Files
aimsight-labeling/auto/analyses/NORDEND_2026-08-10T0115_2026-08-23T0515/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

15 KiB
Raw Blame History

Labels: NORDEND, 2026-08-10T01:15Z → 2026-08-23T05:15Z (UTC)

This is a requalification of the 49 labels of this analysis (work/requalify_input.jsonl) under the new rules of labeling/labeling-rules.md section 1. Only what stands out from the noise of the corrected curves is labelled. I did not read the range again. 28 kept as they are, 0 rewritten, 21 removed. labels.jsonl holds the 28 kept labels exactly as they were: same bounds, family and description, review field included, id field removed. The detailed reading of the first pass is kept in work/rq/analysis.before-requalify.md.

Weather: data/NORDEND/weather/ does not exist, so there is no weather for this station. The first pass used MAR Les Marécottes (4.9 km), found by geocoding the place name. That weather is probably not the quarry's, and it is not used here.

Noise: each target's own noise at the native step over the batch, as ranges.py gives it. It is about 0.005 px on most targets (0.007 on PR04, PR18; 0.011 on CTR01), or 0.16 to 0.35 mm. Threshold: 5 noises.

1. My labels

The 28 kept labels are unchanged. For each one I give only what the requalification measured. The bounds, images and precedents of each label are in work/rq/analysis.before-requalify.md (§0 and §1).

2026-08-10T01:15Z → 2026-08-10T02:30Z (NORDEND-1)

  • Phenomenon of the catalogue: animal or insect very close to the lens
  • Precedents: none useful (requalification)
  • Family: Obstruction, unchanged
  • Measured: PR12 corr_x −8.8 noises at 02:00 (−2.4 mm, 2 measurements), −11.4 from the level of the 3 h before. All other targets stay ≤ 4.8.
  • Images: first pass (01:30 → 02:30)
  • Confidence: high

Daily camera cycles: NORDEND-3, 5, 9, 12, 15, 18, 20, 41, 46, 245

  • Phenomenon of the catalogue: thermal movement of the camera, poorly compensated

  • Precedents: none useful (requalification)

  • Family: Thermal displacement, unchanged

  • Measured: the hourly median of the corrected curve against the level of the 3 h before the start (scripts/thermal_check.py, scripts/thermal_all.py), over targets PR14 to PR18, which carry the cycle every day:

    • 3 (10/08): PR18 x −15.1, PR18 y +11.6
    • 5 (11/08): PR18 x −15.7
    • 9 (12/08): PR18 x −23.2, PR16 x −15.9
    • 12 (13/08): PR18 x −21.0 (−6.9 mm at 08:45)
    • 15 (14/08): PR18 x −19.1
    • 18 (15/08): PR18 x −11.8
    • 20 (16/08): PR18 x −7.5 (−2.8 mm, 09:30 → 13:15)
    • 41 (20/08): PR18 x −6.4 hourly, −8.6 at 07:15 (−2.4 mm); PR08 y +7.0
    • 46 (21/08): PR18 x −13.0
    • 245 (22/08): PR18 x −9.7. On the cold night, PR15 y and PR06 y stay at +7.0 and +8.2 noises (+1.3 and +1.8 mm) until 05:15.

    The ranges.py --rolling 1d output confirms each of them (work/rq/requalify.txt).

  • Images: first pass

  • Confidence: high. The curves keep the cycle beyond 5 noises.

2026-08-17T02:30Z → 2026-08-17T10:30Z (NORDEND-23)

  • Phenomenon of the catalogue: poor light on weak spots (cause not established)
  • Family: Visibility, unchanged
  • Measured: PR06 x −15.0 at 06:00 (−2.8 mm), PR08 x −11.1, PR16 y +10.7, PR16 x −10.4, PR07 x +10.0. These are the jumps that the description names. The label therefore holds on its effect on the curves, not on its rejections.
  • Confidence: medium (cause not established)

2026-08-18T06:15Z → 2026-08-18T12:15Z (NORDEND-27)

  • Family: Visibility, unchanged
  • Measured: PR16 y +20.4 at 07:45 (+3.6 mm), PR07 x +11.3, PR08 y +11.4, PR16 x ±11
  • Confidence: medium (cause not established)

