Files
aimsight-labeling/auto/analyses/WILDSTRUBEL_2026-09-24T0230_2026-10-06T1945/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

22 KiB
Raw Blame History

Labels: WILDSTRUBEL, 2026-09-24T02:30Z → 2026-10-06T19:45Z (UTC)

Task: verify the 94 existing labels of the range (work/existing.jsonl) against the rules of 2026-11-17 (labeling-rules.md, section 1: a movement is labelled when a target leaves its level by more than 5 own noises; a thermal cycle only when the corrected curves keep it; a missing point not for itself; No data only when the station delivers nothing). Result: 46 kept, 6 corrected, 42 removed, 12 added → 64 labels in labels.jsonl.

Context. Station on a road in a town (stations.md), solar battery, six references (PA1_2, PA_8, PB3_2, R1, R2, R3), seven targets at 13.8 m (PB1_2, 5.46 mm/px) to 39.1 m (PA_9, 15.06 mm/px); no image quality metrics (export columns empty); no weather (no weather/ folder: the station has no Swiss position). Cadence 15 min; measurement timestamps carry seconds (e.g. 09:30:12), floored to the minute in my scripts to compare with the label bounds. Own noises of corr (median |diff| over the range): 0.015 to 0.030 px, i.e. 5 noises ≈ 0.07 to 0.15 px ≈ 0.5 to 1.5 mm depending on the target.

Method (scripts in scripts/, outputs in work/):

  • exceed.py: every target measurement more than 5 own noises from its rolling level (median of 8 accepted neighbours on each side, as ranges.py --rolling 8), with the references out and the labels covering it; then per label the max deviation strictly inside (work/exceed.txt on the old labels, work/exceed_new.txt on the new set). labeldev.py (previous session, work/labeldev.txt) adds the deviation against the measurements at the edges of each label.
  • thermal.py: per day and target, the raw deviation of the corrected curves from the night level (else the evening, else the afternoon rest) at clean instants (all references but R3 retained, outside brief labels), and the longest run of consecutive points beyond 5 noises (work/thermal.txt). days_dev.py (previous session, work/days_dev.txt): the same with a 5-point median and the fast daytime agitation against the night agitation.
  • days.py (night levels, byhour), overview.py --rigid on 29/09, summary.py tables, thumbnails, metrics.py.

1. My labels

The 46 labels kept as they are (id, verdict, evidence)

Max deviation strictly inside the label, in own noises (rolling level / edge level).

