Files
aimsight-labeling/auto/analyses/long-labels/NIEDERHORN/analysis.md
T
maximeandClaude Opus 5.5 ddb17c2189 auto: labels requalified under the 5-noise rules (only what stands out from the noise): 32 analyses refined, the 6 paused periods verified, the long labels requalified with the fixes the period analysts found; 1153 -> 961 labels
- 32 analyses: each label kept, rewritten or removed by its analyst (section Requalification of the report; mm magnitudes where the targets have them)
- SNCF01 x3, WILDSTRUBEL 2026-09-01 -> 2026-10-06 x3: verified from their existing labels under the new rules
- long labels (no analysis): auto/analyses/long-labels/<STATION>/; NORDEND-169 and -195 run to 07/11, so NORDEND-283 and -284 (their continuation) are removed
- reading-rules.md, stations.md: the citations of removed or rewritten labels point to their successor, or to [expert]

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-17 23:58:00 +01:00

5.7 KiB
Raw Blame History

Requalification of the long labels: NIEDERHORN, 2026-10-01 → 2026-11-01 (UTC)

Evidence (in work/):

  • inventory.txt: inventory.py over the window, with the rejected measurements of each sensor.
  • ranges_k5.txt and ranges_k5_roll1d.txt: ranges.py --k 5 on corr_x, corr_y and their mm, without and with --rolling 1d, for the 12 labelled targets.
  • ranges_k5_roll1d_neighbours.txt: the same for 8 healthy targets.
  • specific_month.txt: scripts/specific.py, the deviation of each target once the median of 16 healthy targets is removed, compared with the same figure for the healthy targets (leave-one-out).
  • scripts/daynight.py: the hour-of-day profile against the night level, printed in the terminal, not saved.

The month has 721 measurement instants. All 12 labels are Target issue labels of missing points. Under the new rules such a label is kept only in two cases: the target is rejected almost always, or the episode moves the other curves beyond the noise. For the partial ones I also checked the target's own kept points: do they leave 5 noises more than those of the healthy targets?

  • NIEDERHORN-2 (e03): kept as is. Rejected 721 / 721 (100 %); no corrected curve.
  • NIEDERHORN-3 (e04): kept as is. Rejected 721 / 721 (100 %); no corrected curve.
  • NIEDERHORN-4 (e05): kept as is. Rejected 721 / 721 (100 %); no corrected curve.
  • NIEDERHORN-5 (e07): kept as is. Rejected 718 / 721 (almost 100 %); 3 isolated accepted points, within about 1 noise of each other.
  • NIEDERHORN-6 (e08): kept as is. Rejected 720 / 721 (almost 100 %); a single accepted point.
  • NIEDERHORN-7 (e11): kept as is. Rejected 721 / 721 (100 %); no corrected curve.
  • NIEDERHORN-8 (e12): kept as is. Rejected 721 / 721 (100 %); no corrected curve.
  • NIEDERHORN-9 (PWD_RMP_B15_017_C09): removed. Rejected 302 / 721 (42 %), at night only, and not a reference, so its rejections move no other curve.
    • Kept points with the healthy median removed: beyond 5 noises at 8 % of its measurements on corr_x and 27 % on corr_y, against 4–26 % and 5–36 % for the healthy targets. Its extreme corr_y deviation is −18 noises (−5.8 mm, 10-01 12:00); the healthy targets reach −29 to −60 noises (−5.7 to −11 mm) at the same kind of shared instant.
    • The "scattered daytime points that change level" after 28.10 are the afternoon rise of the daily thermal cycle: corr_y +13 to +16 of its own noises from 15:00 to 18:00 UTC (about +2.5 mm). Its neighbours of the same row do the same at the same hours (PWD_RMP_B12_017_C09 up to +11 noises, PWD_RMP_B15_018_C09 up to +6). These afternoons are covered by the Thermal displacement labels (NIEDERHORN-125, -129, -136, -139).
    • Result: missing points, nothing of its own beyond the noise.
  • NIEDERHORN-10 (e01): removed. Rejected 260 / 721 (36 %), by day only, and not a reference.
    • Its afternoon rise on corr_x (hourly median up to about +15 noises at 18:00) has the same shape as its neighbours' e02 (up to +9) and e06 (up to +6.5), inside the daily thermal cycles.
    • Its largest point (+25 noises, 10-06 17:00) also shows on e02 (+22) and e06 (+18) at the same instant.
    • Result: no deviation of its own; missing points.
  • NIEDERHORN-11 (PWD_IMP_B08_024_C09): removed. Rejected 154 / 721 (21 %): by day, mostly in the first half of the month.
    • With the healthy median removed it is beyond 5 noises at 17 % (corr_x) and 10 % (corr_y) of its measurements; the healthy targets show 4–26 % and 5–36 %.
    • Its extremes (−18 noises, −3.7 mm on corr_x on 10-24 19:00; −42 noises, −9.5 mm on corr_y on 10-29 10:00) fall at shared instants: −40 to −45 noises on PWD_RMP_B15_020_C09 and PWD_RMP_B15_019_C09 on 10-29 07:00–10:00, inside the Visibility label NIEDERHORN-127.
    • Result: missing points, near the threshold.
  • NIEDERHORN-70 (PWD_RMP_B12_019_C09): removed. Rejected 142 / 390 over its range from 10-15 01:00 (36 %), at night, and not a reference.
    • Its largest kept point (corr_y +35 noises, +7.3 mm, 10-20 12:00) is the shared single measurement of the Obstruction label NIEDERHORN-93; the healthy PWD_RMP_B12_020_C09 jumps the same way (+36 noises) on such instants.
    • Elsewhere it is beyond 5 noises at 10 % (corr_x) and 8 % (corr_y) of its measurements, inside the healthy range.
  • NIEDERHORN-79 (PWD_IMP_B07_018_C09): removed. Rejected 114 / 352 over its range from 10-17 03:00 (32 %), at night, and not a reference.
    • Its kept points (corr_x up to −18 noises, −3.7 mm) behave like its neighbour PWD_IMP_B07_017_C09 (−19 to −21 noises) and like the other healthy targets.
    • Result: no deviation of its own.

Notes

  • The catalogue still lists "a target rejected every night or every day" as a Target issue (families.md). The requalification instruction overrides it: a label of missing points stays only if the target is rejected almost always. The five removed targets are rejected 21 to 42 % of the time. Rejected targets that are not references do not affect the compensation.
  • My common-mode removal (the median of healthy targets spread across the image) does not cancel the shared movements fully. That is why the healthy targets still leave 5 noises often. So it was used only as a comparison between targets, never as an absolute threshold.
  • Sources read: labeling/labeling-rules.md in full; the Target issue and gap parts of labeling/families.md; the ranges, overview and conventions sections of tools/GUIDE.md; tools/scripts/data.py (load, rejected); labeling/auto/NIEDERHORN.jsonl (the October labels); the other long-label reports (BREITHORN, CLARIDEN, FLORIDA, JUNGFRAU, MICHIGAN) and FLORIDA's analysis.md. Not read: the NIEDERHORN period reports; no thumbnails were opened (the rates and curves were enough).