0002. The velocity rules score the stations the calibration is responsible for¶

Status: accepted (owner, 2026-10-10)
Changes: requirements.GATE.velocity_excluded_classes = ("defo_area", "near_event"); the gate.yaml field velocity_excluded_classes ([] restores the all-station rule); metrics.json gains velocity_los_gate (the statistics the rules used, the excluded classes and how many stations were left out). velocity_los (all stations) and velocity_by_class are unchanged and still reported.

Context¶

DISP-CAL removes long-wavelength secular error. Deforming areas and event footprints are removed before the fit and restored after it (remove-restore, Venti PRD): the calibration surface is not fitted there and does not try to explain that motion. The velocity rules of the gate (RMSE and |median bias| may not rise more than 0.2 mm/yr over the baseline) were computed on every station, including those.

T46 Hawaii (F23211, v0.5 + remove-restore vs gamma) shows what that does:

Stations	n	gamma median / RMSE	v0.5 + remove-restore median / RMSE
all	93	+17.6 / 129.5	−2.3 / 129.8
outside deforming areas and events	18	+18.5 / 18.5	+0.5 / 2.4

The all-station RMSE is one station: ERZ4, installed during the 2018 lower East Rift Zone eruption (record 2018-05 to 2020-12; LOS rate −1229 mm/yr; east −2298, north +1511 mm/yr). It is co-eruptive motion that no calibration should match. With it, the gate failed v0.5 on "RMSE up +0.37 mm/yr" while v0.5 removed the +18 mm/yr bias gamma left on every stable station. The 2018 eruption is now an event in the Venti event DB (kilauea_2018_lerz, venti-dev), so ERZ4 is a near_event station.

Decision¶

The velocity rules use the stations outside defo_area and near_event. The DD sill rule and the golden-epoch check are unchanged (the sill is already a double difference on common dates).
The excluded classes are still reported per class in velocity_by_class and in the report, so a candidate that makes them worse is visible.
A frame's benchmark condition gates that frame (owner, 2026-10-10). ClassTarget.gating: true turns a per-class target miss into a FAIL reason; otherwise the frame's main signal would be checked by no rule once it leaves the velocity rules. Thresholds come from evidence (owner, 2026-10-10), not the PRD's fixed "RMSE < 2 mm/yr": the class RMSE must drop from raw to calibrated (rmse_must_drop_vs_raw). Raw is tied to one reference pixel, so its constant offset (the median raw residual on the gated stations) is removed first; the offset and raw's class metrics are recorded as velocity_by_class_raw. Houston (configs/benchmark/gate_F08882.yaml): |median bias| < 1 mm/yr and the RMSE drop, on defo_area (with the Venti area DB) and subsiding (without it).
A run must pass the same deforming-area and event databases to every label it compares (candidate and baseline): the classes come from the candidate run's databases and are applied to both.

Consequences¶

Nothing changes for a run without defo_geojson / event_geojson: no station is in those classes. The T38 and T46 runs so far passed none, so their verdicts are unchanged until they are re-validated with the Venti databases.
Houston before remove-restore (T46 v0.5 run, no area DB, 82 subsiding stations):

	median	RMSE
raw, as referenced	−5.62	6.37
raw, constant offset removed	+0.09	3.00
gamma	+4.43	5.24
v0.5	+0.92	4.87

Against the fair raw, both calibrations make the subsiding stations worse: without remove-restore the 50 km surface takes out part of the subsidence bowl, as in Central Valley. Houston v0.5 fails this rule until it runs with remove-restore (T46 F08882 v05rr, running 2026-10-10). - T46 with remove-restore, on the gated stations (outside areas and events; Hawaii with the kilauea_2018_lerz event):

frame	gamma median / RMSE	v0.5 + remove-restore median / RMSE	all-station verdict
F36542 Central Valley	+10.6 / 12.8	+2.4 / 5.0	PASS
F23211 Hawaii	+18.5 / 18.5	+0.5 / 2.4	FAIL (ERZ4)
F16941 Ridgecrest	+0.27 / 2.84	+0.45 / 2.16	FAIL (median +0.32)

Ridgecrest's |median| rise on the gated stations is +0.18 (< 0.2). - Hawaii passes the velocity rules with remove-restore (table above); its inside-area RMSE (gamma 43, v0.5 48 mm/yr) is reported, not gated.

Amendment (owner, 2026-10-13): the RMSE limit includes the GNSS reference error¶

Problem. Inside Houston's subsidence polygon, raw DISP with one offset removed fits MIDAS at 1.74 mm/yr. The offset is fitted on the validation stations themselves, so raw gets help the product cannot have. InSAR's relative precision there is better than the absolute GNSS reference the calibration ties to: the UNR grid vs MIDAS is 1.82 mm/yr in that class (TS-G1, F08882). The calibrated RMSE, 2.42, matches √(1.74² + 1.82²) = 2.52. "Below raw" cannot be met there by any surface setting.

Decision. A class target may carry reference_rmse_mm_yr, the grid vs station RMSE of that class taken from the TS-G1 study. The rule is then RMSE ≤ √(raw² + reference²): the calibration adds no error beyond its reference. The validator still never reads the grid; the number is a documented study result in the frame's gate file.

Houston (configs/benchmark/gate_F08882.yaml): defo_area 1.82 and subsiding 4.23 mm/yr.

run	defo_area	subsiding
v0.5 + RR	2.42 ≤ 2.52 ✓	7.30 > 6.13 ✗
v0.5 + RR + regions (venti-dev #32)	2.42 ≤ 2.52 ✓	6.28 > 6.13 ✗ (by 0.15)

The subsiding class outside the polygon still misses. The cause is Galveston's TXGA (a sub-region cycle error) and Matagorda, both items for the edge-case pass.