  • 81 (24/09 02:30 → 01/10 00:00, Obstruction, PB2_2): kept. PB2_2 corr_x night level −13.2 noises on 24/09 (≈ −1.8 mm), +9.9 on 25/09, slow slide to 29/09, rise on 30/09; the rise goes on a little until 01/10 02:00 (one more noise), within the uncertainty of the end.
  • 83: corrected (below).
  • 85 (24/09 05:30 → 06:00, Visibility): kept; PB2_2 latched, huge one-point jump.
  • 86 (24/09 09:00 → 09:30, Obstruction): kept; PB_5 latched, huge jump.
  • 89 (25/09 02:30 → 29/09 23:00, Displacement, PB1_2 corr_y): kept. Night levels of PB1_2 corr_y −0.165 (24/09) → −0.195 (25/09) → −0.250 (29/09) → −0.246 (30/09) px: −4.9 noises on the night medians, the night measurements of 29/09 beyond 5 noises from those of 24/09 (≈ −0.46 mm); days.py night levels.
  • 94 (25/09 12:00 → 12:30): kept; R1, R2 rejected, PB4 +16.6.
  • 96 (26/09 03:45 → 07:45): kept; parked cars, compensation without references, PB1_2 +27.6.
  • 103 (27/09 06:45 → 07:15): kept; PB_5 +7.0.
  • 106 (28/09 04:15 → 12:30, Thermal): kept; PB2_2 corr_y beyond 5 noises from the night level on 6 consecutive clean points 10:00 → 11:45 (max +6.9).
  • 107 (28/09 05:15 → 05:45): kept; PB_5 +9.9.
  • 111 (29/09 02:30 → 05:15, No data): kept; gap of 10 measurements, metrics as described.
  • 113: corrected (below).
  • 116 (29/09 10:45 → 11:15): kept; R1, R2, R3 lost, PB4 +9.5.
  • 118 (29/09 21:00 → 21:30): kept; PA_9 latched on headlights.
  • 119 (30/09 02:15 → 05:45, No data): kept; isolated single measurements at 04:30 and 05:15 (a single measurement is not a resumption).
  • 120 (30/09 02:15 → 16:15, Displacement): kept; PB_5 corr_x +0.08 px from the night level (+5.3 noises, ≈ +0.7 mm), PA3_2 −0.07 px (−3.4 noises, ≈ −0.8 mm).
  • 121 (30/09 04:30 → 12:15, Thermal): kept; PB1_2 corr_x run of 2 beyond 5 noises (05:45-06:00, max −5.8), PB2_2 corr_y run of 2 (07:00-07:15, max +6.2).
  • 123 (30/09 07:30 → 08:15): kept; PB4 +16.6.
  • 125 (30/09 09:15 → 09:45): kept; PB4 +11.4.
  • 127 (30/09 10:45 → 11:30): kept; PB4 +14.9.
  • 131 (01/10 02:30 → 03:15): kept; PB1_2 −17.8, PB2_2 lasting drop.
  • 132 (01/10 03:15 → 05:45, No data): kept.
  • 133 (01/10 04:45 → 12:30, Thermal): kept; PB1_2 corr_x run of 5 (04:45-07:15, max −13.1 ≈ −1.4 mm), corr_y run of 8 (06:30-08:15, max +8.8).
  • 135 (01/10 08:15 → 13:15, Visibility, PA_8): kept; PA2_2 +37.6, PA_9 +32.1 at the loss of all references, steps of PA_9.
  • 136 (01/10 13:00 → 13:30, Displacement, common shift): kept. Night corr_y 30/09 → 01/10: PA2_2 −4.6, PB_5 −4.4, PB4 −3.8, PB2_2 −6.0 noises on the night medians (≈ −0.6 to −0.7 mm, the same px for targets from 14 to 39 m: a base change, not the road); many single points beyond 5.
  • 137 (01/10 21:30 → 02/10 05:15, No data): kept.
  • 138 (02/10 04:45 → 13:00, Thermal): kept; PB1_2 corr_x run of 2 (06:00-06:30, max −7.1) among the few clean instants.
  • 141 (02/10 12:45 → 20:15, Obstruction, PA1_2): kept; jump at 13:00 (PA2_2 −6.2, four targets beyond 5).
  • 142 (02/10 16:15 → 16:45): kept; PB4 latched.
  • 143 (02/10 20:15 → 03/10 05:00, No data): kept.
  • 144 (03/10 05:00 → 07:45): kept; PB1_2 +43.5.
  • 147 (03/10 07:30 → 08:00): kept; PB4 +15.6.
  • 148 (03/10 08:00 → 13:45, Visibility, PA_8): kept; it also holds the PA jump at 08:45 when PA_8 comes back for one measurement (PA2_2 −8.0), formerly under 149.
  • 150 (03/10 10:00 → 10:30): kept; PA3_2 latched.
  • 153 (03/10 15:15 → 04/10 15:45, Displacement, return of the common shift): kept; corr_y of PA2_2, PB4, PB_5 back toward their September level (figure).
  • 154 (03/10 15:30 → 04/10 06:00, No data): kept. The two acquisitions of 04:42:56 and 04:45:24, out of cadence and almost all rejected, then 05:30 alone: I read them as one burst, not a resumption.
  • 157 (04/10 06:45 → 07:15): kept; PA2_2 +9.3.
  • 160 (04/10 13:30 → 14:00): kept; PB4 latched.
  • 161 (04/10 14:00 → 14:30): kept; all targets −13 to −26 together.
  • 162 (04/10 15:45 → 05/10 06:00, No data): kept; single measurements at 04:30 and 05:00.
  • 165 (05/10 08:00 → 08:30): kept; PB4 +18.6.
  • 166 (05/10 08:30 → 09:00): kept; PB4 +5.3/+6.2, PB1_2 +5.5/+5.8 (two targets).
  • 169 (06/10 05:45 → 12:30, Thermal): kept; PB1_2 corr_x 0.11 px below its afternoon level of 05/10 from 07:15 to 08:00 (−5.7 to −6.4 noises on three of four points), before its step of 08:30.
  • 171, 172, 174, 175 (06/10): kept; PB1_2 +11.8, PB4 +18.7, PB_5 +14.3, PB_5 +13.8.
  • 173 (06/10 08:30 → 09:00, Displacement, PB1_2): kept. PB1_2 corr_x ≈ +0.10 on the afternoon of 05/10 and up to 08:30 (apart from the morning dip), ≈ −0.08 to −0.13 from 08:45 until the evening (≈ −9 noises, ≈ −1.0 mm), with R3 rejected both before and after: the step is not R3 entering.

