batches/2026-11-05/ : détecteurs, vérité, rapport du scanner, improve/ (mesures, modification, validation, bilan, tools/instants.py de l'upgrader) ; docs/essais/ : le journal de l'essai et l'audit ; outils/ : les scripts de lecture de l'audit, réutilisables (JOUR=AAAA-MM-JJ). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
3.5 KiB
3.5 KiB
wf-improve: BREITHORN,LENZSPITZE,SUSTENHORN, 2026-11-05 → 2026-11-06
Result: modification kept
The range
- No data: found —, correct 0/286 (0 %) → — (no instant flagged)
- Displacement: 33/33, 33/36 → unchanged
- Obstruction: 31/66, 31/37 → unchanged
- Thermal displacement: 513/513, 513/519 → unchanged
- Visibility: 20/35, 20/34 → unchanged
Elsewhere
Total unchanged, before → end: Displacement 5746/6345, 5746/20444; No data 75/836, 75/100; Obstruction 1044/4061, 1044/3718; Target issue 19080/22627, 19080/19467; Thermal displacement 10897/16142, 10897/14828; Visibility 3412/7207, 3412/4442. No other line changed (benchmark-before and benchmark-after identical).
What changed
- S1 (
scripts/no_data.py): the cadence is now the larger of the median gap between instants and one hour divided by the median number of instants per clock hour, so a station measuring in bursts (SUSTENHORN) no longer reports a whole day of No data. - S2 (
scripts/visibility.py): a diffuse-nuisance episode (point 8) lying entirely inside a Poor visibility episode of point 3 is dropped, removing duplicate Poor visibility events.
Upgrader: changed. Validator: keep.
For the user to decide
- BREITHORN-2026-11-05-D1 (Displacement) and D2 (Obstruction), 00:00 → 04:00: the detectors put the dip of POUI22 in both families. A parked vehicle (Obstruction) is likely, given the history of this station, but only the images can say. The benchmark is unchanged either way.
Limits:
- LENZSPITZE-2026-11-05-S1 (Visibility, 08:35 → 09:55): no rejection, error_value at its usual level except at 09:05–09:10 (two instants), curves only slightly more agitated. Telling a loss of visibility from agitation without a rise of error_value takes the images. The scanner judged it from a figure at a coarse scale.
- SUSTENHORN, the periodic rejection of the same half of the targets (11 out of 22) for one to four measurements, day and night: whether it is something passing (Obstruction), the image (Visibility) or the station itself takes the images. The labels mix the two families on the same mechanism (Obstruction for the label S1, Visibility for D9/D14/D18).
Reservations of the validator:
- No data, S1: the cadence step can only rise, so a station whose hours are only partly filled sees a longer step and needs a longer hole before it reports one. This costs nothing on the reference truth (No data 75/836 and 75/100, unchanged). Watch for missed short holes on stations with an irregular cadence.
Not fixed
- SUSTENHORN-2026-11-05-S1 (Obstruction, 05:45 → 08:45): two rejected measurements in a row fall between Brief obstruction and Poor visibility; a tried rule caught one pair only and chained other Brief obstructions through the one-hour merge (Obstruction correct 81 % → 46 %). Dropped.
- BREITHORN-2026-11-05-D3 (extra Displacement, 08:00, one instant): the return of a target from a dip is counted as a jump; fixing it would need a new host rule for one instant.
Tools
batches/2026-11-05/improve/tools/instants.py <detectors> <station folder> <from ISO> <to ISO>: for each instant, the number of sensors, sensors rejected, sensors "high" (rule 1 ofvisibility.py), the median error_value in MAD and the deviation of the corrected curves. Worth keeping as a workflow tool: it shows directly why a visibility or obstruction rule did or did not fire.- Validator: none.
Errors
None. Files already modified: none. Base: 887b313ba272f42058fdcbc422618fe3d132e3ad.
5438b0b wf-improve: S2
7e3043d wf-improve: S1