# SenseShift-RL Method Card

The environment uses `Discrete(37)` to enumerate every subset of zero, one, or
two sites from eight slots. All actions are budget-valid.

At each step, selected valid values are revealed. A frozen inverse-distance
reconstructor predicts the full valid field with weights
`1 / (1 + distance_km)^2`; the train median is used when no valid site was
revealed. Revealed sites have exact zero reconstruction error.

Default daily reward is the negative mean Huber loss over every valid site minus
`0.05` per selected slot. Invalid selected sites still incur the configured
cost. A day with no valid site is unscored and incurs sensing cost only. Hidden
current-day values never enter the agent observation or info mapping.

Test evaluation is parameter-frozen and reward-blind: policies can update
recurrent state from observations and previous actions, but the full-field
evaluation reward is not returned as a future policy input. The v1.1 research
release also records alternative sensing costs. Neural and fixed-spatial
decisions are re-scored without adaptation; Gaussian information reoptimizes
its declared analytic subset objective at each candidate cost.

Natural episode completion is `terminated=True, truncated=False`. External
time limits are responsible for truncation. See PROJECT_SPEC.md in the source
repository for the estimand, tracks, baselines, and falsification tests.