release/DATA_DICTIONARY.md

# SenseShift-RL Data Dictionary

## worlds.parquet

| Column | Meaning |
|---|---|
| world_id | Stable authored geographic-world identifier. |
| world_role | Train, calibration, or held-out test geography. |
| world_role_sha256 | Hash used to assign the frozen world role. |
| latitude_bin, longitude_bin | Authored fixed-cell indices. |
| cell_center_latitude, cell_center_longitude | Fixed geometric cell center. |
| site_count | Number of sites; exactly eight in environment v1. |

## sites.parquet

| Column | Meaning |
|---|---|
| world_id, slot | World and action-slot key. |
| site_id, monitor_id | AQShift-US/EPA source identifiers. |
| latitude, longitude, elevation_m | Static EPA site metadata. |
| train_valid_coverage | Training-period valid-day fraction used for selection. |
| distance_to_cell_center_km | Authored haversine distance used in panel selection. |

## measurements.parquet

| Column | Meaning |
|---|---|
| track | Allowed protocol partition. |
| world_id, slot, site_id, monitor_id, local_date | Natural row key and provenance. |
| measurement_value_ppm | Valid EPA daily 8-hour ozone maximum; null otherwise. |
| value_normalized | `(value - train median) / train IQR`; null otherwise. |
| valid_measurement | Whether this row may enter reward truth. |
| availability_reason | Validity reason or `no_source_row`. |
| source_last_change_date | EPA record revision date where a source row exists. |

## episodes.parquet

| Column | Meaning |
|---|---|
| episode_id | Unique fixed episode identifier. |
| track, world_id | Protocol partition and world. |
| episode_start_date, episode_end_date | Inclusive 90-day boundary. |
| calendar_year, quarter_index | Calendar grouping. |
| episode_days | Exactly 90 in environment v1. |

## normalization.json

Global training-only median/IQR statistics and the method-lock hash.

## baseline_episode_metrics.parquet

One row per policy, test episode, and policy seed. It contains total and
per-calendar-day return, field Huber loss, query rate, selected-missing rate,
scorable coverage, and episode size.

## baseline_metrics.json and tabular_q_values.npy

Macro-by-world summaries with uncertainty metadata, plus the frozen transparent
Q-table used for test evaluation.