24/09 04:00 → 12:30 (WILDSTRUBEL-83, corrected)

  • Phenomenon of the catalogue: The camera moves through a thermal effect.
  • Precedents: precedents.py unreachable (Difficulties); cases used: the station's own validated thermal labels 106, 121, 133 (same wording for PB1_2 and PB2_2 at the bottom of the image).
  • Family: Thermal displacement: the deviation follows the morning camera wave and comes back with it; it is the old content of 84 (PB1_2 alone by day, Visibility) merged here: one event, one label.
  • Measured: thermal.py: PB2_2 corr_y beyond 5 noises on 15 consecutive clean points 06:45-11:30 (max +9.4); PB1_2 corr_x run of 2 (05:15-06:30, max −8.7 at 08:45), corr_y run of 2 (max +6.0); the other targets ≤ 3.5 noises. Change: description only (adds the corr_x dip of PB1_2 and says PB2_2's spot is partly covered).
  • Images: none new (see 113 for the kerb).
  • Confidence: medium (the wave is just over the threshold on PB1_2; clear on PB2_2, itself under its obstruction 81).

29/09 05:15 → 12:30 (WILDSTRUBEL-113, corrected)

  • Phenomenon of the catalogue: The camera moves through a thermal effect.
  • Precedents: as 83.
  • Family: Thermal displacement rather than Visibility (old 112): the deviation of PB1_2 follows the camera wave, its raw position leaves the rigid movement of the targets at the same hours, the two targets closest to the camera (PB1_2 13.8 m, PB2_2 16.5 m) deviate, the farther ones do not; its error_value by day (0.26 → 0.40) rises like PB2_2's (0.40) and PB4's (0.45 → 0.68), not more (summary.py error_value, 1 h).
  • Measured: PB1_2 corr_x −0.07 at 05:15, −0.16 at 05:45, −0.15 at 07:15, back to +0.01 at 09:15, against a night level ≈ +0.02 to +0.07 (run of 6 points beyond 5 noises 06:00-07:30, max −11.0 ≈ −1.1 mm); PB2_2 corr_x run of 3 (08:00-08:30, max −6.5). overview.py --rigid 2026-09-29T02:00Z: PB1_2 dx −0.055 (05:15) → −0.18 (07:15) → −0.01 (09:30) px; PB2_2 dx ≈ −0.1; PB4, PB_5 within ±0.05. Change: description only.
  • Images: thumbnails PB1_2, PB2_2, PB4 at 02:00, 07:15, 10:00, 14:00 (work/thumbs/…290200-290715-291000-291400.png): a bright sunlit kerb band next to PB1_2 at 07:15 and at 10:00, when the curve is back; spot sharp and whole.
  • Confidence: medium.

