Files
aimsight-labeling/auto/analyses/JUNGFRAU_2026-10-21T1030_2026-11-01T0000/analysis.md
T

156 lines
17 KiB
Markdown

# Labels: JUNGFRAU, 2026-10-21T10:30Z → 2026-11-01T00:00Z (UTC)
Request: verify the 34 existing labels of the range (`work/existing.jsonl`): keep, correct, remove, add. DETECTORS: no.
Result: 36 labels in `labels.jsonl`: 27 kept, 7 corrected, 0 removed, 2 added. The spanning label JUNGFRAU-2 holds within the range.
Weather: MeteoSwiss station Nyon / Changins (CGI), 2.0 km away, height 458 m, hourly (`data/JUNGFRAU/weather/`). Rain on 21/10 (to 10:00, then 13:00-18:00 and 20:00-02:00), 23/10 07:00-12:00 (up to 2.1 mm/h at 09:00), 15:00-16:00, 18:00-20:00, 22:00 → 24/10 01:00, 24/10 03:00-13:00 (light, 0.2-1 mm/h), 24/10 22:00 → 25/10 17:00 (up to 2.4 mm/h); dry from 25/10 18:00 to the end of the range. Sun (60 min/h) on 22/10 and 27/10 from 09:00 to 15:00, 29/10 15:00, 31/10 15:00. Nights of 26, 27 and 28/10: dry, humidity 85-95 %, wind below 10 km/h.
The station: a bridge, camera under the deck fixed to it, three references off the bridge in the open (REF_G on a tripod on the left slope, REF_D at the right fence, REF_CT at the abutment wall: located from their `x`, `y` on the 22/10 10:00 image). No image quality metrics (export columns peak to snr empty). Images every 15 min until 25/10 00:00, then hourly.
## Verdict per old id
| id | verdict | evidence |
|---|---|---|
| JUNGFRAU-66 | kept | median corr_y of the targets down from 10:30 to its minimum at 16:30 (-0.116 px), housing temperature 13 → 15.6 °C at 11:30; the jump after 16:30 is REF_G (label 69) |
| JUNGFRAU-67 | **corrected** (description) | episode confirmed (11 sensors jumping together at 13:00 and 14:00, REF_G error_value 0.08 → 0.32 at 14:00, targets 0.19 → 0.25); rain at Nyon 13:00-15:00 (1.4 mm at 14:00): "cause not established" → rain likely |
| JUNGFRAU-68 | kept | corr_y up from 16:30, peak +0.20 at 22/10 07:30, housing temperature 12.8 → 10.1 °C |
| JUNGFRAU-69 | **corrected** (description) | common jumps 17:00-19:15 (up to 11 sensors), REF_G error_value up to 0.27; rain at Nyon 17:00-18:00 (1.8, 1.2 mm): rain likely |
| JUNGFRAU-70 | **corrected** (description) | REF_G 0.56, REF_D 0.27, back at 02:30-02:45; rain at Nyon 00:00-02:00 (1.3, 0.7, 0.1 mm), none from 03:00: rain likely |
| JUNGFRAU-71 | kept | corr_y peak 07:30, down to -0.25 at 17:15-17:30, corr_x up to +0.146, scale max at 14:00, housing temperature up to 25.9 °C at 14:30, sunny day |
| JUNGFRAU-72 | kept | 15:30: 7 sensors in corr_x, REF_G 0.25; sun 60 min at 15:00 |
| JUNGFRAU-73 | kept | corr_y -0.25 → +0.08 at 23/10 05:30, tilt back to 0.030 |
| JUNGFRAU-74 | kept | corr_y down to -0.047 at 15:30, tilt down from 07:30 |
| JUNGFRAU-75 | **corrected** (bounds 07:30-10:30 → 07:15-12:15, description) | first common jump at 07:30 (REF_CT corr_x 0.128 → 0.180; 12 sensors at 07:45): start 07:15. Error_value of the targets high until 11:30 (BL_G 0.65 at 10:45); BJ_G corr_x -1.15 at 10:30, -0.99 at 11:00, -0.85 at 11:15, -0.59 at 11:30; BJ_D corr_y 0.43 (11:15) → 0.67 → 0.75 → 0.91 (12:15, back): end 12:15. "In clear weather" is contradicted: rain at Nyon 07:00-12:00 (2.1 mm at 09:00), wet road and no shadow on the 09:00 image. "Spots larger and blurrier than the day before" is an exposure effect: they are as large at 13:00, after the episode (`work/th75`) |
| JUNGFRAU-76 | kept | corr_y up until 24/10 02:30, flat to 07:00 (PBJ_C peak at 07:00, within two steps of 06:30) |
