Files
aimsight-labeling/auto/analyses/NORDEND_2026-09-17T0545_2026-09-30T0415/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

11 KiB

Labels: NORDEND, 2026-09-17T05:45Z → 2026-09-30T04:15Z (UTC)

Task (2026-11-17): requalify the 24 labels of this analysis (work/requalify_input.jsonl) under the new rules (labeling-rules.md, section 1; families.md, "What is not labeled"): only what stands out from the noise of corr_x, corr_y is labelled; a missing point is not. Not a new reading: the earlier reading (its verdict table is kept in work/labels.before-requalify.jsonl and in the git history of this file) stands. Result: 15 kept as they are, 1 rewritten, 8 removed; 16 labels remain.

Weather: data/NORDEND/weather/ does not exist (no Swiss position): no MeteoSwiss weather. Station: no image-quality metrics (export peak to snr empty), quality.py not applicable. 1243 instants every 15 min, none missing; references CTR02, CTR03, CTR04; CTR01 a local reference outside the compensation, read as a target (stations.md).

1. My labels

Measures (section "Requalification"): the largest deviation of a target, in its own noises (median step between two accepted measurements over the range, as ranges.py), against three levels: the window median (ranges.py --k 5), the rolling median of ±4 measurements (--rolling 4), and for thermal labels the night level (01:00-05:00 UTC of the start day).

Thermal displacement: NORDEND-126, 128, 130, 131, 134, 136, 138, 140, 142, 143, 145, 146, 147 (kept as they are)

  • Phenomenon of the catalogue: The camera moves through a thermal effect
  • Precedents: none needed: the labels exist, reviewed 2026-11-17; the rule applied is the new section 1 of labeling-rules.md
  • Family: Thermal displacement, unchanged: the corrected curves keep the camera cycle beyond the noise every day
  • Measured: deviation from the night level, PR18 corr_x every day: 126 -10.6 (17/09 09:15), 128 -12.4 (18/09 08:00), 130 -11.6 (19/09 09:00), 134 -18.6 (21/09 09:00), 136 -18.7 (22/09 10:00), 138 -20.6 (23/09 08:00), 140 -9.0 (24/09 08:45, the weakest), 142 -19.1 (25/09 07:45), 143 -19.4 (26/09 09:30), 145 -16.7 (27/09 10:00), 146 -16.5 (28/09 09:30), 147 -13.8 (29/09 11:45); PR16, PR15 corr_x -8 to -15, PR18 corr_y +7.5 to +15. 131 (20/09, weak day): outside the visibility episode 132, PR18 corr_x stays above its night level all evening and night, 42 measurements from 19:45 to 21/09 06:45, up to +9.1 noises (01:00), PR16 +7.4 (30 measurements), PR15 +7.4: the slow return of the camera kept by the curves, as its description says
  • Images: none (not a new reading)
  • Confidence: high: every cycle is well beyond 5 noises, on several targets, for hours

NORDEND-127, 2026-09-17T20:00Z → 22:45Z, Obstruction (kept as it is)

  • Phenomenon of the catalogue: A spider web in front of the lens (a thread hides a few spots)
  • Family: Obstruction, unchanged
  • Measured: rolling: PR16 corr_y -8.7 (22:00), PR15 corr_x +8.3 (20:15), PR07 corr_x -7.9 (20:45), PR14, PR11 corr_x +6.9: two of the four largest isolated deviations (≥ 8 noises) of the whole range fall in this label. The rejections of PR07 are missing points, but the jumps of the others are the label
  • Confidence: high

NORDEND-132, 2026-09-20T08:45Z → 12:00Z, Visibility (kept as it is)

  • Phenomenon of the catalogue: A reference dropped: the base changes; Visibility drops over the whole scene
  • Family: Visibility, unchanged
  • Measured: PR08 corr_x -124 noises, PR06 -91, PR16 -73, PR05 -65, PR14 -53 (window level), CTR04 rejected, compensation on two references
  • Confidence: high

