In reinforcement learning distillation, training on synthetic single-step input-output samples leaves models brittle when real external environments return unexpected errors. Harness distillation trains directly on think, act, and observe feedback loops. When tool execution fails, the model learns intermediate rollback and adaptive correction pathways. In this interactive simulator, clicking Inject Tool Failure forces step three into an error state. The harness trajectory adapts with a purple recovery path, and the tool displays a simulated recovery gain of plus four point eight times.
What changes when you switch modes or inject a tool failure?
Compare the scripted think-act-observe traces, step through the selected scenario, and export its visible trajectory as JSON. Harness mode and synthetic mode show different preset telemetry values. These values illustrate the page’s comparison; the tool does not train a model or measure its generalization.
The density, recovery and generalization readouts are constants selected by the JavaScript mode and failure flag. Changing a scenario, depth or branching setting does not turn them into experimental results. Exported traces are scripted scenario nodes; they are not recorded Claude Code executions or evidence that a distillation method outperforms another.
Start with the default grid scenario and click Synthetic RL Data. Trajectory Density changes from 94.2 to 48.6, and OOD Generalization changes from 78.4% to 32.1%. Then click Inject Tool Failure: the page returns to Harness Trajectories and displays +4.8x Error Recovery Gain. These are preset display values, not measured improvements.
Export Dataset JSON saves the selected scenario and mode, current displayed metrics, and scenario nodes whose step is at or below the current step. Run Trajectory Rollout or move the step control before exporting to compare how much of the scripted trace is included. The export is a demonstration dataset, not an automatically validated training corpus.
ARC-AGI is designed around learning unfamiliar tasks using core knowledge priors. Its official overview explains this benchmark design. The grid scenario here is an illustration; this page does not execute the ARC evaluation protocol, report an ARC score, or substantiate the displayed generalization percentages. ARC Prize: benchmark design and core knowledge priors
Explains the benchmark’s purpose and core knowledge priors; does not validate this simulator’s metrics.