Files
aimsight-labeling/auto/analyses/JUNGFRAU_2026-09-30T0000_2026-10-10T1615/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

19 KiB

Labels: JUNGFRAU, 2026-09-30T00:00Z → 2026-10-10T16:15Z (UTC)

Request: verify the 31 existing labels of the range (work/existing.jsonl), keep, correct, remove, add. DETECTORS: no. Result: 36 labels in labels.jsonl: 29 kept as they are, 2 corrected (description), 0 removed, 5 added. The spanning label JUNGFRAU-2 holds within the range.

Requalification, 2026-11-17 (new rules: only what stands out from the noise of the corrected curves; labeling-rules.md §1). All 36 labels re-measured with ranges.py --k 5: 35 kept as they are, 1 rewritten (JUNGFRAU-108, its description rested on a compensated camera movement), 0 removed. Details in "## Requalification" below.

Weather: data/JUNGFRAU/weather/, MeteoSwiss station Nyon / Changins (CGI), 2.0 km, hourly, every hour of the range present. Rain: 01/10 13:00-18:00 (up to 1.9 mm/h, humidity 96-97 % until 20:00, 81 % at 21:00); 07/10 10:00-11:00 (trace), 15:00-17:00 (up to 2.9 mm/h), 22:00-00:00; 08/10 02:00-07:00 (up to 4.3 mm/h, gusts 26-33 km/h), 08:00-09:00, 12:00-15:00 (2 mm at 12:00), gusts to 40 km/h on the morning of 08/10; humidity 66 % at 16:00 on 08/10; 10/10 03:00-04:00 (trace), 13:00 (0.3 mm), gusts 30-38 km/h all afternoon. Sun: 02/10, 05/10, 06/10 full; 03/10, 04/10, 09/10 partly; 30/09, 01/10, 07/10, 08/10 none.

Verdict per existing label

id verdict evidence
JUNGFRAU-1 kept night rise of corr_y of the bridge targets until morning, tilt rising, housing temperature 21.6→20.3 °C; pan leaves its level at 07:45
JUNGFRAU-3 kept pan +0.0067° at 13:45, scale +1.2e-4 at 14:15, tilt down; median corr_y of the targets -0.11 at 17:00; housing 20.7→25.1 °C; no sunshine at Nyon: "little sun" right
JUNGFRAU-4 kept corr_y rises all night, pose returns, temperature drops
JUNGFRAU-5 corrected (description) bounds right: first touched point 11:30 (ev of BL_C, PBL_C, BJ_G up, corr_y of BJ_G and BJ_C down), PM_C rejected 13:45→20:00 (26), ev back 20:45-21:30. Cause: rain 13:00-18:00 at Nyon, humidity 96-97 % until 20:00 and 81 % at 21:00, when the ev falls: "cause not established (perhaps rain or mist)" becomes "deduced from the weather"
JUNGFRAU-6 kept housing 20.6→16.6 °C, corr_y rises; overlaps JUNGFRAU-7 (the cycle is visible on both sides)
JUNGFRAU-7 kept REF_G ev 0.07→0.36 at 02:00, 0.52 02:15-03:15, 0.15 at 03:30, 0.05 at 03:45; REF_D ev up to 0.29; pan -0.005°, scale +6.8e-5 02:15-03:15; corr_x BJ_G +0.23, PBJ_C +0.13, back at 03:45; raw positions of the targets flat
JUNGFRAU-8 kept sunny (60 min/h from 10:00); median corr_x +0.27 at 14:45, corr_y -0.30 at 18:15; the step of the BJ targets at 07:15 (corr_y +0.13) comes with the scale rise at the start of the camera cycle: a remainder of the camera movement, not apart
JUNGFRAU-9, -10, -11 kept 02/10 11:30, 12:15, 13:30 absent for all sensors, no image at those minutes, last_sync entries at those minutes, BusVoltage 13.4-14.2 V; FLORIDA, CLARIDEN, NORDEND, WILDSTRUBEL measure at those minutes (NIEDERHORN, hourly, misses 12:00, not the same minute): only station affected
JUNGFRAU-12 to -16 kept daily warming and cooling, bounds on the turning points of pose and corr (days.py --bounds: pan leaves at 07:15 on 03/10, 08:30 on 04/10), sun as in the weather
JUNGFRAU-17 kept 05/10 06:15 absent, image 20261005T0615Z present, sync at 06:15, voltage normal, other stations present
JUNGFRAU-18 kept sunniest day (60 min/h 07:00-16:00), median corr_x +0.32 at 15:00
JUNGFRAU-19 kept 05/10 10:45 absent, no image, sync at 10:45
JUNGFRAU-20 kept REF_G ev 0.10→0.22 and REF_CT 0.09→0.24 at 16:15, BJ_G corr_x -0.13 in one step, partial return
JUNGFRAU-21 kept cooling 05/10 evening → 06/10 morning
JUNGFRAU-22 kept zone thumbnail 06:00: a truck trailer covers the place of REF_G; REF_G rejected (ev 0.97), the other two with empty inlier (0 retained); all targets jump (BJ_G corr_y +0.9), back at 06:15
JUNGFRAU-23, -24 kept warming and cooling of 06/10, sunny
JUNGFRAU-25 kept housing 17.3→19.5 °C, tilt 0.0401→0.0322 (05:00→12:00), corr_y of the BJ targets sinks; overcast; overlap with JUNGFRAU-26 stated
JUNGFRAU-26 corrected (description) bounds right: first agitated measurement 08:45 (BJ_G corr_x, corr_y drop after the ev rise of 08:15-08:30), ev back at 16:00 on 08/10 (humidity 66 % at 16:00). Rain 07/10 15-17 h, 22-00 h, 08/10 02-09 h (heavy, gusts), 12-15 h: cause "deduced from the weather", the expert's words kept
JUNGFRAU-27 kept corr_y of the bridge targets rises from 14:30 on 08/10 to 07:00 on 09/10 (median 0.19→0.475), housing 16.8→13.6 °C
JUNGFRAU-28 kept tilt leaves at 06:45, pan 07:00, scale 07:15; the BJ targets rise with the scale until 08:30 (remainder of the camera movement), the median of the targets sinks from 08:00; housing 13.9→22.6 °C
JUNGFRAU-29 kept 09/10 12:00 absent, no image, sync at 12:00
JUNGFRAU-30, -31 kept cooling night 09-10/10, warming 10/10 (pan from 08:00)
JUNGFRAU-32 kept 10/10 10:45 absent, no image, sync at 10:45
JUNGFRAU-2 (spanning) holds within the range PM_G rejected at all 1019 instants; its ev (median 0.41) between PM_D (0.32) and PM_C (0.47), threshold 0.3

