# Frozen thermal comparison data

See `../RIGHTS.md` for source attribution, changes and **CC BY-SA 4.0** terms. Keep that notice with CSV/JSON exports. No customer motor or changed-design performance is represented.

## Files and units

- `summary.json`: schema version, three main model IDs, seven aggregate comparisons, original-resolution per-session rotor metrics, split IDs, frozen criteria, exposure/initialization/band definitions, source URLs and source hashes.
- `profile-metrics.csv`: 98 rows, seven candidates × 14 sessions. Temperature errors and band widths are °C; coverage is a fraction from 0 to 1; sample count excludes the initial state.
- `model-comparison.csv`: seven rows. Equal-profile RMSE is the square root of the arithmetic mean of the 14 session MSEs. Sample-weighted RMSE is separately labeled. Positive bias means the prediction is warmer than the observation. Maximum underprediction is max(0, measured − predicted) over scored samples. Both acceptance outcomes remain separate.
- `traces/profile-<id>.json` and `.csv`: selected original samples with original sample index, time in seconds from session start, measured rotor temperature in °C, and prediction/lower/upper values in °C for physical, causal and hybrid models. The figure displays time in minutes. Raw measurements were sampled at 2 Hz.

The JSON `columns` array defines the order of each array in `rows`. JSON and CSV display values are identical. For each of 120 time bins, the export retains the union of min/max indices for observations, the three predictions, their lower/upper bounds and their measured-minus-predicted errors, plus both endpoints. Values are rounded to six decimal places. Global extrema survive to that display precision. Connecting retained points omits fine time structure; this is a display export, not a replacement for the source records.

All scores and coverage use original 0.5-second samples and exclude sample 0, the measured initial state. There are 324,540 held-out rows including initialization, and 324,526 scored rows per model. Display export contains 12,848 rows across 14 sessions. The 14 held-out profiles are 4, 6, 7, 9, 11, 15, 26, 27, 29, 36, 53, 63, 75 and 79.

Bands are development-derived constant-width envelopes, not guaranteed 95% coverage. Original frozen fields include “95” in their names; the figure uses neutral band labels to avoid implying calibration. Metadata includes no calibrated sensor uncertainty for these scores. These are released evaluation outcomes, not fresh unseen data for future revisions.
