n/a | n/a | n/a |
| A1 · Connectomics / data-flow graph | causal / intervention | 0.111 | n/a | n/a | 0.891 | 1.000 |
| A6 · Granger causality | gradient / correlational | 0.083 | n/a | n/a | 0.818 | 1.000 |
| Attribution / edge patching | gradient / correlational | 0.000 | n/a | n/a | 0.941 | 1.000 |
| Linear probing + control tasks | probing | 0.000 | n/a | n/a | n/a | 0.625 |
| Sparse autoencoders | dim_reduction | 0.000 | n/a | n/a | 0.215 | 1.000 |
Source: site_data.json
· generated by gen_site_data.py from the
committed §R records + leaderboard.json.