| JUNGFRAU-77 | **corrected** (start 22:30 → 21:45, description) | first common jump at 22:00 (REF_CT corr_x 0.145 → 0.168, BJ_G -0.58 → -0.45, BL_G 0.19 → 0.23; 5 sensors), back at 22:15; error_value of the targets rising from 21:15 (BL_G 0.47 → 0.62 at 23:15). "Images not looked at": now looked at, 22:45 thumbnail, REF_G spot a little smaller than at 21:00 and 00:00 (`work/th77`). Rain at Nyon 22:00-23:00 (1.2, 1.4 mm) |
| JUNGFRAU-78 | **corrected** (description) | episode confirmed (common jumps at 04:00, 06:15-06:30, 08:00-08:15, 09:00-10:45, REF_G up to 0.32); rain at Nyon every hour 03:00-13:00: rain likely. Bounds kept: BJ_G and BJ_D stay apart in corr_x until about 15:00, the edge-target pattern of every cool night, not a common jump |
| JUNGFRAU-79 | kept | 25/10: 11 to 14 sensors out of level 07:45-16:45, rejections of PM_C, PM_D, BL_G, BL_C, PBL_C; rain at Nyon all day (up to 2.4 mm/h), stops at 17:00 with the return 16:45-17:15 |
| JUNGFRAU-80 | kept | as described; the thermal part cannot be separated from the rain |
| JUNGFRAU-81 | **corrected** (end 09:00 → 10:00) | median corr_y still rising after 09:00: +0.155 at 09:00, +0.196 at 09:45, plateau +0.19 to 10:30, then down; JUNGFRAU-83 starts at 10:00: the phases touch (cut in the flat); 09:00-10:00 was left without label while the curves still moved |
| JUNGFRAU-82 | kept | REF_G error_value 0.07 → 0.21 at 00:15, rejected 02:45-06:00, reprojection at zero; last common jumps at 07:15; dry night |
| JUNGFRAU-83 | kept | corr_y down from 10:30 to the minimum at 19:00 (-0.087) |
| JUNGFRAU-84 | kept | no measurement at 14:30 and 14:45; 14:45 missing on CLARIDEN, FLORIDA, LENZSPITZE, NORDEND too, 14:30 on NORDEND (`stations.py`); no image between 14:00 and 15:15 |
| JUNGFRAU-85 | kept | corr_y almost flat, BJ_G down and BJ_D up in corr_x, tilt up |
| JUNGFRAU-86 | kept | REF_G 0.27 at 22:45, 0.47 at 23:00, REF_D up from 23:00; a corr_y hop of a few hundredths at 22:45 only, within one step; no pose 02:00-06:45; last common jumps 07:30 |
| JUNGFRAU-87 | kept | corr_y down to -0.375 at 16:15, corr_x up to +0.27, housing temperature 22 °C, sunny |
| JUNGFRAU-88 | kept | 09:00 missing; same minute on LENZSPITZE, NORDEND, 09:15 on CLARIDEN, FLORIDA; no image between 08:00 and 09:15 |
| JUNGFRAU-89 | kept | 15:15: 5 sensors in corr_x, REF_G 0.24; sun 60 min at 15:00 |
| JUNGFRAU-90 | kept | corr_y -0.37 → -0.18 (comes back only partly: the housing temperature stays at 13-14 °C all night), BJ_G/BJ_D apart |
| JUNGFRAU-91 | kept | REF_G 0.07 → 0.12 at 01:45, 0.46 at 02:00, rejected 03:15-04:30; REF_D still raised to 06:00 |
| JUNGFRAU-92 | kept | BJ_G and BJ_D come back in steps 08:00-10:45, corr_y down to the flat trough 16:15-18:00 |
| JUNGFRAU-93 | kept | corr_y -0.21 → 0.00 at 29/10 06:15 |
| JUNGFRAU-94 | kept | corr_y minimum at 16:15 (-0.161), tilt minimum at 13:30 |
| JUNGFRAU-95 | kept | corr_y back to 0 at 21:15 then slowly to +0.029 |
| JUNGFRAU-96 | kept | corr_y within 5 noises, but tilt -6.7 noises at 13:30: a cycle to label |
| JUNGFRAU-97 | kept | corr_y up from 19:15, +0.116 at 31/10 06:15 |
| JUNGFRAU-98 | kept | corr_x +0.116 and corr_y -0.116 at 16:15, pan -21 noises at 15:30 |
| JUNGFRAU-99 | kept | corr_y up to +0.14 at 22:15 |
| JUNGFRAU-2 (spanning, not changed) | holds | PM_G rejected at all 1012 measurements of the range, error_value 0.42 against its threshold 0.3, its neighbours PM_C 0.49, PM_D 0.37 |
## 1. My labels
Only the corrected and the added labels are detailed; the kept ones are justified in the table above.
