Files
aimsight-labeling/auto/analyses/WILDSTRUBEL_2026-09-13T0000_2026-09-24T0230/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

67 lines
12 KiB
Markdown
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
# Labels: WILDSTRUBEL, 2026-09-13T00:00Z → 2026-09-24T02:30Z (UTC)
Task: verify the 10 existing labels of the range (`work/existing.jsonl`, WILDSTRUBEL-71 to -80) under the rules of 2026-11-17, and check the 3 spanning labels (WILDSTRUBEL-46, -69, -70) inside the range.
Context. From 13/09 00:00 to 22/09 08:45 the station delivers nothing: this is the spanning No data WILDSTRUBEL-69 (09/09 21:45 → 22/09 08:45). The measured part of the range is **22/09 08:45 → 24/09 02:30**, 168 instants at 15 min, no missing step, a pose at every instant (`inventory.py`). No weather: `data/WILDSTRUBEL/weather/` does not exist (station without a Swiss position). No image quality metrics (export columns peak to snr empty), so no `quality.py`. Housing temperature (`metrics.py`): about 35 °C at night, 46–48 °C by day both days.
Noise (own, window 22/09 08:45 → 24/09 02:30, `ranges.py`): corr_x 0.015–0.021 px for the targets, corr_y 0.012–0.022 px. mm/px (`targets.csv`): PB1_2 5.46 (13.8 m), PB2_2 6.57 (16.5 m), PB_5 9.15 (22.9 m), PB4 9.35 (23.4 m), PA2_2 9.67 (24.5 m), PA3_2 11.74 (29.5 m), PA_9 15.06 (39.1 m). References: PA1_2, PA_8, PB3_2, R1, R2, R3 (no mm).
## Verdict per old label
| Old id | Verdict | Evidence |
|---|---|---|
| WILDSTRUBEL-71 (22/09 08:45 → 12:30, Thermal) | **corrected** (description) | The rule now requires the corr to keep the thermal movement. They do: corr_y of PB2_2 is +7 to +8 noises above its night level 09:00–09:45 (−0.154 to −0.140 px against about −0.245), PB1_2 +5.6 at 09:00, falling with the tilt (0.13° → 0.073° at 11:30). The old text said "the corrected curves do not show this wave": wrong. Bounds kept (the pose returns about 12:30). |
| WILDSTRUBEL-72 (22/09 08:45 → 13:00, Visibility, PB1_2) | **corrected** (description) | The curve holds: PB1_2 corr_x −0.16 px at 09:00 (−12 noises against the window level, −13 against its night level of about +0.05), back by about 12:00–13:00. Rigid fit (`overview.py --rigid 2026-09-22T22:00Z --follow PB1_2`): PB1_2 dx −0.17 px at 09:00, −0.11 to −0.14 until 11:45, −0.04 to −0.08 until 14:30, 0 at 14:45; it moves **alone** in the image. Thumbnails of 22/09 09:00 and 11:00: its spot sits on the sunlit kerb band; at 14:45 the kerb is in shade. The old claim "its error_value rises much more than that of PB2_2 and PB4" is false: PB1_2 goes from 0.13 at night to 0.30 by day, PB2_2 from 0.10 to 0.21, PB4 from 0.20 to 0.40, all about ×2. The old text also cited 28/09 thumbnails, outside the range. Bounds kept: with the night level as reference, the last points beyond 5 noises fall between 11:45 and 14:00 depending on the level chosen (window level: 11:45; night level: 12:45 −5.1, 14:00 −5.0); 13:00 sits inside that uncertainty. |
| WILDSTRUBEL-73 (22/09 08:45 → 15:15, Visibility, day) | **corrected** (description) | The day agitation is real and beyond the threshold: isolated points beyond 5 noises by day (PA3_2 cx +5.5 at 11:30, PA1_2 and PB3_2 references), none at night; error_value about ×2 on all sensors 08:45 → 15:00, and 12–13 sensors above median+5 MAD (common panel) until 15:00, 0 from 15:15. Bounds kept. Changes: the restart jump at 08:45 now has its own label (rule "what falls under a long label is labeled apart"); "the compensation follows" is replaced, since the remainder on PB2_2 now goes in the thermal label. |
| WILDSTRUBEL-74 (22/09 17:00 → 17:30, Obstruction) | **kept** | R1 raw jumps (+21 px x, +17 px y) at 17:15, inlier 0; PB4 corr_x +0.043 against −0.057 to −0.085 around (+6 noises against its local level, +5.0 against the window level), PB_5 +4, PB1_2 cy +4.7. Back at 17:30. The description is right. |
