Experiment 12.3 · Model Comparison

Matched image viewer

Choose one diagnostic

Fixed-configuration capability ladder

Open CSV

Win threshold normalized gain 25% Worst-position floor five-choice tasks 20%

Flexible comparison

Exact saved trials

Choose a model and task.

Model Task Condition Layer Coefficient Injected / clean Search displayed fields Hide trials where the expected answer is “No injected thought”

Parsed outcome

Parsed outcome

Source bundles