2026-08-18T15:15Z → 2026-08-18T16:45Z (NORDEND-29)

  • Family: Visibility, unchanged
  • Measured: PR16 x −14.6 at 16:15 (−2.6 mm), PR13 x −10.2, PR06 y −8.9
  • Confidence: medium

2026-08-18T17:45Z → 2026-08-18T18:15Z (NORDEND-31)

  • Family: Obstruction, unchanged (the spider hides CTR02, so the compensation runs without it)
  • Measured: PR01, PR11, PR17 corr_y about +11 noises at 18:00 (PR01 +4.6 mm), all targets together
  • Confidence: high

2026-08-18T22:15Z → 2026-08-19T01:45Z (NORDEND-32)

  • Family: Obstruction, unchanged (CTR03 is hidden by a thread of cobweb)
  • Measured: PR18 y −15.9, PR17 y −14.9 (−5.1 mm), PR10 y −12.0, PR11 y −11.6, all on a plateau from 22:30 to 01:00
  • Confidence: high

2026-08-19T20:15Z → 2026-08-20T03:45Z (NORDEND-37)

  • Family: Visibility, unchanged (cobweb night)
  • Measured: PR07 x −10.3 at 20:15 (−2.1 mm). PR08 x is at −9.8 from 00:00 to 03:15 (−1.6 mm, 11 measurements).
  • Confidence: high

2026-08-20T16:15Z → 2026-08-20T17:15Z (NORDEND-44)

  • Family: Visibility, unchanged (sun reaching the wall)
  • Measured: PR15 x +16.0, PR14 x +15.9 at 16:30 (+3.4 mm), PR16 x +12.6
  • Confidence: high

2026-08-21T15:45Z → 2026-08-22T07:30Z (NORDEND-48)

  • Family: Visibility, unchanged (cobweb)
  • Measured: PR08 x −24.5 at 20:15 (−4.0 mm, 45 measurements until 06:30), PR14 x +23.0 at 18:45 (+4.9 mm), PR16 x +20.5, PR07 x +18.9
  • Confidence: high

No data: NORDEND-17, 33, 36, 38, 39, 244, 42, 45, 50

  • Family: No data, unchanged

  • Measured: inventory.py, "Steps absent for all sensors" over the window (work/rq/inventory_window.txt) gives exactly these nine gaps:

    • 15/08 01:00–01:30 (1 step)
    • 19/08 01:15–01:45 (1)
    • 19/08 14:15–20:15 (23)
    • 19/08 20:30–21:00 (1)
    • 20/08 00:30–01:00 (1)
    • 20/08 03:45–06:15 (9)
    • 20/08 08:00–09:00 (3)
    • 21/08 00:30–01:15 (2)
    • 22/08 16:30–17:00 (1)

    "Present instants where sensors have no measurement: none". So the station delivers nothing at all at these instants.

  • Confidence: high

2. What I left without a label

The 21 removed labels:

id bounds (UTC) family why deviation
NORDEND-2 10/08 05:30 → 06:00 Visibility missing points only: CTR01 rejected once at sunrise —
NORDEND-6 11/08 06:15 → 06:45 Visibility missing points only (CTR01, sunrise) —
NORDEND-8 12/08 05:45 → 06:15 Visibility missing points only (CTR01, sunrise) —
NORDEND-11 13/08 05:45 → 06:15 Visibility missing points only (CTR01, sunrise) —
NORDEND-14 14/08 05:45 → 06:30 Visibility missing points only (CTR01, two steps at sunrise) —
NORDEND-4 10/08 13:30 → 17:00 Visibility missing points only: CTR01 rejected on a sunny afternoon —
NORDEND-7 11/08 12:45 → 17:00 Visibility missing points only (CTR01, sunny afternoon) —
NORDEND-10 12/08 13:30 → 17:00 Visibility missing points only (CTR01, sunny afternoon) —
NORDEND-13 13/08 13:30 → 17:00 Visibility missing points only (CTR01, sunny afternoon) —
NORDEND-16 14/08 13:30 → 17:00 Visibility missing points only (CTR01, sunny afternoon) —
NORDEND-19 15/08 13:00 → 17:00 Visibility missing points only (CTR01, intermittent afternoon) —
NORDEND-22 16/08 13:00 → 17:00 Visibility missing points only (CTR01, afternoon) —
NORDEND-26 17/08 13:30 → 15:30 Visibility missing points only (CTR01, afternoon) —
NORDEND-43 20/08 14:15 → 14:45 Visibility missing points only (CTR01, one step) —
NORDEND-21 16/08 09:15 → 09:45 Visibility missing points only (CTR01, dull morning, one step) —
NORDEND-34 19/08 06:45 → 12:15 Visibility missing points only (CTR01, dull morning) —
NORDEND-47 21/08 12:15 → 13:30 Visibility missing points only (CTR01, two steps) —
NORDEND-30 18/08 17:30 → 19/08 04:00 Visibility missing points only outside its two nested labels (31, 32) max 4.6 noises (PR12, PR09), PR04 aside
NORDEND-25 17/08 07:30 → 18/08 08:15 Thermal displacement camera cycle absorbed by the compensation PR14–18 max 4.5 noises
NORDEND-28 18/08 08:15 → 19/08 07:15 Thermal displacement camera cycle absorbed by the compensation PR14–18 max 3.7 noises
NORDEND-35 19/08 07:15 → 20/08 03:45 Thermal displacement camera cycle absorbed by the compensation all targets ≤ 4.9 hourly

