Skip to content

0001. Median bias in the gate; "coastal" means the sea

  • Status: accepted (owner, 2026-10-09)
  • Changes: requirements.GATE.bias_statistic = "median"; the gate.yaml field bias_statistic; classes.classify_stations(sea_min_km2=500).

Context

Bias. The gate fails a candidate when its |bias| against MIDAS rises by more than 0.2 mm/yr over the baseline (PRD §2.7). The release target is |bias| ≤ 1 mm/yr. Both used the mean station residual. On F08882 the TS-U1 unwrap run corrected five Galveston stations by 3–6 mm/yr toward MIDAS. They had been large negative outliers on a frame whose bias is positive (issue #2), so:

  • the mean rose from +1.14 to +1.37 mm/yr, and the gate failed (+0.22);
  • the RMSE fell from 4.79 to 4.31, and the median stayed at +0.95.

A rule that fails a candidate for fixing outliers judges the outliers, not the calibration.

Coastal. The station classes (PRD §3.4) report metrics separately for coastal stations, because the coast is where calibration fails differently: water gaps, islands, unwrapping across water, and the land/sea troposphere. The rule was "within 10 km of any water in the DISP water mask". Those masks hold thousands of ponds, reservoirs and rivers; Oklahoma City has ~20 000 water bodies. As a result:

  • every F08886 station was "coastal";
  • no stable class was left on F08886 or F08882, which is the class TS-B1 needed.

Decision

  1. The bias rules use the median of the station residuals. That covers the rise vs baseline (0.2 mm/yr, unchanged), the release target (1 mm/yr, unchanged, reported) and the per-class targets. The mean is still reported next to it. The RMSE rule is unchanged, so a candidate that makes outliers worse still fails. gate.yaml can set bias_statistic: mean for a release that needs the old rule.
  2. "Coastal" is within 10 km of the sea. The sea is a water body of at least 500 km² in the water mask, together with the bays and tidal channels joined to it.
  3. On the four frames the seas are 900–15 000 km²: the Gulf of Mexico, Galveston Bay, the Pacific, and the Atlantic with Long Island Sound.
  4. The largest inland water body is ~300 km².
  5. sea_min_km2=0 restores the old rule.

Consequences on the T38 runs (v0.5 vs gamma)

Gamma → v0.5, mean / median LOS bias in mm/yr:

frame mean bias median bias
F08882 +4.05 → +1.14 +4.35 → +0.95
F08886 +0.90 → +0.06 +0.83 → −0.11
F16940 +2.95 → +0.65 +2.17 → +0.61
F08622 +0.71 → +0.25 +0.73 → +0.11
  • All four frames still PASS (re-run 2026-10-09).
  • The 1 mm/yr release target is now met on all four frames; with the mean, Houston missed it at +1.14.
  • The TS-U1 unwrap run passes the gate under the median. Its no-go stands on the other grounds in studies/TS-U1/REPORT.md, and the correction stays off.
  • Stable stations reappear: Houston 24, Oklahoma City 24, Los Angeles 209, New York 39. Class tables in the reports change.