02/10 05:15 → 13:00 (WILDSTRUBEL-140, corrected)

  • Phenomenon of the catalogue: A reference loses its brightness, without an obstacle (Family of the cause: A reference dropped).
  • Precedents: the station's 135 and 148 (PA_8 poorly seen by day, Visibility, validated).
  • Family: Visibility, as 135 and 148.
  • Measured: start moved from 06:00 to 05:15: at 05:15 (first measurement after the No data 137), PA1_2 and PA_8 are rejected: PA2_2 corr_y +18.0, PB1_2 +12.9, PB_5 +9.9, PB2_2 +9.1 noises; at 05:30, PA_8 alone: PA2_2 corr_x +7.6, PB2_2 +5.8; these jumps were only under the removed shimmer label 139 and the thermal 138 ("labelled separately", which they were not). Then as before.
  • Images: none new.
  • Confidence: high.

03/10 10:45 → 11:15 (WILDSTRUBEL-151, corrected) and 03/10 11:45 → 12:15 (WILDSTRUBEL-152, corrected)

  • Phenomenon of the catalogue: A pass-by hides the targets for the length of one measurement, with a reference lost.
  • Family: Obstruction, unchanged.
  • Measured: 151: at 11:00, R2 rejected (with PA_8 out under 148): PB2_2 corr_x +6.5, PA2_2 corr_y +7.0, PB1_2 corr_x +5.6; the old text said "the other targets do not move". 152: at 12:00, PA1_2 is rejected too (inventory) with PA_8 out: PA3_2 corr_y +5.7, PA2_2 +5.3 (edge +7.8); the old text said "the other targets do not move". Bounds unchanged.
  • Images: as described by the old labels; PA1_2 at 12:00 not looked at, so no cause is stated for it.
  • Confidence: high.

05/10 15:30 → 06/10 05:45 (WILDSTRUBEL-168, corrected) and 06/10 06:00 → 06:30 (added, No data)

  • Phenomenon of the catalogue: The station shuts down for lack of energy; The station runs, the measurement is missing.
  • Family: No data.
  • Measured: the old label ran to 06:30 over the measurements of 05:45 and 06:00 (two measurements back = a resumption, labeling-rules.md section 3). Split: the night outage ends at 05:45; the 06:15 measurement alone is missing between 06:00 and 06:30. Metrics: stop after 15:00, back at 04:00 (temperature 20.5, voltage 12.44 V), 06:00 voltage 12.45 V; images at 05:06, 05:34, 05:45, 06:00, 06:17, 06:30.
  • Confidence: high.

24/09 04:45 → 05:15, 25/09 04:15 → 05:00, 27/09 04:15 → 05:30, 28/09 04:15 → 05:15 (added, Visibility)

  • Phenomenon of the catalogue: The sun in the field or behind a spot (Family of the cause: A reference dropped).
  • Precedents: the station's 85 (low sun on the fixed scene at sunrise, Visibility, validated) and the PA_8 labels 135, 148.
  • Family: Visibility: nothing passes, the low sun lights the background behind PA_8.
  • Measured: PA_8 (with PA_9) rejected at sunrise: 24/09 05:00 → PA2_2 corr_x +6.7; 25/09 04:30-04:45 → PA2_2 corr_x +7.7, PB1_2 +6.2, PB_5 +5.3, then PB2_2 +5.2, PA2_2 +5.2; 27/09 04:30-05:15 → PA2_2 corr_y +9.7, corr_x +8.1, PA3_2 +7.8, PB_5 +6.3; 28/09 04:30-04:45 → PB4 corr_x −14.2 (−0.25 px ≈ −2.3 mm, PB4 left under its threshold), corr_y −8.4, PA2_2 +6.4, PA3_2 +5.8, PB1_2 +5.4, PB2_2 +5.3; 05:00 PA2_2 corr_y −9.5 with PA_8 back at error 0.57 (threshold 0.6). These were inside the removed shimmer and thermal labels only.
  • Images: PA_8 at 04:15, 04:45, 05:00, 05:30 on 27/09; at 05:00 on 24/09; at 04:30 on 25/09: bright streaks around PA_8 at sunrise, dark at 04:15, fading by 05:30. 28/09 not looked at: cause deduced.
  • Confidence: high (24, 25, 27/09), medium (28/09).