NORDEND-264, 2026-09-17T14:30Z → 15:00Z, Visibility (rewritten)

  • Phenomenon of the catalogue: The sun in the field or behind a spot; a target left under the threshold keeps its jump
  • Family: Visibility, unchanged
  • Measured: CTR01 corr_x -6.3 noises rolling (-6.0 window) at 14:45, not rejected (error_value just under threshold); the other targets at most 4.1 (PR17 corr_y)
  • Rewritten: the substance holds (CTR01 beyond 5 noises), but its description cited "the other sunny afternoons, where CTR01 is rejected (labelled separately)": those labels are removed. The sentence now says those afternoons only miss CTR01's curve. It had no review field
  • Confidence: medium: one point, just above the threshold, but at the sun hour with the cause seen on the thumbnail

2. What I left without a label

Removed labels (missing points only; no effect on the other curves beyond the noise). Baseline: outside the morning of 20/09, the targets (PR04 aside) make 121 isolated rolling exceedances of 5 noises over the 13 days, about 9 a day, almost all single measurements of 5 to 7 noises scattered over all targets and hours; a single point of 5 to 6 noises during an episode is that baseline, not an effect of the episode. CTR01 is a local reference outside the compensation: its rejection cannot move the other targets.

Id Range (UTC) Family Why removed Largest deviation of another target (rolling ±4)
NORDEND-129 18/09 14:45 → 16:00 Visibility CTR01 rejected (sun): missing points PR02 corr_y 5.0 (15:30), single point
NORDEND-133 20/09 15:15 → 15:45 Visibility CTR01 rejected one measurement: missing point PR09 corr_x 6.0, PR02 corr_x 5.8, corr_y 5.6, all at 15:30, error_value at level (0.07 to 0.09)
NORDEND-135 21/09 14:30 → 16:15 Visibility CTR01 rejected: missing points PR16 corr_y 6.2 (15:00), single point
NORDEND-137 22/09 14:30 → 16:15 Visibility CTR01 rejected: missing points PR01 corr_y 6.1 (15:30), PR02 corr_y 5.6 (15:00); CTR01 corr_y 5.5 at 14:30, the bound measurement
NORDEND-139 23/09 14:15 → 15:45 Visibility CTR01 rejected two measurements: missing points targets ≤ 5.1 (PR13 corr_y 14:30); the references CTR02 corr_x jump (10 noises, 15:00-15:15) is a reference, no target follows (≤ 4.8)
NORDEND-141 24/09 11:15 → 26/09 12:45 Obstruction cobweb: CTR01, PR09 rejected almost without stopping, PR01 at times: missing points; CTR02, PR17 error_value up without effect PR09 corr_y 7.2 at 26/09 08:00 (an accepted point between rejections), but PR09 makes -6.5 at 24/09 09:30, before the web: its usual scatter; PR18 corr_x 7.8 at 25/09 07:45 is the camera cycle (142)
NORDEND-144 26/09 14:30 → 16:00 Visibility CTR01 rejected: missing points PR01 corr_y 5.6 (15:15), single point
NORDEND-148 29/09 14:30 → 16:00 Visibility CTR01 rejected: missing points PR01 corr_y 6.0 (15:15), single point

The 141 episode is not a Target issue: CTR01 and PR09 are rejected almost always only during the two days of the web, seen on the images, not under the ordinary regime.

Remark: PR01 corr_y makes a single point of 5 to 7 noises (window level) around 15:15-15:30 on several sunny afternoons (22, 24, 26, 29/09), the hour CTR01 drowns in the sun; at the threshold, one measurement, error_value at its level: not labelled.

What the earlier reading already left without a label stays so (one-measurement error_value rises; 25/09 14:00 common x dip of 5 to 7 noises under 142; 22/09 10:00-10:30 under 136; morning steps of the cycle; references' corr of thousandths of a pixel).

