Counterfactuals

Read this page as counterfactual evidence for finite structural reward parameters with a neural continuation-value approximation. The network is not the reward being intervened on.

NNES estimates structural reward parameters and a neural value approximation. The counterfactual evidence comes from the simulation harness, which re-solves oracle counterfactuals and compares them with policies implied by the recovered NNES structural object.

See the Simulation Study for the generator, links to results files, and numerical checks.

Counterfactual Families

Type	Intervention	Purpose
Type A	Shift rewards and hold transitions fixed.	Payoff counterfactual.
Type B	Change transitions and hold rewards fixed.	State-dynamics counterfactual.
Type C	Disable one non-anchor action.	Action-set or design counterfactual.

Primary Results

These rows are the high-dimensional primary-cell results from nnes_results.json.

Counterfactual	Policy TV	Policy KL	Value RMSE	Regret
Type A	0.021578	0.002165	0.007880	0.004865
Type B	0.021651	0.002128	0.007916	0.005559
Type C	0.013053	0.000920	0.002325	0.001314

All three regrets are below the 0.05 threshold. The table compares intervened policies from the recovered NNES object with counterfactual policies from the fully specified data-generating process.

Low-Dimensional Sanity Check

Counterfactual	Policy TV	Policy KL	Value RMSE	Regret
Type A	0.005080	0.000083	0.000271	0.000224
Type B	0.005505	0.000089	0.000335	0.000332
Type C	0.002695	0.000021	0.000076	0.000055

The public NNES wrapper exposes fitted rewards, policies, values, and choice probabilities. It does not currently expose a wrapper-level counterfactual() method, so this page reports the harness-level counterfactual evidence rather than promising a public counterfactual API.