01/10 13:45 → 14:15, 04/10 11:45 → 12:15, 06/10 08:00 → 08:30 (added, Visibility)

  • Phenomenon of the catalogue: A reference loses its brightness, without an obstacle.
  • Precedents: 135, 140, 148.
  • Family: Visibility.
  • Measured: PA_8 rejected one measurement: 01/10 14:00 → PA2_2 corr_x +7.4, PB1_2 +6.3, PA_9 +5.6, PA3_2 +5.3; 04/10 12:00 → PA2_2 +7.5, PB1_2 +7.3, PB_5 +6.3, PB2_2 +5.5; 06/10 08:15 (R3 taken) → PB1_2 corr_x −30.0 (−0.59 px ≈ −3.2 mm), PA2_2 +14.7, PA_9 +13.1, PA_9 corr_y +5.6, PB1_2 −5.1.
  • Images: PA_8 thumbnails at 01/10 14:00 (faint spot, dark background), 04/10 12:00 (faint on light background), 06/10 08:15 (barely visible) and 08:30 (sharp).
  • Confidence: high.

05/10 10:15 → 11:15, 05/10 13:15 → 13:45, 06/10 08:45 → 09:45 (added, Visibility, R1)

  • Phenomenon of the catalogue: A reference loses its brightness, without an obstacle (boundary: a reference at its threshold, nothing in front: Visibility, families.md).
  • Family: Visibility; no object seen in front of R1.
  • Measured: 05/10: R1 rejected 10:30-10:45 → PB4 corr_x +6.0; at 11:00 the vehicle of the removed 167 (PA3_2, PA_9, PB4 rejected, no effect beyond 4.6 noises), included as the same continuous episode. 05/10 13:30: R1 rejected alone → PB4 +6.5, PB_5 +5.3. 06/10 09:00: R1 rejected → PB1_2 corr_x −28.6, PB2_2 corr_y +9.5, PA3_2 +7.8, PB1_2 corr_y −7.6, PA2_2 −6.7/+6.2, PB_5 +6.2, PB2_2 corr_x −5.7; 09:15 PA3_2 +6.7, PA2_2 +6.2 with all references in; PA3_2 back at 09:45 (0.22 → 0.20 → 0.13 → 0.06).
  • Images: R1, R2 thumbnails at 05/10 10:15, 10:30, 10:45, 13:30 and 06/10 09:00: a dark band thickening above R1 at 10:30-10:45; at 13:30 R1 and R2 much brighter and larger than usual; at 09:00 on 06/10 R1 faint.
  • Confidence: medium (cause not established for 10:30; a shadow is possible).

06/10 05:45 → 06:00 (added, Visibility, R2)

  • Phenomenon of the catalogue: The sun in the field or behind a spot.
  • Family: Visibility.
  • Measured: first measurement after the night outage, R2 and PB4 rejected (as every morning at this hour): PB_5 corr_x +7.1; back at 06:00. Start at the resumption measurement (labeling-rules.md, section 2).
  • Images: R2 at 05:45: bright streaks of the low sun around it.
  • Confidence: high.

