Files
aimsight-labeling/auto/analyses/WILDSTRUBEL_2026-10-06T1945_2026-10-20T0000/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

9.7 KiB
Raw Blame History

Labels: WILDSTRUBEL, 2026-10-06T19:45Z → 2026-10-20T00:00Z (UTC)

Requalification (2026-11-17) of the 23 verified labels of this range under the new rule (labeling/labeling-rules.md, section 1): only what stands out from the noise of the corrected curves, at least one target beyond 5 of its own noises. The reading was not redone; the earlier verification table is in work/labels.before-requalify.jsonl and in git history. Result: 12 kept as they are (with their id and review), 1 rewritten (194), 10 removed. 13 labels remain.

Noise used: the target's own noise (median of the absolute differences between consecutive accepted measurements) over the measured batch 07/10 05:00 → 09/10 12:30, about 0.03 to 0.05 px in corr_x and 0.02 to 0.055 px in corr_y depending on the target. The station has no night measurements in this range (it shuts down every night), so that noise is a daytime noise. That is the station's ordinary regime here.

1. My labels

2026-10-06T19:45Z → 2026-10-07T07:00Z (WILDSTRUBEL-176), 2026-10-07T13:30Z → 2026-10-08T05:30Z (182), 2026-10-08T16:00Z → 2026-10-09T05:00Z (192), 2026-10-09T12:30Z → 2026-10-15T08:00Z (197)

  • Phenomenon of the catalogue: station out of energy (176, 182, 192); station silent (197)
  • Precedents: not rerun (requalification)
  • Family: No data. No sensor delivers anything there, apart from the isolated measurements already stated in 176 and 182 (05:00, 06:00, 06:30; 14:15).
  • Measured: inventory, steps absent for all sensors (work/inventory.txt); unchanged since the verification.
  • Images: none (requalification)
  • Confidence: high. Kept as they are.

2026-10-07T11:45Z → 2026-10-07T12:30Z (180)

  • Phenomenon of the catalogue: passage in front of the references
  • Family: Obstruction, unchanged
  • Measured: at 12:00, PA3_2 corr_x +0.25 px (6.7 noises, about 2.9 mm), PB2_2 corr_x +0.34 px (6.4), PB_5 corr_x +0.26 px (5.9), PB1_2 corr_y +0.17 px (5.9).
  • Confidence: high. Kept.

2026-10-07T13:30Z → 2026-10-08T13:30Z (183)

  • Phenomenon of the catalogue: movement of single targets (cause not established)
  • Family: Displacement, unchanged
  • Measured: ranges.py --outside 2026-10-07T13:30Z/2026-10-09T12:30Z (level and noise of 07/10 by day): PB2_2 corr_x beyond 5 noises from 11:00 to 15:00 on 08/10 (7 measurements, extreme -0.36 px, -7.9 noises, about -2.4 mm). It stays there after PA1_2 comes back (13:15), with five references kept. PB1_2: -0.43 px (-9.3) from 08:30 to 11:00, then -6.9 at 12:00–13:00. On 09/10 they are still between 5 and 6 noises below. Measured against the batch noise, the lasting step after the return of PA1_2 is about 4.9 noises (PB2_2) and 3.6 (PB1_2): at the edge of the threshold.
  • Confidence: medium. The step is just at the threshold once PA1_2 is back. Kept because the prescribed tool (k 5) places PB2_2 beyond the threshold for 7 measurements, including after the reference returns. It is also a possible real movement of the kerb.

2026-10-08T06:00Z → 2026-10-08T13:15Z (186)

  • Family: Visibility (reference PA1_2 lost in the sun), unchanged
  • Measured: PB1_2 corr_x -0.31 px at 08:30 (6.1 noises against the 08/10 afternoon level). At 13:00 (PA1_2 rejected, 4 references), PB_5 corr_y -0.21 px (9.3 noises), PA2_2 -0.26 px (8.1), PA_9 -0.41 px (7.6, about 6 mm), PB2_2 (5.9). These moves are what happens to the other curves when the reference is lost, so the label holds under the missing-points rule.
  • Confidence: high. Kept.

2026-10-08T11:30Z → 12:00Z (187), 14:00Z → 14:30Z (190), 14:30Z → 15:00Z (191)

  • Measured: 187: PB4 corr_x +0.56 px (16.8 noises, about 5.2 mm), PB_5 +0.36 px (8.0). 190: PB_5 corr_x +0.45 px (10.1), PB1_2 corr_y (6.1). 191: PB4 corr_x +0.50 px (15.0), PB_5 (8.0), PA3_2 (7.1).
  • Confidence: high. Kept.

2026-10-09T05:00Z → 2026-10-09T12:30Z (194), rewritten

  • Phenomenon of the catalogue: loss of the references at sunrise (low sun)
  • Family: Visibility, unchanged
  • Measured: at sunrise, no reference is retained at 05:00 to 05:30 (inventory, retained references). PA2_2 corr_x +0.34 px (8.4 noises, about 3.3 mm) and corr_y +0.24 px (7.5), PB1_2 corr_x +0.33 px (6.6) and corr_y (6.7), PB2_2 corr_x (5.4). Everything is back from 06:15. At 09:45, without PA_8, PA2_2 corr_x is at 5.1 noises (ranges.py, batch window), 4.9 with my script. From 06:15, leaving out the passages, the largest deviation is 4.9 noises: the daytime agitation stays within the noise.
  • Why rewritten: the old text rested on the daytime agitation (heat shimmer), which no longer crosses the threshold. What crosses it is the sunrise loss of the references. The bounds are kept (the label holds over its range). The description now speaks about that effect, so the review field is dropped.
  • Confidence: medium (the bounds stay wider than the event, as asked).