1. My labels

Only the corrected and added labels are detailed; the kept ones are in the table above.

2026-10-01T11:15Z → 2026-10-01T21:30Z (JUNGFRAU-5, corrected)

  • Phenomenon of the catalogue: Rain
  • Precedents: JUNGFRAU-79 (reviewed: rain, wet road, spots shrinking): same way of stating the cause
  • Family: Visibility (whole scene, ev up on most sensors)
  • Measured: as the existing label; added the weather: Nyon 2 km, rain 0.2-1.9 mm/h 13:00-18:00, humidity 92-97 % 13:00-20:00, 81 % at 21:00; ev falls 20:45-21:30
  • Images: as the existing label
  • Confidence: high on bounds and family; cause likely (weather 2 km away, drops not seen)

2026-10-07T08:30Z → 2026-10-08T16:00Z (JUNGFRAU-26, corrected)

  • Phenomenon of the catalogue: Rain; visibility drops over the whole scene
  • Precedents: JUNGFRAU-79, JUNGFRAU-58 (reviewed)
  • Family: Visibility
  • Measured: ev of BL_C, PBL_C, PM_C 0.50→0.66 from 08:15, BJ_G corr_x -0.09 and corr_y -0.11 at 08:45; PM_C rejected 14:45-17:00, 17:30-00:00, 02:30-07:00; ev back at 16:00 on 08/10. Weather above
  • Images: as the existing label (16:00 wet road)
  • Confidence: high on bounds; cause likely rain, wind also measured (the expert's word kept)

2026-10-01T06:15Z → 2026-10-01T07:45Z (added)

