Files
aimsight-labeling/auto/analyses/JUNGFRAU_2026-10-10T1615_2026-10-21T1030/analysis.md
T

17 KiB

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

Request: verify the 33 existing labels of the range (work/existing.jsonl): keep, correct, remove, add. DETECTORS: no. Result: 26 kept, 7 corrected, 0 removed, 0 added. The spanning label JUNGFRAU-2 (PM_G, Target issue) holds within the range.

Weather: MeteoSwiss station Nyon / Changins, 2.0 km, 458 m (data/JUNGFRAU/weather/, hourly). Dry from 10/10 to 17/10 except a few tenths of a mm on 17/10 (12:00-13:00, 17:00, 20:00-23:00); nights 13/10 to 17/10 very humid (86-94 %). Continuous rain from 18/10 11:00 to 19/10 03:00 (up to 3.6 mm/h, humidity 91-97 %); dry 19/10 04:00 to 20/10 14:00 (humidity 78-91 %); rain again 20/10 15:00 → 21/10 10:00 (up to 1.9 mm/h, humidity 90-98 %). Sunny afternoons on 12, 14, 15/10 (60 min/h) and 16/10 (32-50 min/h).

No image quality metrics on this camera (the six columns peak … snr are empty, inventory), so quality.py gives nothing; visibility was read on the error_value, the images and the weather.

