Files
aimsight-labeling/auto/analyses/JUNGFRAU_2026-10-10T1615_2026-10-21T1030/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

10 KiB

Labels: JUNGFRAU, 2026-10-10T16:15Z → 2026-10-21T10:30Z (UTC)

Request: requalify the 33 labels of this analysis (work/requalify_input.jsonl) under the new rules (labeling-rules.md, section 1): only what stands out from the noise of the corrected curves. Not a new reading. Result: 33 kept as they are, 0 rewritten, 0 removed. The detailed reading of the previous pass (verification of 2026-11-17, its verdict table and measures) is kept in work/analysis.before-requalify.md.

Weather: MeteoSwiss Nyon / Changins, 2.0 km, 458 m (data/JUNGFRAU/weather/): dry 10-17/10 (a few tenths of a mm on 17/10), very humid nights 13-17/10; rain 18/10 11:00 → 19/10 03:00 and 20/10 15:00 → 21/10 10:00; sunny afternoons 12, 14, 15, 16/10. No image quality metrics on this camera.

Targets (inventory, targets.csv): PBJ_C 3 m (0.61 mm/px); BJ_C, BJ_D, BJ_G 8 m (1.6 mm/px); PM_C, PM_D 21 m (4.3 mm/px); BL_C, BL_D, BL_G 34 m (6.8 mm/px); PBL_C, PBL_D 39 m (7.8 mm/px). Noise of the corr over the range: 0.007 to 0.023 px depending on the target (work/requalify/full_rolling3.txt).

1. My labels

All the labels are kept with their id, bounds, family, description and review field. Per label, the largest deviation of a target (references excluded) in its own noises, at k 5 (scripts/requalify.py, work/requalify/deviations.txt): roll against the median of the 3 measurements on each side (as ranges.py --rolling 3), pre/post against the level of the hour before the start / after the end (the night level of a thermal phase). Measurements at the bounds excluded.

Thermal phases of the bridge (16 labels)

JUNGFRAU-33, 34, 100, 36, 38, 40, 41, 43, 46, 48, 51, 102, 56, 60, 63, 65

  • Phenomenon of the catalogue: The structure expands and contracts with the temperature; A rise over several days with the drop in temperature (JUNGFRAU-56)
  • Precedents: none needed: the rule is a threshold on the corr
  • Family: Thermal displacement, unchanged: the corrected curves keep the movement, far beyond 5 noises from the night level; it is not a camera movement absorbed by the compensation
  • Measured (deviation from the night level, noises): 33: 52 (PBJ_C corr_y); 34: 51; 100: 60; 36: 81 (PM_D corr_y); 38: 70; 40: 53; 41: 36; 43: 72; 46: 83 (with the reference loss inside; 51 at its start); 48: 66; 51: 36; 102: 41 (post level; starts at the resumption); 56: 71; 60: 27 with the BJ targets, 23 without them and before the BJ label (PBJ_C corr_y, 10:15 → 16:15); 63: 29 with, 11 without BJ and before the rain label (PBJ_C corr_y, 11:00 → 14:15); 65: 45 (BJ_C corr_y), 36 without BJ. In mm: about 1.5 to 4 mm a phase on PM_C, PM_D (corr_y), up to 2 mm on BL_G.
  • Images: none (previous pass)
  • Confidence: high: the weakest (63) still passes at twice the threshold outside any overlap

Sun behind REF_G, one-measurement jump (4 labels)

JUNGFRAU-37 (12/10 15:45 → 16:15), 45 (14/10 15:45 → 16:15), 50 (15/10 15:30 → 16:15), 55 (16/10 15:30 → 16:15)