Requalification

Commands (from the repository root; outputs in work/):

  • python3 tools/scripts/ranges.py data/NORDEND --columns corr_x,corr_y --k 5 --start 2026-09-17T05:45Z --end 2026-09-30T04:15Z → work/rq_ranges_k5.txt
  • the same with --rolling 4 → work/rq_ranges_k5_roll4.txt
  • python3 scripts/requalify.py work/requalify_input.jsonl (per label: top deviations against the night, rolling and window levels, and the rejections) → work/rq_per_label.txt; the same on 131 split around 132 (work/rq_131_parts.jsonl)
  • python3 scripts/look.py <start> <end> 4.5 on 14:15-16:15 of each afternoon from 17 to 29/09 (per instant, the sensors beyond 4.5 noises of the rolling level, the rejections)
  • python3 labeling/scripts/validate.py labels.jsonl: 16 labels, 0 errors

Noises (corr_x, px): PR18 0.0068, PR16 0.0055, PR15 0.0052, PR14 0.0049, CTR01 0.0094, PR08 0.0038.

Id Verdict Largest target deviation (noises)
NORDEND-126 kept PR18 corr_x -10.6 night level
NORDEND-264 rewritten CTR01 corr_x -6.3 rolling
NORDEND-127 kept PR16 corr_y -8.7, PR15 corr_x +8.3 rolling
NORDEND-128 kept PR18 corr_x -12.4 night level
NORDEND-129 removed 5.0 (missing points of CTR01)
NORDEND-130 kept PR18 corr_x -11.6 night level
NORDEND-131 kept PR18 corr_x +9.1, 42 measurements (night return)
NORDEND-132 kept PR08 corr_x -124
NORDEND-133 removed 6.0 single instant (missing point of CTR01)
NORDEND-134 kept PR18 corr_x -18.6 night level
NORDEND-135 removed 6.2 single point
NORDEND-136 kept PR18 corr_x -18.7 night level
NORDEND-137 removed 6.1 single point
NORDEND-138 kept PR18 corr_x -20.6 night level
NORDEND-139 removed 5.1
NORDEND-140 kept PR18 corr_x -9.0 night level
NORDEND-141 removed PR09 7.2 single accepted point, within its pre-web scatter
NORDEND-142 kept PR18 corr_x -19.1 night level
NORDEND-143 kept PR18 corr_x -19.4 night level
NORDEND-144 removed 5.6 single point
NORDEND-145 kept PR18 corr_x -16.7 night level
NORDEND-146 kept PR18 corr_x -16.5 night level
NORDEND-147 kept PR18 corr_x -13.8 night level
NORDEND-148 removed 6.0 single point

3. The documentation

  • Helped: labeling-rules.md section 1 and families.md "What is not labeled" (a missing point is not labelled; thermal only when the corr keep it), the boundary "A gap" (the episode labelled only if it moves the other curves).
  • Unclear: the threshold "at least one target beyond 5 noises" with the noise taken as the median step: at this station about 9 single points a day pass it anywhere. The rule does not say whether a single point at 5 to 6 noises during a missing-points episode is "the episode moving the other curves"; I judged by the baseline and by the mechanism (CTR01 outside the compensation). If the user reads the rule literally, 133, 135, 137, 144, 148 would come back rewritten on single points of PR01, PR02, PR09, PR16.
  • NORDEND-264 has the same weakness (one point at 6 noises), kept because the moving curve is the one of the sun episode itself, with the cause seen.

Difficulties

none

Sources read

  • instructions/analyst.md, the request, work/requalify_input.jsonl, the earlier analysis.md and report.md of this folder
  • tools/GUIDE.md, tools/DATA.md; labeling/families.md, labeling/labeling-rules.md, labeling/reading-rules.md, labeling/stations.md (in full)
  • data/NORDEND/targets.csv (distances, mm per px); inventory.py (window)
  • Tools: ranges.py (k 5, plain and --rolling 4); my scripts scripts/requalify.py, scripts/look.py (through tools/scripts/data.py)