Verdict per old id

Old id Verdict Evidence
JUNGFRAU-33 kept corr_y of the targets rises from 10/10 16:00 (median -0.17) to 11/10 07:00 (+0.26), tilt 0.0230 → 0.0353, housing temperature 18.6 → 12.5 °C; starts at the range start, continuing JUNGFRAU-31.
JUNGFRAU-34 kept corr_y max 11/10 07:00, sinks to its trough by 16:00 (-0.04, hourly), tilt 0.0353 → 0.0244, temperature up to 21.8 °C.
JUNGFRAU-35 corrected: start 18:00 → 16:15 The only thermal pair of the range that did not touch: from 16:15 the tilt rises again (0.0246 → 0.0269 at 18:00) and the temperature drops (17.5 → 15.5 °C) while the targets' corr_y stays at its trough (y_targets -0.10 to -0.145 from 14:45 to 18:15). The pose moves, the cooling has begun: the cut goes at the turning point of the tilt, in the flat trough (labeling-rules 2, the phases touch). Description adapted.
JUNGFRAU-36 kept corr_y 12/10 07:30 max (+0.36) → 17:15 min (-0.22), corr_x of the targets up to +0.31 at 15:00, temperature to 20.4 °C.
JUNGFRAU-37 kept REF_G error_value 0.232 at 16:00 (level 0.085, 53 noises); corr_x jump of the three references (5-8 noises) and BJ_G, BL_C, BL_G (5-8 noises) at 16:00 only; 15:45 and 16:15 calm (rolling ranges, k 4).
JUNGFRAU-38 kept corr_y 17:15 min → 13/10 07:00 max, tilt 0.0271 → 0.0380, temperature 16.6 → 11 °C.
JUNGFRAU-39 kept REF_G error_value 0.18 at 01:00, 0.51-0.55 from 01:15 to 02:15, 0.41 at 02:30, 0.09 at 02:45; corr_x of the references shift at 01:15 (REF_CT 0.095 → 0.148), back at 02:45; BJ_G, BL_G, PBJ_C, PBL_C shift with them. REF_D stays at 0.33-0.35 until 03:45 (back 04:15) with no visible effect on the corr after 02:45.
JUNGFRAU-40 kept corr_y max 07:00, min 18:00 (-0.13), tilt 0.038 → 0.0204, temperature up to 21.5 °C.
JUNGFRAU-41 kept corr_y flat about -0.08 until 01:00, then rises to +0.07 at 08:00; tilt 0.0223 → 0.0322.
JUNGFRAU-42 kept REF_G 0.295 at 23:45, 0.53 from 00:00; references' corr_x shifted from 00:00 to 05:15 (REF_CT 0.114 → 0.142, then 0.08), back at 05:30 when REF_D returns (0.17 at 05:15, 0.13 at 05:30).
JUNGFRAU-43 kept corr_y max 07:15 (+0.07), min 17:00 (-0.52); corr_x of the targets +0.26 at 14:00; temperature up to 25.9 °C (11:00).
JUNGFRAU-44 kept One missing instant 09:30, image gap 09:15 → 09:45 (inventory).
JUNGFRAU-45 kept REF_G error_value 0.124 at 16:00 (45 noises), common corr_x jump at 16:00 only.
JUNGFRAU-46 kept corr_y 17:00 min (-0.52) → rises all night; tilt 0.0303 → 0.0336 at 06:00; end at the morning turning point (tilt down from 07:30, temperature up from 08:00).
JUNGFRAU-47 corrected: end 07:45 → 08:00, description Rejections end at 06:15, but REF_D, accepted again, is at 0.55, 0.44, 0.16, 0.50, 0.62, 0.34 (06:30 → 07:45), 0.27 at 08:00, 0.157 at 08:15 (level 0.14). The targets stay shifted in jolts with it (BJ_D corr_x 0.39 → 0.47 at 07:15; PBJ_C corr_x -0.074 until 07:45, -0.024 at 08:00): 07:45 is still touched, the end is 08:00. The description now says REF_D is seen poorly in jolts until early morning.
JUNGFRAU-48 kept corr_y 08:30 → min 17:00 (-0.52), corr_x +0.31 at 14:00, temperature to 23.4 °C.
JUNGFRAU-49 kept One missing instant 09:45, image gap 09:30 → 10:00.
JUNGFRAU-50 kept REF_G 0.142 at 15:45, still high at 16:00, common corr_x jump.
JUNGFRAU-51 kept corr_y rises from 17:00 (-0.52) to 22:00 (-0.27), stops at the gap.
JUNGFRAU-52 kept 46 missing instants 22:45 → 10:30, image gap of 705 min (inventory); metrics present all night.
JUNGFRAU-53 corrected: description "the camera housing temperature is the highest of the batch" is false: hourly maximum 25.2 °C on 16/10 (14:00), 25.9 °C on 14/10 (11:00). Replaced by "rises clearly in the early afternoon". Bounds and the rest hold (corr_y -0.05 → -0.44 at 17:00).
JUNGFRAU-54 kept One missing instant 13:45, image gap 13:30 → 14:00.
JUNGFRAU-55 kept REF_G 0.17 at 15:45 (62 noises), corr_x jump of BJ_G, BL_C, BL_G, PBJ_C and the references at 15:45 (5-8 noises).
JUNGFRAU-56 kept corr_y -0.44 (16/10 17:00) → +0.21 (18/10 08:00), weak dip 17/10 13:00-15:00; temperature 20 → 13 °C with only a weak bump on 17/10 (18 °C).
JUNGFRAU-57 kept REF_G 0.155 at 00:30, 0.39-0.54 from 00:45 to 02:45, back 03:00; REF_D 0.18-0.44 from 00:30, back 04:30; references' corr_x shifted from 00:45, last touch 04:30 (REF_CT 0.065 at 04:15, 0.110 at 04:30, 0.126 at 04:45). The start at 00:15 is one step early (00:30 barely moves), within the uncertainty.
JUNGFRAU-58 corrected: description (cause) "Mist probable" is contradicted: rain measured at Nyon, 2 km away, from 11:00 to 19/10 03:00 (up to 3.6 mm/h, humidity 96-97 %); on the 14:00 image the road is wet and the scene slightly veiled; 21:00 image veiled in grey with halos; 19/10 12:00 image dry and sharp. Bounds hold: error_value of the targets up from 10:15 (BJ_C 0.205 → 0.215 → 0.256 at 10:30), still high at 08:45 before the gap, back at 10:15 (BJ_C 0.187). The precedent JUNGFRAU-79 (wet road, same station) gives the same cause.
JUNGFRAU-59 corrected: description A No data states the image of the instant (labeling-rules 5): the image of 09:00, the first missing measurement, exists (inventory: image gap 09:00 → 10:15); added, it confirms the station runs.
JUNGFRAU-60 kept corr_y of the targets 0.20 (10:00) → 0.00 (20:00), tilt 0.0207 → 0.0191 at 11:00, temperature 15.4 → 16.5 °C; weak but present. From 16:15, the BJ part of the sinking is the nuisance of the corrected JUNGFRAU-61 (overlap allowed). Confidence medium.
JUNGFRAU-61 corrected: family Thermal displacement → Visibility, start 20:00 → 19/10 16:15, description Housing temperature flat (14.0-15.0 °C), pose flat (tilt 0.0212-0.0227), corr_y of the other targets flat: no thermal cycle that night. What moves is BJ_C, BJ_D, BJ_G, in step with their own error_value (mean 0.19 at 16:00 → 0.27 at 00:00 → 0.29 at 07:00 → 0.20 at 11:00): BJ_D - BJ_G corr_x 0.81 → 1.69 → 1.86 → 1.00, BJ_C corr_y -0.14 → -0.57 → -0.59 → -0.21. "Ordinary error_value" in the old description is false. Humid night after the rain (86-90 %), 00:00 image slightly veiled. First touched point 16:30, back at 11:00.
JUNGFRAU-62 kept Missing instants 20:45 and 21:15; images at 20:30, 21:00, 21:30 only.
JUNGFRAU-63 kept Tilt 0.0205 (09:00) → 0.0164 (13:00), temperature 15.8 → 17.3 °C, corr_y 0.083 (11:00) → -0.046 (16:00).
JUNGFRAU-64 corrected: description (cause) Rain at Nyon at 15:00 (0.1 mm) and from 17:00 to 21/10 10:00 (up to 1.9 mm/h, humidity 93-98 %); error_values of the targets up from 14:30 (BJ_C 0.219 → 0.228 at 14:15 → 0.256 at 14:30), REF_D and REF_G from 17:00; 21/10 00:00 image veiled in grey, 09:00 image damp and duller. "Mist probable" replaced by the rain; bounds hold.
JUNGFRAU-65 kept corr_y -0.06 (17:00) → +0.25 (21/10 10:00), temperature 16 → 11.4 °C; mixed with JUNGFRAU-64, as it says.
JUNGFRAU-2 (spanning) holds within the range PM_G rejected at all 978 instants; its error_value (level 0.44) between those of PM_C (0.51) and PM_D (0.35), following their variations.