  • Phenomenon of the catalogue: The sun in the field or behind a spot (here behind a reference: the base wobbles)
  • Family: Visibility, unchanged
  • Measured (rolling ±3): 37: BL_G corr_x 6.7, BJ_G 5.6 at 16:00; 45: PM_D corr_x 7.2 at 16:00 (15 against the hour before, PBJ_C); 50: PM_D corr_x 9.4 at 16:00; 55: PBJ_C corr_x 8.4, BL_G 8.4, BJ_G 7.8 at 15:45. Confirmed by ranges.py --rolling 3 --k 5 on the whole range. About half a mm to 0.7 mm on BL, PBL.
  • Confidence: high: above 5 noises on several targets at the same measurement

References seen less well at night (4 labels)

JUNGFRAU-39 (13/10 01:00 → 02:45), 42 (13/10 23:45 → 14/10 05:30), 101 (14/10 23:00 → 15/10 08:00), 57 (17/10 00:15 → 04:45)

  • Phenomenon of the catalogue: A reference loses its brightness, without an obstacle; Two references lost: the camera is blind (101)
  • Family: Visibility, unchanged; these are not labels of missing points: the episodes move the corr of the other targets (the rule's exception, their descriptions already state that effect)
  • Measured: 39: rolling 7.8 (BL_C corr_x 01:30), PM_C 7.7, PBL_C 7.6; 42: rolling 5.6 (BL_C corr_x 00:15), BL_G 5.4 at 02:45; against the levels outside, BJ_G corr_x 17.7 at 01:30, PBJ_C corr_y 16.5; 101: rolling 58 (BJ_G corr_y 01:45), up to about 3 mm on BL_G; 57: rolling 7.3 (PBJ_C corr_x 04:15), BL_C 6.5, PBJ_C 6.7 at 01:45.
  • Confidence: high for 39, 101, 57; medium for 42, whose rolling deviation is close to the threshold (5.6), but whose shift of BJ_G and PBJ_C against the levels before and after (16-18 noises) holds

Rain, the whole scene (2 labels)

JUNGFRAU-103 (18/10 10:00 → 19/10 10:15), 106 (20/10 14:15 → 21/10 09:15)

  • Phenomenon of the catalogue: Rain; Visibility drops over the whole scene
  • Family: Visibility, unchanged. The rejections of PM_C, PM_D, PBL_C that the descriptions mention are not what is labelled; the agitation of the corr is, and it passes far beyond the threshold
  • Measured: 103: rolling 24 (PM_D corr_x 21:15), 49 against the levels before and after; up to about 2.7 mm (BL_C corr_x). 106: rolling 17 (PM_C corr_x 21/10 08:00), 40 (BL_C corr_x 22:30); up to about 3 mm (BL_C corr_x).
  • Confidence: high

Near spots seen less well on a humid night (1 label)

JUNGFRAU-105 (19/10 16:15 → 20/10 11:00)

  • Phenomenon of the catalogue: Part of the scene goes dark at night (near spots)
  • Family: Visibility, unchanged
  • Measured: BJ_D corr_x 30 noises against the level before, 25 against the level after (06:30); rolling 6.6 (BJ_D 02:30). In mm, about 1 mm on BJ_D, BJ_G.
  • Confidence: medium-high (as before: the link with their error_value is strong, the mechanism not seen)

No data (6 labels)

JUNGFRAU-44, 49, 52, 54, 104, 62

  • Phenomenon of the catalogue: The station runs, the measurement is missing (44, 49); A gap common to several stations (52, 54, 104, 62)
  • Family: No data, unchanged: no measurement of any sensor at the missing instants (no accepted value inside 44, 49, 52, 54, 104; for 62, only the single measurement of 21:00 between two missing ones; inventory: image gaps at the same instants)
  • Confidence: high

2. What I left without a label

Removed labels: none. Every label passes the threshold of 5 own noises of at least one target over its range, or is a No data where the station delivers nothing. No label of this analysis rests only on missing points (the rejections of PM_C, PM_D, PBL_C in the rain and of REF_G, REF_D at night come with movements of the other curves beyond the noise), and no thermal label is a camera movement absorbed by the compensation (the corr keep 23 to 83 noises from the night level).