| WILDSTRUBEL-75 (23/09 04:00 → 12:15, Thermal) | **corrected** (description) | Same as 71: corr_y of PB1_2 −0.075 to −0.11 px 05:30–07:45 against −0.188 at night (+5 to +8 noises, about 0.5–0.6 mm), PB2_2 −0.155 to −0.19 against −0.25 (+5 to +7 noises), back about 10:45–11:00 with the tilt (peak 0.173° at 06:15, trough 0.077° about 11:30). The old text said the corrected curves do not show the wave: wrong. Also "the tilt rises, with its scale and its rotation": the rotation dips first (−0.077° at 06:00) then rises above its level (−0.021° at 10:30); the pan dips then rises. Bounds kept. |
| WILDSTRUBEL-76 (23/09 04:00 → 15:15, Visibility, day) | **corrected** (description) | Agitation by day beyond 5 noises, none at night (PA_9 cx 05:30, 07:15; PA3_2 10:15, 13:30; PB1_2 cx 10:45 −5.8; PB2_2 cy 10:30; PB_5 cy +6.0 at 12:30). error_value high 04:15 → 15:00. Bounds kept. Changes: the sunrise reference loss (04:15–04:45) now has its own label; "PB4 rejected almost the whole first hour with R2" was wrong in time: PB4 is rejected 05:30–06:30, R2 05:45–06:00, an hour and a half after the start; the remark on the compensation is replaced. |
| WILDSTRUBEL-77 (23/09 08:30 → 09:00, Obstruction) | **kept** | R1 rejected, R2 inlier 0 (raw −23 px x, −26 px y); 08:45 thumbnails: a large bright reflection next to R1, a bright spot near R2. PB4 cx +12.7 noises (+0.23 px, 2.2 mm), PB_5 +9.3, PB3_2 +13, PB1_2 cy +11, PB2_2 cy +11, PA3_2 +6.7. Back at 09:00. The description is right. |
| WILDSTRUBEL-78 (23/09 09:15 → 09:45, Obstruction) | **kept** | R1 rejected at 09:30 (raw −34 px x); PB4 cx +0.074 against about −0.04 (+6 to +7 noises, 1.1 mm). Back at 09:45. |
| WILDSTRUBEL-79 (23/09 13:15 → 13:45, Obstruction) | **corrected** (description) | 13:30 image: a translucent vehicle crossing the road, seen in several overlaid positions. PB4 and R2 rejected; under the new rule a missing point does not justify a label, what counts is the effect: without R2 the basis changes and PA3_2 cx (+5.3 noises against the window level), PA2_2 cy (+4.2), PB1_2 cy (+3 to +5 against its local level) move together at this instant only. That is just beyond the threshold. The old description said nothing of an effect on the curves; it now states it. Bounds kept. |
| WILDSTRUBEL-80 (23/09 13:45 → 14:15, Obstruction) | **removed** | At 14:00, PA1_2, PA_8 (references), PA2_2, PA_9, PB2_2 and PB_5 are rejected; the targets still measured (PA3_2, PB1_2, PB4) do not move: PA3_2 cx −1 noise, cy +1.7; PB1_2 cx −2.4, cy 0; PB4 cx +1.3, cy 0, against their local level. The rejected points are only missing points, filtered by the station: not labelled under the new rule. The description was also wrong: PB1_2 is not rejected at 14:00 (`inventory.py`). |
## 1. My labels
### 2026-09-22T08:45Z → 2026-09-22T09:00Z (added)
- **Phenomenon of the catalogue**: "A reference dropped: the base changes", here three at once; the family is that of the cause, an image darker over the whole scene.
- **Precedents**: the search (`precedents.py`) did not answer in time (the embedding container was starting); the bounding follows WILDSTRUBEL-77 and -74 (one measurement, base change, back at the next).
- **Family**: Visibility: the whole image is darker (backgrounds of the thumbnails 92/55/32/55/52 against 125/90/72/87/87 at 08:39 and 120/86/25/83/79 at 09:00) and error_value is high on most sensors (PA3_2 0.75, PB4 0.74, R1 0.70, R2 0.86, R3 0.90). No object in front: not Obstruction.
- **Measured**: first measurement after the outage; R1, R2, R3, PB4, PA3_2 rejected; 3 references retained. PB_5 cx +0.206 px (+14 noises, 1.9 mm), PB3_2 (ref.) cx +13, cy +12; PB2_2 cx +6.6, cy +11 (0.95 mm); PB1_2 cy +12 (0.94 mm); PA2_2, PA_9 cx +5. At 09:00 PB_5, PB2_2 cy and PB3_2 are back (PB2_2 cy stays raised but that is the thermal remainder of WILDSTRUBEL-71; PB1_2 cx goes down, its own label WILDSTRUBEL-72).
- **Images**: 22/09 08:39, 08:45, 09:00 (whole image and fixed thumbnails of R1, R2, PB4, PA3_2, PB_5): 08:45 darker, nothing in front of the references, a moving vehicle blurred far on the road.
- **Confidence**: high for the event and its bounds; medium for the cause (an exposure not settled after the restart is deduced, not proven).