2026-10-09T11:45Z → 12:15Z (196), 2026-10-15T07:45Z → 08:15Z (202)

  • Measured: 196: PA3_2 and PB4 about 29 px off on both axes (hundreds of noises; about 0.35 m and 0.28 m), a false lock on reflections. 202: PB1_2 corr_y +1.03 px (35 noises, about 5.6 mm), PB_5 +0.69 px (30.5, about 6.3 mm), PA_9 +0.96 px (17.9, about 14 mm), with a single reference kept.
  • Confidence: high. Kept.

2. What I left without a label

Removed labels (the old ids), with the largest deviation over their range, in batch noises, leaving out the instants of the labels that stay:

id bounds family deviation why
WILDSTRUBEL-177 07/10 05:00 → 12:00 Thermal displacement 2.8 noises (PB4 corr_y, against the 06/10 evening level) camera movement absorbed by the compensation
WILDSTRUBEL-178 07/10 05:00 → 14:15 Visibility 3.5 (PB1_2, PB2_2 corr_x) daytime agitation within the noise; the 12:00 spike is 180
WILDSTRUBEL-179 07/10 11:15 → 11:45 Obstruction 3.1 (PA_9 corr_x at 11:30) missing points only (PA3_2, PB4, PB_5, PA1_2, R2 rejected)
WILDSTRUBEL-181 07/10 12:30 → 13:00 Obstruction 4.1 (PB1_2 corr_x at 12:45) missing points only (PB4, R2)
WILDSTRUBEL-184 08/10 05:30 → 13:00 Thermal displacement 4.2 (PB4 corr_y), against the 08/10 afternoon level, outside 08:30 compensated; the 08:30 dip of PB1_2 (6.1) happens with PA1_2 lost and is covered by 186. The rotation that does not come back stays in the pose, not in the corrected curves
WILDSTRUBEL-185 08/10 05:30 → 15:00 Visibility 4.2 (PB2_2, PB1_2 corr_x) daytime agitation within the noise; the spikes are 186, 187, 190, 191
WILDSTRUBEL-188 08/10 12:00 → 12:30 Obstruction 4.2 (PB2_2 corr_x at 12:00) missing points only (PA3_2, PA_9, PB4, R2)
WILDSTRUBEL-189 08/10 12:30 → 13:00 Obstruction 2.2 (PA_9 corr_y at 12:30) missing points only (PB1_2, PB2_2, R2); the 13:00 step is 186
WILDSTRUBEL-193 09/10 05:00 → 12:30 Thermal displacement 4.7 from 06:15, outside the reference drop-outs (5.8 for PA2_2 at 09:45, but that is the PA_8 drop-out, now in 194) compensated
WILDSTRUBEL-195 09/10 09:15 → 09:45 Obstruction 3.8 (PB1_2 corr_y at 09:30) missing points only (PB_5, PB3_2)

Remark on the thermal labels: with the noise of the calm 06/10 evening alone (ranges.py --outside, 19 measurements, a noise two to four times smaller), PB1_2 and PB2_2 cross 5 noises on 07/10 by day. That is only about 0.1 px (under 1 mm). The evening noise is not the station's ordinary noise in this range: every measured hour is by day, and the nights are gaps. I keep the batch noise, the same for every label.

Also left out, as in the verification: the isolated rejections of PA_9 (07/10 10:45–11:00) and of PA_9, PA3_2, PB2_2 (08/10 08:30–09:00) are missing points. The spanning labels WILDSTRUBEL-130 (R3, Target issue) and 199 (No data) are outside this analysis and unchanged.

3. The documentation

The new rule (section 1 of labeling-rules.md) is clear about missing points and compensated thermal effects. It says nothing about which noise to use when a station has no night measurements at all. The calm evening and the batch give thresholds that differ by a factor of two to four. I used the batch noise; a sentence in the rules would settle it. A bound rule conflicts with requalification: 194 keeps day bounds that are wider than its event (sunrise), because the bounds of a label that holds must not move.

Requalification

Commands (output in work/requal/):

  • python3 tools/scripts/ranges.py data/WILDSTRUBEL --columns corr_x,corr_y --k 5 --start 2026-10-06T19:45Z --end 2026-10-20T00:00Z --sensors PA2_2,PA3_2,PA_9,PB1_2,PB2_2,PB4,PB_5 → ranges_k5.txt. Over the batch, the only ranges beyond 5 noises are at 07/10 12:00, 08/10 08:30, 11:45, 13:00, 14:15, 14:45, 09/10 05:00–06:00, 09:45, 12:00 and 15/10 08:00.
  • The same with --outside around each day label (r177.txt, r183.txt, r184.txt, r193.txt) and with --rolling 4 --instants 2 (rroll.txt).
  • python3 scripts/requal.py '<label windows>' (in this folder) → requal.txt: for each label, the largest deviation per target in batch noises, leaving out the instants of the other labels.

Values: kept 180 (6.7), 183 (7.9 with the 07/10 noise; 4.9 in batch noise after PA1_2 is back), 186 (9.3), 187 (16.8), 190 (10.1), 191 (15.0), 196 (>700), 202 (35.3); No data 176, 182, 192, 197 (no measurement). Rewritten 194 (8.4 at sunrise). Removed: see section 2.

Difficulties

none

Sources read

instructions/analyst.md; labeling/labeling-rules.md (sections 1 to 5); tools/GUIDE.md (ranges, inventory, days, metrics); tools/DATA.md (start); the earlier analysis.md, report.md, work/requalify_input.jsonl, work/inventory.txt; inventory.py and ranges.py outputs on the range; tools/scripts/common.py (own_noise) and data.py (loader). No weather folder for WILDSTRUBEL.