  • Phenomenon of the catalogue: Visibility drops over the whole scene, without an obstacle (milky dawn): the closest, not a clear match
  • Precedents: none useful (the closest, JUNGFRAU-37 and -45, are the sun behind REF_G in the afternoon)
  • Family: Visibility: ev rising on most targets, sharp end, pose still (not Thermal: no camera movement; not Displacement: a whole-scene degradation signature)
  • Measured (scripts/dump.py): corr_y BJ_G -0.785 (06:00) → -0.832, -0.890, -0.892, -0.893, -0.971 (07:30) → -0.846 (07:45); BJ_C -0.556 → -0.759 → -0.646; BJ_D alike; corr_x BJ_G -0.62 → -0.77 → -0.65, BJ_D +0.10 then back; ev BL_C 0.51→0.54, PBL_C 0.46→0.52, PM_C 0.45→0.50, BJ_G 0.16→0.195, all back at 07:45; pan, tilt, scale constant (tilt 0.0348-0.0352 until 07:00), references flat. Humidity 81-83 %, no rain, no sun
  • Images: thumbnails 06:00, 07:00, 07:30, 07:45 and 30/09 07:30 (BJ_G, BJ_C, BJ_D, BL_D, REF_G): 06:00 night halos, 07:00-07:45 day, spots alike; whole image 07:30: clear scene, no fog
  • Confidence: medium (cause unknown; the bump is clear, about 8 noises on BJ_G corr_y, with a sharp return)

2026-10-01T07:45Z → 2026-10-01T11:15Z (added; JUNGFRAU-108, description rewritten at the requalification)

  • Requalification: kept, bounds and family unchanged, description rewritten. The old one said "the corrected curves stay almost flat" and rested on the compensated camera tilt, no longer labeled by itself. Against the level of the night before (02:00-06:15, --outside 06:15/11:15), the corrected curves do keep a slow drift: BL_G corr_y -0.213 → -0.263 px (09:00-11:15, 9 measurements, up to -18 night noises; about -9 against the label's own noise 0.0055), PBL_D corr_y +0.031 (07:45-11:15, 13 measurements, +12.7 noises), PBJ_C corr_y 0.09 → 0.00 and PM_D 0.08 → 0.047 from 10:15, BL_C corr_x 0.027 → 0.090 → 0.049. The new description is built on these curves.
  • Phenomenon of the catalogue: The camera moves through a thermal effect
  • Precedents: JUNGFRAU-25 (same station: weak warming of an overcast morning, ended where the visibility takes over)
  • Family: Thermal displacement (the pose follows the housing temperature, compensated; always labelled, D-117)
  • Measured: housing 20.4 (07:00) → 23.1 °C (10:00); tilt 0.0348 (07:00) → 0.0333 (07:15) → 0.0320 (10:00-11:15), pan -0.0509 (07:45) → -0.0491 (11:00), scale 1.000238 → 1.000297; raw positions of all sensors together; corr of the targets within ±0.03 of their level, median corr_y flat. Start: pan leaves its level at 08:00 (tilt at 07:15), the later → 07:45; end at the start of JUNGFRAU-5
  • Images: none (compensated movement)
  • Confidence: medium (small amplitude, but beyond 5 noises of the night level on tilt and pan)

2026-10-04T19:30Z → 2026-10-04T20:00Z (added)

  • Phenomenon of the catalogue: The base changes without a reference lost
  • Precedents: FLORIDA-74 (a reference measured shifted for one measurement, targets jump together), JUNGFRAU-67 (base hesitating)
  • Family: Displacement (no cause seen; REF_D far from its threshold, nothing in front: "Displacement without a cause", families.md)
  • Measured: 19:45: corr_x PM_C -0.07, BL_G -0.07, BJ_D -0.06, BJ_C -0.05, PM_D -0.05, PBL_D -0.05 (5-7 noises), REF_CT -0.04, REF_D +0.03; raw x of REF_D +0.08 px alone (targets 0); scale +1.5e-5; ev flat; back at 20:00
  • Images: fixed thumbnails REF_D, REF_CT 19:30, 19:45, 20:00: same spots, nothing in front
  • Confidence: medium (small, but common to all targets and explained by REF_D alone)

2026-10-09T00:45Z → 2026-10-09T01:15Z (added)

  • Phenomenon of the catalogue: The base changes without a reference lost
  • Precedents: as above
  • Family: Displacement
  • Measured: 01:00: corr_x BJ_D -0.10, PM_D -0.05, BJ_C, BL_D, PBL_D, REF_CT -0.04, REF_D +0.02; raw x REF_D +0.09 alone; ev REF_D 0.123 against 0.11; scale +3.1e-5; back at 01:15. BJ_C and BJ_D corr_y a little higher 01:00-01:30 (about three noises): left aside
  • Images: fixed thumbnails 00:45, 01:00, 01:15: same spots
  • Confidence: medium

2026-10-10T12:15Z → 2026-10-10T13:30Z (added)