1. My labels

The 26 kept labels are not repeated here (verdict table above). The corrected ones:

2026-10-11T16:15Z → 2026-10-12T07:30Z

  • Phenomenon of the catalogue: The structure expands and contracts with the temperature
  • Precedents: none useful beyond the neighbouring labels of the station (JUNGFRAU-36/38, 40/41 touch)
  • Family: Thermal displacement: the pose follows the housing temperature
  • Measured: tilt 0.0246 (16:00) → 0.0269 (18:00) → 0.0380 (12/10 06:00); y_targets flat -0.10 to -0.145 from 14:45 to 18:15, then rising; temperature 17.5 °C (16:00) → 11.5 °C (12/10 06:00)
  • Images: none (nothing else moves)
  • Confidence: high

2026-10-14T23:00Z → 2026-10-15T08:00Z

  • Phenomenon of the catalogue: Two references lost: the camera is blind
  • Precedents: the existing label
  • Family: Visibility (kept): nothing in front of the references, a very humid night
  • Measured: REF_G 0.156 at 23:00 → rejected 01:45-04:00; REF_D rejected 02:00-06:15, then 0.55/0.44/0.16/0.50/0.62/0.34/0.27 (06:30-08:00), 0.157 at 08:15; no pose 02:00-04:00; targets jump at each switch (BJ_G corr_y +59 noises at 01:45); last shifted point 07:45
  • Images: not looked at again
  • Confidence: high on the end (REF_D and PBJ_C read at each step)