### 2026-10-21T12:45Z → 2026-10-21T14:45Z (JUNGFRAU-67, corrected)
- **Phenomenon of the catalogue**: Rain (the references, in the open, suffer it; the targets under the deck less)
- **Precedents**: NORDEND-182 (validated: reference error_value up, corrections jumping together, calm when it falls back), for the signature
- **Family**: Visibility: a diffuse nuisance on the reference basis, nothing in front
- **Measured**: `ranges.py --rolling 4 --instants 3`: 11 sensors at 13:00 and 14:00, 8 at 13:15, 6 at 13:45 and 14:15; REF_G error_value 0.08 → 0.20 (13:45) → 0.32 (14:00); targets 0.19 → 0.25. Rain at Nyon 0.1 mm at 13:00, 1.4 at 14:00, 0.8 at 15:00
- **Images**: 12:00 and 14:00 thumbnails (old label)
- **Confidence**: medium: rain measured a few km away, not seen at the camera
### 2026-10-21T16:45Z → 2026-10-21T19:45Z (JUNGFRAU-69, corrected)
- **Phenomenon of the catalogue**: Rain
- **Precedents**: as above
- **Family**: Visibility
- **Measured**: common jumps 17:00 to 19:15 (11 sensors at 17:45 and 18:30), REF_G error_value up to 0.27; rain 1.8 mm at 17:00, 1.2 at 18:00
- **Images**: 18:00 thumbnail (old label)
- **Confidence**: medium
### 2026-10-22T00:15Z → 2026-10-22T02:45Z (JUNGFRAU-70, corrected)
- **Phenomenon of the catalogue**: Rain; "A reference loses its brightness, without an obstacle" fits the image, the weather gives the likely cause
- **Precedents**: none useful beyond the old label
- **Family**: Visibility
- **Measured**: REF_G error_value 0.56, REF_D 0.27 from 01:45 to 02:15, back at 02:30-02:45; rain 1.3, 0.7, 0.1 mm at 00, 01, 02 h, none from 03:00
- **Images**: 01:30 and 02:15 thumbnails (old label)
- **Confidence**: medium
### 2026-10-23T07:15Z → 2026-10-23T12:15Z (JUNGFRAU-75, corrected)
- **Phenomenon of the catalogue**: Rain
- **Precedents**: NORDEND-173, NORDEND-213 (validated, days of rain: error_value up over the whole scene, rejections, corrections jumping)
- **Family**: Visibility: the whole scene, wet road, no object
- **Measured**: `scripts/q.py`: REF_CT corr_x 0.128 (07:15) → 0.180 (07:30) → 0.220 (07:45), 12 sensors out at 07:45 and 09:45; error_value of BL_G, PBL_C, BL_C, PM_C 0.45-0.60 → 0.60-0.72 from 07:30 to 11:00, back about 12:00; BJ_G corr_x -0.69 (07:15) → -1.15 (10:30) → -0.59 (11:30); BJ_D corr_x up to 0.72, corr_y 0.80 → 0.32 (10:30) → 0.91 (12:15); PM_C rejected at 08:45, 09:00, 09:30. Rain at Nyon 0.3, 1.1, 2.1, 1.7, 0.3, 0.2 mm from 07:00 to 12:00. Figure `rain_morning_2026-10-23.png`
- **Images**: 09:00 whole image (grey scene without shadow, wet road); thumbnails of REF_G, PM_C, BJ_G, BL_D at 22/10 09:00 and 23/10 09:00, 11:00, 13:00 (`work/th75`): nothing in front; the spots are larger all through the overcast day, 13:00 included, than on sunny 22/10: an exposure effect, not a proof
- **Confidence**: high for the bounds, medium-high for the cause (image and weather agree)
### 2026-10-23T21:45Z → 2026-10-23T23:45Z (JUNGFRAU-77, corrected)
- **Phenomenon of the catalogue**: Rain
- **Precedents**: as for 67
- **Family**: Visibility
- **Measured**: 22:00: REF_CT corr_x 0.145 → 0.168, BJ_G -0.58 → -0.45, BL_G 0.19 → 0.23, back at 22:15; jumps again at 22:45, 23:00, 23:30; back at 23:45 (REF_CT 0.140). Error_value of the targets up from 21:15 (BL_G 0.47 → 0.62 at 23:15); REF_G 0.25 at 22:45. Rain 1.2 mm at 22:00, 1.4 at 23:00
