Paper 2 · per-game results

Jamesbond

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.jamesbond).

← back to the Ground Truth ROMs · Paper 2 audit · method catalogue

Jamesbond screen

Ground truth

External label source
OCAtari
Verified causal labels
5
Labels that move a sprite
4
Scorable (enough true causes)
yes
Position regime
yes — a moving tracked sprite at this frame

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.

Every method's faithfulness on Jamesbond

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).

MethodFamilyF (all)content-Fposition-FSM
A8 · Whole-state recordingdescriptive1.000n/an/a1.0000.055
ACDC — automatic circuit discoverycausal / intervention1.000n/an/a0.3851.000
Activation patching / causal mediationcausal / intervention1.000n/a1.0001.0000.909
Causal scrubbingcausal / intervention1.000n/an/a1.0000.500
Interchange interventions / DAScausal / intervention1.000n/an/a1.0000.769
A2 · Single-unit lesionscausal / intervention0.985n/an/a0.6001.000
Logit / tuned lenscausal / intervention0.764n/an/an/a0.105
A5 · Local field potentialsgradient / correlational0.743n/an/an/an/a
Occlusioncausal / intervention0.7240.7470.7000.8930.964
KernelSHAP / Shapleygradient / correlational0.7190.7470.6910.8930.288
LIMEgradient / correlational0.6840.7270.6400.8930.288
RISEgradient / correlational0.6440.6850.6030.8930.288
A7 · Dimensionality reduction (NMF/PCA)dim_reduction0.600n/an/an/an/a
Sparse autoencodersdim_reduction0.580n/an/a-0.0970.875
On-distribution counterfactualcausal / intervention0.5780.7470.4080.8931.000
Grad×Input / DeepLIFTgradient / correlational0.4470.8650.0280.8931.000
Extremal / meaningful perturbationcausal / intervention0.4020.4890.3140.8930.964
Expected Gradientsgradient / correlational0.3880.7470.0280.8931.000
Integrated Gradientsgradient / correlational0.3880.7470.0280.8931.000
Vanilla gradient (saliency)gradient / correlational0.3880.7470.0280.8931.000
SmoothGradgradient / correlational0.3880.7470.0280.8931.000
Guided Backpropgradient / correlational0.3740.7470.0000.8931.000
Linear probing + control tasksprobing0.362n/an/an/a0.800
A4 · Pairwise correlationsgradient / correlational0.272n/an/a0.3450.107
A3 · Tuning curvesgradient / correlational0.188n/an/a0.1300.875
Attribution / edge patchinggradient / correlational0.174n/an/a0.9171.000
A1 · Connectomics / data-flow graphcausal / intervention0.000n/an/a0.981n/a
A6 · Granger causalitygradient / correlational0.000n/an/a0.8530.150
NMF/PCA dictionariesdim_reduction0.000n/an/a0.9151.000
Path patching / IOI circuitcausal / intervention0.000n/an/an/an/a

Source: site_data.json · generated by gen_site_data.py from the committed §R records + leaderboard.json.