Details:

  • The CTR01 labels (17). CTR01 is a local reference read as a target (stations.md; pattern ref). It carries no inlier: only CTR02, CTR03, CTR04 do, so the compensation does not use it. Its rejection cannot move the other curves. It is rejected in 4,815 of 13,209 measurements (36 %). That is often, but not "almost always", so it is not a Target issue.
    • Inside the afternoon labels, other targets reach 7 to 10 noises around 15:00–17:00: PR01 y, PR05 y, PR10 x, PR14 x, PR18 y. That is the late-afternoon second bump of the camera cycle, which the thermal labels of those days already describe. It is not an effect of CTR01's loss.
    • Inside the sunrise labels, the deviations of PR18 and PR16 (8 to 16 noises) are the morning jump of the camera cycle, which the thermal labels hold.
    • Inside NORDEND-34, PR04 x +13.7 at 10:00 is PR04's own agitation, which its spanning Target issue NORDEND-24 holds.
    • Inside NORDEND-26, 43 and 47, nothing exceeds 5 noises apart from PR04.
  • NORDEND-30. Outside the spider's passage (31) and CTR03's loss (32), every target stays ≤ 4.6 noises: PR12 y +4.6, PR09 x −4.6. The exception is PR04 y, at +6.6 at 19:15 and −7.6 at 03:45 against the batch level, as isolated points. Against its local level (--rolling 1d) PR04 has no range there, and PR04 is under its own NORDEND-24. What this label added on top of 31 and 32 is CTR01, PR01, PR02, PR03, PR09 rejected at times: missing points. The cobweb's two effects that stand out stay labelled (31, 32).
  • NORDEND-25, 28, 35 (overcast days 17, 18, 19/08). The camera turns but the compensation absorbs it.
    • Hourly median of PR14–PR18 against the end-of-night level: 4.5 (PR17 y), 3.7 (PR18 x), 3.6 (PR18 y). Over all targets, outside the nested labels, the only values above 5 are:
      • PR04, which has its own agitation and NORDEND-24
      • PR13 y +6.1 at 17/08 16h, against a level taken during NORDEND-23's morning; ranges.py, plain and --rolling 1d, has no PR13 range that afternoon
      • one point at 19/08 05:15 (PR02 x +7.6, PR05 x +6.7), a single-step jolt at the end of night, not the cycle
    • The descriptions said "PR18 drops a little / markedly". Measured, that stays under the threshold.

Other things under the threshold, not labelled: the short single-step deviations of CTR02, CTR03, CTR04 at 5 to 7 noises (references; the dusk pose turn at about 18:00 on 10–12/08).