- **Images**: thumbnails of REF_G, REF_CT, BL_D at 21:00, 22:45, 23:00, 00:00 (`work/th77`): REF_G a little smaller at 22:45, nothing in front
- **Confidence**: medium
### 2026-10-24T03:00Z → 2026-10-24T12:45Z (JUNGFRAU-78, corrected)
- **Phenomenon of the catalogue**: Rain
- **Precedents**: as for 75
- **Family**: Visibility
- **Measured**: common jumps at 04:00, 06:15, 06:30, 08:00 (10 sensors), 08:15, 09:00, 09:45, 10:45; REF_G up to 0.32; rain 0.2 to 1.0 mm every hour from 03:00 to 13:00
- **Images**: 08:00 thumbnail (old label)
- **Confidence**: medium
### 2026-10-24T06:30Z → 2026-10-24T17:00Z (added)
- **Phenomenon of the catalogue**: The structure expands and contracts with the temperature
- **Precedents**: JUNGFRAU-25, JUNGFRAU-60, JUNGFRAU-63 (validated: weak warming of overcast days labelled on this station)
- **Family**: Thermal displacement: slow and smooth, each target its amplitude, back in the evening
- **Measured**: `summary.py --step 1h`: PBJ_C corr_y 1.52 (07:00) → 1.17 (16:00-17:00), PM_D 0.826 (06:00) → 0.740 (17:00), BL_D 0.211 → 0.129; housing temperature 9.5 °C (05:00) → 12.5 (12:00), then 11.4-11.9; pan -0.0532 → -0.0489 (15:00). `days.py --bounds`: median corr_y beyond 5 noises from 11:15, minimum -14.5 noises at 18:15 against the night: a cycle to label (`labeling-rules.md`, section 1). No thermal label covered 24/10 06:30 → 25/10 07:30: the only day of the range without one. Figure `thermal_2026-10-24.png`
- **Images**: none (deduced)
- **Confidence**: medium: weak cycle, the visibility episode of the morning superimposed
### 2026-10-24T17:00Z → 2026-10-25T00:15Z (added)
- **Phenomenon of the catalogue**: The structure expands and contracts with the temperature
- **Precedents**: as above
- **Family**: Thermal displacement
- **Measured**: PBJ_C corr_y 1.16 (17:00) → 1.49 (00:15), PM_D 0.735 → 0.813, BL_D 0.119 → 0.189, PBL_D -0.101 → -0.024; housing temperature 12.4 (19:00) → 10.0 °C (00:00, the minimum of the night). After 00:15 the curves are dominated by the rain of the night (JUNGFRAU-79: 12 sensors jumping at 00:15, 03:00), PBJ_C goes back down to 1.30 by 07:00; that part is not labelled thermal (possibly the start of the warming of 25/10, JUNGFRAU-80 starts at 07:30; not separable)
- **Images**: none
- **Confidence**: medium
### 2026-10-25T16:45Z → 2026-10-26T10:00Z (JUNGFRAU-81, corrected)
- **Phenomenon of the catalogue**: The structure expands and contracts with the temperature
- **Precedents**: none needed
- **Family**: Thermal displacement
- **Measured**: `scripts/turn.py` (median of the targets, smoothed over 5 measurements): corr_y +0.144 (08:30), +0.155 (09:00-09:30), +0.196 (09:45), +0.19 to 10:30, +0.162 at 11:00, down after; tilt 0.0276 (07:00) → 0.0206 (10:00); housing temperature 8.1 °C (05:30) → 13.8 (10:00). The turning point of the structure is at 10:00-10:30, where JUNGFRAU-83 starts
- **Images**: none
- **Confidence**: high
## 2. What I left without a label
- 26/10 11:00: no camera registration and no reference with an inlier, but the corrections are continuous with the measurements before and after: nothing on the curves.