Left without label by the previous pass, unchanged: 17/10 16:00-21:15 (REF_G error_value up, corr at most 4-5 noises on one or two targets); 14/10 14:30 (one-measurement wobble, about 6 noises rolling, under JUNGFRAU-43); the daytime error_value rises of REF_D and REF_G; single-instant flags of the rolling ranges (10/10 18:00, 12/10 18:45, 15/10 22:45, 16/10 10:30). Not reopened (no new events sought).

3. The documentation

  • labeling-rules.md section 1 and families.md ("A gap", "What is not labeled") were clear for the decisions. The thresholds are generous for this bridge: the thermal phases are 25 to 80 noises, so no thermal label is threatened.
  • What is not said: over which noise to measure (that of the label window, or of the batch). I used the batch noise (the noise of ranges.py over the whole range), steadier than the noise of a window of three measurements; the window noise of ranges.py on 12/10 14:00-18:00 gives the same verdict for JUNGFRAU-37 (BL_G 7.9 noises).
  • For a label that overlaps another (60 and 105, 63 and 106), the rule does not say whether the threshold must be passed outside the overlap; I checked it outside, and it holds.

Difficulties

none

Sources read

  • instructions/analyst.md, the request, work/requalify_input.jsonl, the previous analysis.md and report.md (kept in work/)
  • tools/GUIDE.md, tools/DATA.md, labeling/families.md, labeling/labeling-rules.md, labeling/reading-rules.md, labeling/stations.md
  • data/JUNGFRAU: inventory (range), ranges (corr_x, corr_y, k 5, whole range, plain and rolling 3), the export through scripts/requalify.py and scripts/mm_magnitudes.py, weather of 18/10 (header)

Requalification

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

  • python3 tools/scripts/ranges.py data/JUNGFRAU --columns corr_x,corr_y --start 2026-10-10T16:15Z --end 2026-10-21T10:30Z --rolling 3 --k 5 --merge 0min → full_rolling3.txt; the same without --rolling → full_level.txt
  • python3 <folder>/scripts/requalify.py data/JUNGFRAU <folder>/work/requalify_input.jsonl 2026-10-10T16:15Z 2026-10-21T10:30Z → deviations.txt; with BJ_C,BJ_D,BJ_G excluded → deviations_without_BJ.txt; on the sub-ranges 19/10 10:15-16:15 and 20/10 11:00-14:15 (sub.jsonl)
  • python3 <folder>/scripts/mm_magnitudes.py data/JUNGFRAU <folder>/work/requalify_input.jsonl → mm.txt
  • python3 tools/scripts/inventory.py data/JUNGFRAU --start … --end … → inventory.txt
id family max deviation (noises) verdict
33 Thermal 52 (night level) kept
34 Thermal 51 kept
100 Thermal 60 kept
36 Thermal 81 kept
37 Visibility 6.7 rolling kept
38 Thermal 70 kept
39 Visibility 7.8 rolling kept
40 Thermal 53 kept
41 Thermal 36 kept
42 Visibility 5.6 rolling, 17.7 against outside levels kept
43 Thermal 72 kept
44 No data no measurement kept
45 Visibility 7.2 rolling kept
46 Thermal 51 at the start, 83 overall kept
101 Visibility 58 rolling kept
48 Thermal 66 kept
49 No data no measurement kept
50 Visibility 9.4 rolling kept
51 Thermal 36 kept
52 No data no measurement kept
102 Thermal 41 kept
54 No data no measurement kept
55 Visibility 8.5 rolling kept
56 Thermal 71 kept
57 Visibility 7.3 rolling kept
103 Visibility 24 rolling, 49 outside kept
104 No data no measurement kept
60 Thermal 23 outside the BJ label kept
105 Visibility 30 against the level before kept
62 No data no measurement (one between) kept
63 Thermal 11 outside the rain label, without BJ kept
106 Visibility 17 rolling, 40 outside kept
65 Thermal 45 kept

python3 labeling/scripts/validate.py <folder>/labels.jsonl: 33 labels, 0 errors.