The 42 removed labels

  • Daily heat shimmer, Visibility (82, 90, 97, 101, 105, 114, 122, 134, 139, 146, 156, 164, 170): removed. Outside the brief labelled moments and the instants with a reference other than R3 out, the daytime agitation (deviation from the median of 4 clean neighbours on each side) reaches at most 3 to 5.5 noises per target and day, 0 to 3 points just above 5 per day over all targets (work/days_dev.txt), against 1.5 to 3 at night; on the rolling level of 8, the single points above 5 without any reference event are 25/09 11:45 PB4 +5.4, 13:45 PA_9 −5.0, 27/09 13:00 PB4 +5.3, 04/10 06:15 PA_9 +5.1, 06:30 PB4 +5.2, 10:30 PB_5 +5.3, 11:45 PA2_2 −5.1 (section 2). What these labels held beyond 5 noises were the reference losses at sunrise and by day, now labelled apart (added labels and 140).
  • Daily camera cycle, Thermal displacement (91 25/09, 98 26/09, 102 27/09, 145 03/10, 155 04/10, 163 05/10): removed; the corrected curves keep no wave beyond 5 noises: no two consecutive clean points beyond 5 on any target (work/thermal.txt): 25/09 max PA3_2 +7.2 single (PB2_2's +13.5 is its step of 81, not the wave); 26/09 PA3_2 +5.4 single; 27/09 PB2_2 corr_y +7.7 and corr_x −7.4 single, PB1_2 corr_y +5.3 single, PB2_2 and PB1_2 corr_y peaks ≈ +4 to +5 on the 5-point median; 03/10 PB1_2 corr_y +6.6 single (afternoon rest); 04/10 PA2_2 corr_y −7.6 single, PB1_2 corr_x −5.0 on two points; 05/10 PB_5 +6.5 single. The camera wave is real every day (pose, as described), but compensated: not labelled (labeling-rules.md, section 1). Note: 102 is cited in reading-rules.md and stations.md as the source of the expert's word ("a road, the movement of the camera"); its removal drops that id from auto/.
  • PB1_2 alone by day, Visibility (84 24/09, 112 29/09): removed, merged into 83 and 113 (same deviation, same hours as the camera wave; error_value not higher than its neighbours'; the kerb band is lit after the curve has come back, 113).
  • Passages and one-measurement losses under the threshold (max rolling / edge, in own noises): 87 (PB_5 2.3 / PB2_2 5.2 at the edge only; the PB2_2 value at the bound is the end of the 86 spike), 88 (PA_9 4.6 / 4.8), 92 (PB4 3.1 / 3.6), 93 (2.3 / 2.6), 95 (3.5 / 3.3), 99 (3.2 / 4.4), 100 (3.0 / 3.3), 104 (3.5 / 3.9), 108 (PA_9 3.8 / 4.6; two references lost at 06:00, the remaining targets under 5), 109 (3.3 / 2.8), 110 (3.3 / 3.6), 115 (PB1_2 5.4 / 4.6, a single point at the threshold), 117 (PA2_2 5.3 / 4.7, idem), 124 (2.8 / 4.9), 126 (2.8 / 2.9), 128 (3.9 / 4.0), 129 (4.7 / 4.7), 149 (the PA jump at 08:45 is PA_8 coming back, held by 148; PB1_2 and PB2_2 only missing), 158 (4.7 / 5.3, PB3_2 lost without effect), 159 (4.5 / 5.2, idem), 167 (4.6 / 4.4; replaced by the R1 label of 05/10 10:15 → 11:15 which includes its vehicle). Their causes (vehicles, passers-by) are right; what they did is missing points only.

2. What I left without a label

  • Daytime agitation every day from sunrise to mid-afternoon: 3 to 5 noises (≈ 0.05 to 0.13 px, ≈ 0.4 to 1.5 mm), single points at 5.0 to 5.4 noises without any reference event (listed above); 25/09 06:45 PA3_2 corr_y +5.1 (PA_8 rejected, single point), 29/09 05:15 PA_9 corr_x +5.5 (first measurement after the No data 111, R2 rejected as every morning), 03/10 08:00 PA3_2 corr_y −5.3 (between 147 and 148), 06/10 11:00 PA_9 −5.0 (end of 174): single points at the threshold, not significant.
  • Camera wave of 25, 26, 27/09 and 03, 04, 05/10: compensated (deviations above). On 27/09, PB1_2 and PB2_2 corr_y peak at ≈ +4 to +5 noises (≈ +0.45 to +0.5 mm), just under.
  • PB1_2 corr_x after 131: lower by ≈ 6 noises on the evening of 01/10, back to its level by 02/10 (night levels +0.031 → −0.090 → +0.029): the drop is bounded by 131; the return is hidden in the outages and the PA_8 losses; not labelled.
  • References: their corr (PA_8, PA1_2, PB3_2 higher by 6 to 13 noises by day, days.py byhour) is not labelled; it shows the base deforming by day, compensated.
  • Rejections alone (PB3_2 at 09:00 and 11:00 on 04/10, the passages of the removed labels): missing points.