### 2026-09-23T04:00Z → 2026-09-23T05:00Z (added)
- **Phenomenon of the catalogue**: "The sun in the field or behind a spot" (Visibility), with "A reference dropped: the base changes".
- **Precedents**: none read (see above); WILDSTRUBEL-76 mentioned it inside the day label.
- **Family**: Visibility: the sun in the field, seen.
- **Measured**: PA_8 (ref.) rejected 04:15–04:30 (error_value 0.77, 0.69), PA_9 rejected 04:30–04:45; error_value jumps on all sensors 04:15–04:45 and falls back at 05:00. PA_9 cx +0.19 px (+9 noises, 2.8 mm, the farthest target at 39 m), PA2_2 cx +7.5 (1.2 mm), cy +5.3; PA1_2 (ref.) +10; PB1_2 cx +4.5 to +5 against its night level. Values back at 05:00 for all (PA3_2 cy −4.8 at 04:45, back at 05:00).
- **Images**: 23/09 04:15 whole image: the sun just above the buildings, flare across the image, the scene underexposed.
- **Confidence**: high.
### The corrected and kept labels
See the table above; the full texts are in `labels.jsonl`. For WILDSTRUBEL-71 and -75 the family stays Thermal displacement: the corr_y remainder of PB1_2 and PB2_2 rises and comes back with the tilt on both days, smooth, error_value without a step at that moment, the housing temperature at 46–48 °C (rule "deviation specific to targets that follows the camera on the same dates is the same event"). WILDSTRUBEL-72 stays Visibility (boundary "the lit background behind a spot", same family as the sun in the field).
## 2. What I left without a label
- 22/09 21:45: R1 rejected alone; PB4 cx +2, PB_5 +1.7, PB2_2 +1.3 noises against their local level: under the threshold.
- 23/09 05:30–06:30: PB4 rejected (missing points), R2 rejected 05:45–06:00; PB3_2 (ref.) cy +5 to +7 and PB1_2/PB2_2 cy +5 to +6 are the thermal remainder (WILDSTRUBEL-75); PA_9 cx +5.1 at 05:30 is part of the day agitation (WILDSTRUBEL-76). Not labelled apart.
- 23/09 14:00 (old WILDSTRUBEL-80): measured targets within ±2.4 noises; missing points only.
- Night 22–23/09 and night 23–24/09 (15:15 → 04:00, 15:15 → 02:30): no target beyond 5 noises on corr_x or corr_y (`ranges.py`, 34 off-level instants listed, all by day or at the labelled instants).
- 13/09 00:00 → 22/09 08:45: no measurement, covered by the spanning WILDSTRUBEL-69.
## 3. The documentation
- Helpful: `labeling-rules.md` §1 (thermal labelled only when the corr keep it), §2 (event at the resumption starts at that measurement), §1 "what falls under a long label is labeled apart" (which led to the two added labels); `families.md` "A gap" (missing points not labelled), which removes WILDSTRUBEL-80; `tools/GUIDE.md` `overview.py --rigid --follow`, which shows PB1_2 moving alone.
- Precedents: `precedents.py` did not answer within the call time (the Ollama container was starting); the reviewed labels of the range itself (74, 77, 78) served as the model for the bounding of one-measurement base changes.
- The window level vs the night level: `ranges.py` measures against the window median, which on a short window half of daytime is pulled toward the day; the thermal remainder of PB1_2 and PB2_2 only appears clearly against the night level (`summary.py` 15 min tables). The rules say "from the night level" for the thermal: worth stating in the guide that `ranges.py --outside <day>` or `days.py` is needed for it.
- The old labels written under the earlier rules described thermal waves as fully compensated without measuring against the night level: contradicted by the data here.
- Spanning labels, checked within the range: WILDSTRUBEL-46 (R3 Target issue) holds: R3 rejected 22/09 08:45–15:00 and 23/09 04:15–15:00 almost continuously, accepted at night. WILDSTRUBEL-69 (No data) holds: nothing from 13/09 to 22/09 08:45, first measurement at 08:45. WILDSTRUBEL-70 (Displacement during the outage, end 22/09 09:00) holds within the range; its text says PB1_2's raw position shifted alone, consistent with what the rigid fit shows on the first day.
## Difficulties
- `precedents.py` did not return within 2 min (container start); run in the background.
- No weather folder for this station.
## Sources read
`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`, `request.json`; tools: `inventory.py`, `ranges.py --instants 2`, `overview.py` (residuals, `--rigid 2026-09-22T22:00Z`, `--follow PB1_2,PB2_2,PB4`), `summary.py` (corr_x, corr_y, error_value at 15 min), `metrics.py`, `thumbnails.py` (PB1_2, PB2_2, PB4; R1, R2, PB4, PA3_2, PB_5 `--fixed`), `plot.py`; images 22/09 08:45, 23/09 04:15, 23/09 13:30.