  • Phenomenon of the catalogue: Visibility drops over the whole scene (short); nothing in the catalogue for wind
  • Precedents: JUNGFRAU-75 (reviewed: short drop in visibility, spots smaller)
  • Family: Visibility (ev up on almost all targets, spots smaller, sharp edges)
  • Measured: ev BL_G 0.51→0.62 (12:45), PBL_C 0.41→0.48, BL_C 0.48→0.56, PM_C 0.47→0.53, back 13:15-13:30; corr_y BJ_C +0.16→-0.14 and BJ_G +0.12→-0.13 between 12:30 and 12:45, back 13:00-13:30; corr_x BJ_G -0.13, -0.11 at 13:00, 13:15, PBJ_C -0.07, BL_G -0.06; scale +4.6e-5, +2.6e-5, -3.4e-5, +5.4e-5 (12:30-13:15). Weather: gusts 30-31 km/h, 0.3 mm at 13:00, sunshine 14 min at 12:00, 0 at 13:00
  • Images: thumbnails 12:15, 12:45, 13:00, 13:15, 13:45 (BJ_C, BJ_G, BL_G, PBL_C, BL_D): spots smaller 12:45-13:15; whole image 13:00: trees blurred by the wind, road dry
  • Confidence: medium (anomaly clear, cause not)

2. What I left without a label

Requalification: no label removed. Every movement label has at least one target beyond 5 of its own noises inside its bounds, every thermal label at least one target beyond 5 noises from the night level, every No data label is an instant missing for all sensors (values in "## Requalification"). No label rests on missing points alone: PM_C rejected in JUNGFRAU-109 and -111 is only part of those descriptions, the agitation of the other targets is the labeled effect (up to 25 noises).

From the first analysis:

  • 01/10 04:45: small camera step (pan, tilt, raw positions of all sensors together), corr unchanged except BL_D and BL_G corr_x by about five noises: well compensated.
  • 03/10 07:45, 07/10 06:00, 09/10 06:00: one-measurement blips of the BJ targets (corr_y +0.05 to +0.09, corr_x of BJ_G and BJ_D moving apart), four to five noises, nothing on the images (09/10 06:00 checked), around the change from night to day. Below what this station's labels mark; see the report.
  • 04/10 17:45-18:00: BJ_C corr_y -0.06 for two measurements, ev of the BJ targets up a little, at nightfall; 06/10 19:00: BL_C corr_y -0.07 once. One or two sensors, at most six noises.
  • 10/10 16:00: REF_G ev 0.18 once (sun behind it, as JUNGFRAU-20), PBJ_C corr_x -0.05 then back: four noises, mixed with the evening return, at the edge of the range.
  • REF_D ev peaks without effect on corr (06/10 05:15 up to 0.56, 05/10 06:45, 09/10 12:30, 10/10 11:00-13:30): not labelled (families.md, what is not labelled).

Requalification

Commands (outputs in work/requal/):

  • thermal labels: python3 tools/scripts/ranges.py data/JUNGFRAU --columns corr_x,corr_y --k 5 --min-measurements 6 --start <start> --end <end>; for JUNGFRAU-108 and -25 also against the night before: --start <night> --end <end> --outside <start>/<end>;
  • movement labels: ranges.py … --k 5 --start <start-3h> --end <end+3h> --outside <start>/<end> --rolling 4 (level = median of the 4 neighbours on each side, noise outside the label);
  • No data: inventory.py --start 2026-09-30T00:00Z --end 2026-10-10T16:15Z, missing steps for all sensors.

Largest target deviation inside each label (signed, own noises):