3. The documentation

  • labeling-rules.md §1 (new) decided everything. The point "a missing point is not labelled for itself" removes the 17 CTR01 labels. "Thermal only when the corrected curves keep it" removes the three overcast days.
  • stations.md (CTR01 a local reference read as a target) and the inventory (inlier on 3 sensors only) proved that CTR01 cannot move the other curves.
  • The thermal threshold is "5 noises from the night level". On NORDEND the pose comes back all night, and a level taken just before the start can fall inside a nuisance (17/08 morning). I read ranges.py --rolling 1d alongside so that a skewed level does not decide.
  • The noises are those of the whole batch, so a short label is judged against the station's ordinary noise, not against its own window.
  • Kept labels whose description now rests on something no longer labelled: the thermal labels 3, 5, 9, 12, 15, 18 say "(the rejections of CTR01 are labelled separately)", which is no longer true. That is a side note, not the substance. I kept them exactly as they were so that they keep the user's validation. They could be rewritten if the user prefers.

Requalification

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

python3 tools/scripts/ranges.py data/NORDEND --columns corr_x,corr_y --k 5 --start 2026-08-10T01:15Z --end 2026-08-23T05:15Z          > ranges_k5.txt
python3 tools/scripts/ranges.py data/NORDEND --columns corr_x,corr_y --k 5 --rolling 1d --start 2026-08-09T01:15Z --end 2026-08-24T05:15Z > ranges_k5_r1d.txt
python3 tools/scripts/ranges.py data/NORDEND --columns corr_x_mm,corr_y_mm --k 5 --start 2026-08-10T01:15Z --end 2026-08-23T05:15Z    > ranges_k5_mm.txt
python3 tools/scripts/ranges.py data/NORDEND --columns corr_x,corr_y --k 5 --rolling 8 --sensors PR04,PR01,CTR01,PR09,PR02,PR03 --start 2026-08-18T12:00Z --end 2026-08-19T08:00Z
python3 tools/scripts/inventory.py data/NORDEND --start 2026-08-10T01:15Z --end 2026-08-23T05:15Z > inventory_window.txt
python3 scripts/requalify.py     > requalify.txt      # per label: ranges overlapping it (plain, rolling 1d); thermal: hourly deviation vs end of night
python3 scripts/inwindow.py NORDEND-30:2026-08-18T22:15Z/2026-08-19T01:45Z NORDEND-30:2026-08-18T17:45Z/2026-08-18T18:15Z > inwindow.txt
python3 scripts/thermal_check.py > thermal_check.txt  # PR14-PR18, first 12 h, outside the nested labels
python3 scripts/thermal_all.py NORDEND-20 NORDEND-25 NORDEND-28 NORDEND-35 NORDEND-41 > thermal_all.txt
python3 labeling/scripts/validate.py labels.jsonl     # 28 labels, 0 errors

Values, per label: max deviation of a target in noises (in-window / thermal hourly).

id family max verdict
1 Obstruction PR12 x −8.8 kept
3, 5, 9, 12, 15, 18 Thermal PR18 x −15.1, −15.7, −23.2, −21.0, −19.1, −11.8 kept
20 Thermal PR18 x −7.5 kept
25, 28, 35 Thermal 4.5, 3.7, 4.9 (PR04 aside) removed
41 Thermal PR18 x −6.4 hourly, −8.6 point kept
46 Thermal PR18 x −13.0 kept
245 Thermal PR18 x −9.7; cold night PR06 y +8.2 kept
23, 27, 29, 44, 48 Visibility 15.0, 20.4, 14.6, 16.0, 24.5 kept
37 Visibility PR07 x −10.3, PR08 x −9.8 kept
30 Visibility 4.6 outside 31, 32 (PR04 aside) removed
31, 32 Obstruction PR11 y +11.1; PR18 y −15.9 kept
17 CTR01 labels Visibility missing points only removed
9 gaps No data all sensors missing kept

Difficulties

  • pandas is not installed in python3 or .venv, so the scripts use the standard library only.
  • data/NORDEND/weather/ is absent. The first pass's weather came from geocoding the place name, probably the wrong place.

Sources read

  • instructions/analyst.md; labeling/labeling-rules.md §1–5 (rules of bounds, events, description); labeling/stations.md (NORDEND row); tools/GUIDE.md (ranges, days, inventory, thumbnails, quality sections)
  • This analysis: work/requalify_input.jsonl, the previous analysis.md and report.md (copies in work/rq/)
  • data/NORDEND/displacement/NORDEND_advanced_202608*.csv through ranges.py, inventory.py and my scripts