2026-10-16T10:30Z → 2026-10-16T16:45Z

  • Phenomenon of the catalogue: The structure expands and contracts with the temperature
  • Precedents: none useful
  • Family: Thermal displacement
  • Measured: temperature 20.4 → 25.2 °C (14:00), corr_y -0.05 → -0.44 (17:00), corr_x +0.17 at 14:00; the hourly maximum of the range is 25.9 °C, on 14/10 11:00
  • Images: none
  • Confidence: high

2026-10-18T10:00Z → 2026-10-19T10:15Z

  • Phenomenon of the catalogue: Rain; Visibility drops over the whole scene
  • Precedents: JUNGFRAU-79 (wet road at 15:00, rain, same station): same signature and cause; NORDEND-213 (rain then a night of poor visibility)
  • Family: Visibility: the whole scene
  • Measured: as the old label; rain at Nyon 11:00 → 19/10 03:00
  • Images: 14:00 (road wet, scene slightly veiled, near spots haloed), 21:00 (background grey, halos), 19/10 12:00 (dry road, sharp)
  • Confidence: high

2026-10-19T08:45Z → 2026-10-19T10:15Z

  • Phenomenon of the catalogue: A gap common to several stations
  • Precedents: the existing label
  • Family: No data
  • Measured: 5 missing instants 09:00-10:00; image of 09:00 present
  • Images: file listing
  • Confidence: high

2026-10-19T16:15Z → 2026-10-20T11:00Z

  • Phenomenon of the catalogue: Part of the scene goes dark at night / Visibility drops over the whole scene (here the three near spots above all)
  • Precedents: NORDEND-150 ("a poorer view at night of the left-edge targets", Visibility); LENZSPITZE-16 (the large nearby spots jump when visibility drops)
  • Family: Visibility rather than Thermal displacement (families.md, boundary "Thermal displacement or Visibility"): an error_value that rises clearly on the moving targets and a little on the others, a slightly veiled image, no temperature or pose change; there is no thermal cycle to explain the separation
  • Measured: evBJ (mean of BJ_C, BJ_D, BJ_G) 0.193 (16:00), 0.197 (16:30), 0.217 (18:00), 0.275 (00:00), 0.293 (07:00), 0.237 (10:00), 0.200 (11:00), then 0.19-0.21 until 14:00; BJ_D - BJ_G corr_x 0.81, 0.87, 1.08, 1.69, 1.86, 1.36, 1.00; BJ_C corr_y -0.14, -0.18, -0.30, -0.57, -0.59, -0.40, -0.21 (BJ_D and BJ_G corr_y the same way); other targets (BL_D, PBL_D, PBJ_C) 0.193 → 0.224 → 0.185; temperature 14.0-15.0 °C; tilt 0.0212-0.0227; references calm (REF_G 0.07-0.09)
  • Images: 20/10 00:00 (background slightly grey, nothing in front); BJ thumbnails 16/10 20:00, 19/10 14:00, 20/10 00:00 and 07:00 (large saturated spots, background greyer on the night of 20/10)
  • Confidence: medium-high: the step-by-step link with the error_value is strong (scripts/bj_spread.py), the physical mechanism is not seen

2026-10-20T14:15Z → 2026-10-21T09:15Z

  • Phenomenon of the catalogue: Rain; Visibility drops over the whole scene
  • Precedents: JUNGFRAU-79
  • Family: Visibility
  • Measured: as the old label; rain at Nyon 15:00 and 17:00 → 21/10 10:00
  • Images: 21/10 00:00 (grey veil), 21/10 09:00 (scene damp, duller), 20/10 15:00 (clear before)
  • Confidence: high on the cause; medium on the start (the error_values rise from 14:30, before the rain of 17:00, with 0.1 mm at 15:00)