id family target, column deviation verdict
JUNGFRAU-1 Thermal PM_D corr_y -13.0 kept
JUNGFRAU-3 Thermal PM_D corr_y -16.0 kept
JUNGFRAU-4 Thermal PM_C corr_y -18.9 kept
JUNGFRAU-107 Visibility BJ_G corr_y / corr_x -14.5 / -13.9 kept
JUNGFRAU-108 Thermal BL_G corr_y vs night -18 (≈-9 on label noise); PBL_D +12.7 rewritten
JUNGFRAU-109 Visibility PM_D corr_x +25.3 kept
JUNGFRAU-6 Thermal PM_D corr_y -12.4 kept
JUNGFRAU-7 Visibility PBL_C corr_y; PBJ_C, BL_G corr_x -9.4; +8.9 kept
JUNGFRAU-8 Thermal BJ_D corr_y -29.0 kept
JUNGFRAU-9, -10, -11 No data 02/10 11:30, 12:15, 13:30 missing for all sensors - kept
JUNGFRAU-12 Thermal PBJ_C corr_y -36.4 kept
JUNGFRAU-13 Thermal PM_C corr_y -15.4 kept
JUNGFRAU-14 Thermal PBJ_C corr_y -28.8 kept
JUNGFRAU-15 Thermal PM_C corr_y -25.0 kept
JUNGFRAU-16 Thermal PBJ_C corr_y -27.8 kept
JUNGFRAU-110 Displacement PM_C corr_x -8.2 kept
JUNGFRAU-17 No data 05/10 06:15 missing for all - kept
JUNGFRAU-18 Thermal BJ_C corr_y -21.1 kept
JUNGFRAU-19 No data 05/10 10:45 missing for all - kept
JUNGFRAU-20 Visibility BJ_G corr_x -5.8 kept (just above)
JUNGFRAU-21 Thermal PBJ_C corr_y -45.8 kept
JUNGFRAU-22 Obstruction PBJ_C corr_x; BJ_G corr_y -65.4; +50.2 kept
JUNGFRAU-23 Thermal PM_D corr_y -23.0 kept
JUNGFRAU-24 Thermal BJ_D corr_y -27.7 kept
JUNGFRAU-25 Thermal PM_C corr_y; before the rain (06:30-08:30, vs night) PM_C corr_x, PM_D corr_y -13.3; -14.7, +10.9 kept
JUNGFRAU-111 Visibility BJ_G corr_x +15.0 kept
JUNGFRAU-27 Thermal BL_G corr_y -34.4 kept
JUNGFRAU-112 Displacement BL_D corr_x -10.0 kept
JUNGFRAU-28 Thermal BL_C corr_y -16.2 kept
JUNGFRAU-29 No data 09/10 12:00 missing for all - kept
JUNGFRAU-30 Thermal PBJ_C corr_y -29.1 kept
JUNGFRAU-31 Thermal BJ_G corr_y -14.0 kept
JUNGFRAU-32 No data 10/10 10:45 missing for all - kept
JUNGFRAU-113 Visibility BJ_G corr_y -10.6 kept

The thermal deviations are against the median of the label window (≈ half the cycle amplitude), so the deviation from the night level is larger still.

3. The documentation

  • Requalification: labeling-rules.md §1 (5-noise threshold, thermal kept only on the corrected curves, missing points not labeled) decided JUNGFRAU-108: kept, but its old wording (compensated camera, "D-117", "always labelled") no longer fits the rule and was rewritten on the curves.

  • families.md, boundaries "A hindered reference" and "A jump then the return", decided the Displacement family of 04/10 and 09/10 (no cause seen, reference not at its threshold).

  • labeling-rules.md §1 ("thermal movement is always labeled", the 5-noise rule) decided the 01/10 morning camera label; §2 "the later of tilt and pan" its start.

  • reading-rules.md §5 ("what moves in the image is read before the cause"): the raw positions showed REF_D alone shifting at 19:45 and 01:00.

  • The weather, new for this station, settles the cause of the two long Visibility episodes better than the images; the existing descriptions predate it. On 01/10 the end matches the humidity drop to the quarter hour.

  • Missing: the catalogue has no phenomenon for a brief degradation at daybreak (01/10), nor for wind. tools/DATA.md says the centroid offset changes at night when the spot is saturated: the BJ targets (large, close spots) may carry a night/day offset; not established.

  • Precedents of the same station (JUNGFRAU-25, -75, -79) helped for the form; none for the dawn bump; none contradicted by the data.

  • One-measurement gaps: "only station affected" checked against FLORIDA, CLARIDEN, NORDEND, WILDSTRUBEL at the same minutes: right.

Difficulties

  • days.py --bounds output exceeds a call: read through grep. plot.py takes --zone once per zone.

Sources read

Requalification: instructions/analyst.md, labeling/labeling-rules.md §1-2, tools/GUIDE.md (ranges, summary, inventory), work/requalify_input.jsonl, this analysis, the old report.md; data: ranges.py per label, summary.py (01/10 morning), inventory.py. First analysis: agents/agents/analyst.md, 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, work/spanning.jsonl, labeling/auto/JUNGFRAU.jsonl (the station's other labels, for consistency), precedents (precedents.py, two queries). Data: inventory, overview (range and 07-09/10), ranges --rolling 4 --instants 3, days --bounds, metrics (temperature, BusVoltage), weather CSVs, last_sync, other stations' measurement instants at the gap minutes, scripts/dump.py tables, thumbnails (01/10 dawn, 04/10 19:45, 09/10 01:00 and 06:00, 10/10 12:15-13:45, 06/10 06:00 zone), whole images 01/10 07:30 and 10/10 13:00.