Paper 2 · per-game results

Frostbite

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.frostbite).

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

Frostbite screen

Ground truth

External label source
AtariARI + OCAtari
Verified causal labels
12
Labels that move a sprite
9
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 Frostbite

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/a1.0000.109
Activation patching / causal mediationcausal / intervention1.000n/a1.0001.0000.900
Causal scrubbingcausal / intervention1.000n/an/a1.0001.000
Interchange interventions / DAScausal / intervention1.000n/an/a1.0000.593
A2 · Single-unit lesionscausal / intervention0.803n/an/a0.7211.000
Occlusioncausal / intervention0.7910.7240.8570.5001.000
KernelSHAP / Shapleygradient / correlational0.7830.7630.8040.4000.395
LIMEgradient / correlational0.7020.5800.8240.5000.395
ACDC — automatic circuit discoverycausal / intervention0.667n/an/a0.4211.000
NMF/PCA dictionariesdim_reduction0.659n/an/a-0.8641.000
RISEgradient / correlational0.6190.5340.7040.2500.395
Logit / tuned lenscausal / intervention0.568n/an/an/a0.137
On-distribution counterfactualcausal / intervention0.5110.8260.1960.5501.000
Extremal / meaningful perturbationcausal / intervention0.4930.5600.4250.1251.000
A3 · Tuning curvesgradient / correlational0.476n/an/a0.3461.000
A4 · Pairwise correlationsgradient / correlational0.362n/an/a-0.0390.821
A6 · Granger causalitygradient / correlational0.345n/an/a0.9010.812
Sparse autoencodersdim_reduction0.275n/an/a0.0591.000
Linear probing + control tasksprobing0.275n/an/an/a1.000
Grad×Input / DeepLIFTgradient / correlational0.2720.4840.0590.1501.000
Integrated Gradientsgradient / correlational0.2720.4840.0590.1501.000
Vanilla gradient (saliency)gradient / correlational0.2720.4840.0590.1501.000
SmoothGradgradient / correlational0.2720.4840.0590.1501.000
Guided Backpropgradient / correlational0.2420.4840.0000.1501.000
A7 · Dimensionality reduction (NMF/PCA)dim_reduction0.200n/an/an/an/a
A5 · Local field potentialsgradient / correlational0.138n/an/an/an/a
Attribution / edge patchinggradient / correlational0.031n/an/a0.9431.000
A1 · Connectomics / data-flow graphcausal / intervention0.000n/an/a0.977n/a
Expected Gradientsgradient / correlational0.0000.0000.0000.125n/a
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.