3. The documentation

  • Helped: labeling-rules.md section 1 (the 5-noise rule, thermal only when the curves keep it, missing points), section 2 (bounds at a resumption, a rejected reference at a bound), section 3 (two measurements make a resumption: the split of 168); families.md boundaries (a reference at its threshold with nothing in front: Visibility); tools/DATA.md (mm per px, distances: the two nearest targets carry the camera remainder, which settled 84/112 against 83/113).
  • Missing: the rule for the thermal threshold does not say whether a single agitated point beyond 5 noises from the night level is enough, or whether the wave must hold over several measurements. I required at least two consecutive clean points beyond 5 (or the 5-point median), so that the daytime agitation does not make a thermal label every day; a sentence in labeling-rules.md would settle it. Likewise, a single-point deviation of 5.0 to 5.5 noises with no cause is left unlabelled here (it is the tail of the daytime agitation); the rules do not say so.
  • Contradicted: 84 and 112 said PB1_2's error_value rises "much more than that of its neighbours": on 29/09 it does not (hourly medians). 151 and 152 said "the other targets do not move": they do, beyond 5 noises, through a reference lost at the same measurement. 138 and 139 said the first measurements without PA1_2 or PA_8 were "labelled separately": they were not (now under 140). 168 covered two measurements.
  • The measurement timestamps carry seconds (tools/DATA.md does not say it): a script that compares them with minute bounds must floor them, otherwise the point at a bound counts as inside.
  • Spanning labels: WILDSTRUBEL-46 and WILDSTRUBEL-130 (R3 Target issue) hold within the range: R3 rejected by day every measured day (inventory: 04:30-05:45 to 13:30-15:15), almost never at night. On 06/10, R3 is taken at 08:15 and 08:45 between its rejections; the jumps of PB1_2 at those instants come from PA_8 (08:15) and from its own step (173), not from R3.
  • Detectors: not read (DETECTORS: no).

Difficulties

  • precedents.py could not reach its Ollama service (http://10.0.2.140:11435 timed out; the container ollama-bge-m3 runs locally but its port is not published; localhost:11434 answers 404 for the model): no precedent from other stations; I used the station's own validated labels as cases.
  • A previous session had stopped on the usage limit (work/progress.md); its scripts labeldev.py, days_dev.py and outputs were reused and checked.

Sources read

  • instructions/analyst.md; tools/GUIDE.md; tools/DATA.md; labeling/families.md; labeling/labeling-rules.md; labeling/reading-rules.md; labeling/stations.md.
  • work/existing.jsonl (94 labels), work/spanning.jsonl (46, 130), work/progress.md.
  • data/WILDSTRUBEL/: inventory (work/inventory.txt), metrics (work/metrics.txt, metrics.py on 05-06/10), displacement exports (raw CSV around 04/10 04:42 and 06/10 05:45), images list, thumbnails (29/09 PB1_2-PB2_2-PB4; 27/09 PA_8-PA_9-PA2_2; PA_8-PA1_2 at six instants; R1-R2 at six instants); no weather/.
  • Tools: ranges.py --rolling 8, days.py (night levels, byhour), overview.py --rigid (29/09), summary.py, plot.py, thumbnails.py, metrics.py, inventory.py; scripts exceed.py, thermal.py, labeldev.py, days_dev.py, clipplot.py, build_labels.py.
  • labeling/scripts/precedents.py (unreachable), labeling/scripts/validate.py.