2. What I left without a label

  • 17/10 16:00-21:15: REF_G error_value up to 0.30 (level 0.085), with a shower at Nyon at 17:00 (1 mm). Effect on the corr at most 4-5 noises on one or two targets at a time (PBJ_C corr_x 17:45-18:15, BL_C), against 6-8 noises for the labelled reference nuisances (JUNGFRAU-39, 57): left under JUNGFRAU-56. A candidate if smaller ones are wanted.
  • 14/10 14:30: one-measurement base wobble (REF_CT corr_x -0.03, PBJ_C and PM_C +6 noises against the rolling level), REF_D 0.209 and REF_G 0.110 barely above their afternoon level, sunlight patches on the wall behind REF_CT on the thumbnail, in the strongest thermal afternoon. Left under JUNGFRAU-43; borderline.
  • 14/10 21:30: REF_D one-measurement error_value spike (0.16), no effect on any corr.
  • 12/10 12:15-13:15, 14/10 09:45-15:45, 15/10 10:30-15:00, 16/10 11:00-15:30: daytime rises of the REF_D / REF_G error_value (up to 0.19), no corr jump at 4 noises.
  • 18/10 day: no thermal phase: the housing temperature stays at 12.4-14 °C, and the drop of the tilt and of the corr_y comes with the rain (JUNGFRAU-58), which distorts the pose; the corr_y drop 08:00 → 10:00 (0.206 → 0.149) is small and ends where JUNGFRAU-58 begins.
  • Single instants flagged by the rolling ranges at 10/10 18:00, 12/10 18:45, 15/10 22:45, 16/10 10:30-11:00: BJ jitter, noise, or the one-sided window at the edge of a gap (no jump on the 15-min table).

3. The documentation

  • Helpful: families.md, boundary "Thermal displacement or Visibility" (decided JUNGFRAU-61); labeling-rules 2 on touching structure phases (JUNGFRAU-35) and on reading the end on the curves, not on the rejections (JUNGFRAU-47); rule 5 on stating the image of a No data (JUNGFRAU-59); the instructions on the weather (JUNGFRAU-58, 64).
  • A finding for the knowledge: on JUNGFRAU, the separation of BJ_D and BJ_G in corr_x (and the corr_y of BJ_C, BJ_D, BJ_G) follows the error_value of these three targets measurement by measurement: per-day correlation of (BJ_D - BJ_G) with their mean error_value 0.86-0.99 every day except 15/10 (0.62) and 20-21/10 (0.74); step to step r 0.63, slope about 7 px per unit of error_value (scripts/bj_spread.py, 970 steps). Their error_value is low by day and high at night, so their separation is larger every night. The existing thermal labels describe "BJ_D and BJ_G move apart on cool nights" as thermal (old JUNGFRAU-61, 63, 65); at least part of it is a measurement effect of these three large saturated spots nearest the camera, not a movement. Not settled: whether a thermal part remains (compare at equal error_value over several nights). stations.md could say that BJ_C, BJ_D, BJ_G are the large spots on the beams close to the camera.
  • The precedent JUNGFRAU-79 (rain, wet road) and the weather contradicted the "mist probable" of JUNGFRAU-58 and 64.
  • Missing: no tool plots the weather (read from the CSV).

Difficulties

  • summary.py loses its header when piped through tail: the column order was recovered from the --sensors order.

Sources read

  • agents/agents/analyst.md, instructions/analyst.md, the request, work/existing.jsonl, work/spanning.jsonl
  • tools/GUIDE.md, tools/DATA.md, labeling/families.md, labeling/labeling-rules.md, labeling/reading-rules.md, labeling/stations.md
  • labeling/auto/JUNGFRAU.jsonl (labels at the edges: JUNGFRAU-30 to 32, 66 to 69)
  • precedents.py (2 queries): NORDEND-150, LENZSPITZE-16, JUNGFRAU-79, NORDEND-213, NORDEND-173, WILDSTRUBEL-120, FLORIDA-76
  • data/JUNGFRAU: inventory, overview (whole range; rigid on 17/10), ranges (references error_value; rolling corr with instants), days --bounds, metrics, summary tables, weather CSVs (Nyon / Changins, 2.0 km), images 16/10 20:00, 18/10 14:00, 18/10 21:00, 19/10 12:00, 20/10 00:00, 20/10 15:00, 21/10 00:00, 21/10 09:00, thumbnails (references 14/10 and 12/10; BJ targets)
  • scripts/bj_spread.py, scripts/bj_table.py, scripts/build_labels.py (this folder)