- 21/10 21:45 and 23:45, 22/10 03:15 and 19:45: three or four sensors at 5 to 11 noises for one measurement, PM_C or the BJ group, no reference raised: the noise of these targets.
- 26/10 17:45, 19:15, 20:00 and 27/10 07:45-08:30: BJ_C, BJ_D, BJ_G alone (the noisiest targets, at the edges of the image), under the thermal labels.
- 28/10 05:00-10:45: error_value of BL_C, BL_G, PBL_C close to their threshold (night regime, falling with the daylight), PBL_C corr_x jumps of about a tenth of a pixel 08:15-09:45: within JUNGFRAU-91 and JUNGFRAU-92.
- 24/10 12:45-15:00: the tail of JUNGFRAU-78 (BJ_D corr_x back from 0.55 to 0.2, error_value of the targets declining), left in the thermal label of the day, no common jump.
- Rain hours without disturbance: 23/10 15:00-16:00 and 18:00-20:00 (0.3-0.7 mm/h): REF_G at its level, no common jump. Rain does not always disturb REF_G, hence "likely" and not "seen".
- From 28/10 the pose, sigma0 and the references are calmer than before (sigma0 of the pose about 0.0042 → 0.0035): dry weather, no step in the corrections, not a change of configuration.
## 3. The documentation
- `families.md`, "Thermal displacement or Visibility" and the list of phenomena: decisive to keep the opposite movements of BJ_G and BJ_D at night in the thermal labels, and the REF_G episodes in Visibility.
- `labeling-rules.md`, section 2 ("the phases of the structure are cut at the turning points ... the two phases touch") gave JUNGFRAU-81; section 1 ("no thermal label on a day when neither the pose nor the corr deviate by more than 5 noises") showed 24/10 needed one; section 2 ("start is the measurement just before the first touched point") gave the bounds of 75 and 77; section 6 (seen / deduced / not established) and the weather instruction gave the "rain likely" of five labels.
- `labeling-rules.md`, section 5 ("bigger spots ... only after comparison with a calm day of the same image regime"): confirmed on 75, whose "larger and blurrier spots than the day before" compared a sunny day with an overcast one.
- `reading-rules.md`, section 6 ("references that do not move in the rain do not contradict the rain") is the reverse case here: at JUNGFRAU the references are in the open and the targets under the deck, so rain shows first on the references, REF_G above all. Worth a line in `stations.md`.
- The old labels were written before the weather was read: every Visibility episode of the first week with "cause not established" (67, 69, 70, 75, 77, 78) falls in rain hours at Nyon; the nights of reference losses (82, 86, 91) are dry and calm: their cause stays not established (condensation possible, not seen).
- No quality metrics on this camera: occultation and veil rest on the thumbnails.
- `precedents.py --exclude JUNGFRAU <start> <end>` still returns the other JUNGFRAU labels (outside the range): useful here.
## Difficulties
- The measurement times of the export carry seconds (`00:00:58`): a script matching them to the camera poses must round to the minute.
- No tool gives a per-instant table of one column for all sensors together with the pose; I wrote `scripts/q.py` (and `scripts/turn.py` for the smoothed median of the targets with the housing temperature).
## Sources read
- `agents/agents/analyst.md`, `instructions/analyst.md`, `work/prompt.txt`, `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` (in full)
- `labeling/auto/JUNGFRAU.jsonl` (labels around the range: JUNGFRAU-58 to -99, and -2)
- Precedents: NORDEND-213, NORDEND-182, NORDEND-173, NORDEND-65, JUNGFRAU-25, JUNGFRAU-60, JUNGFRAU-63
- Data: `inventory.py` (window), `overview.py` (window, and `--rigid 2026-10-21T11:00Z`), `days.py --bounds --smoothing 5`, `ranges.py --rolling 4 --instants 3`, `metrics.py`, `stations.py` (CLARIDEN, FLORIDA, LENZSPITZE, NORDEND, 26-27/10), `summary.py`, `plot.py` (21-22/10, 23/10 morning, 23-25/10, 28/10 morning), `thumbnails.py` (`work/th75`, `work/th77`), whole images 22/10 10:00 and 23/10 09:00, weather CSVs 21/10 to 01/11; scripts `scripts/q.py`, `scripts/turn.py`, `scripts/build.py`