Files
aimsight-labeling/auto/analyses/NIEDERHORN_2026-10-01T0000_2026-10-11T0500/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

283 lines
34 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: NIEDERHORN, 2026-10-01T00:00Z → 2026-10-11T05:00Z (UTC)
Request: verify the 42 existing labels of the range (`work/existing.jsonl`), keep / correct / remove, and add what they miss; the 10 spanning Target issue labels (`work/spanning.jsonl`) are checked, not rewritten. Result: 52 labels in `labels.jsonl` (29 kept unchanged, 13 corrected, 0 removed, 10 added), built by `scripts/build_labels.py`; `validate.py`: 0 errors (3 overlaps flagged, all between distinct events: plant / train).
Context: hourly cadence (60 min), 241 instants, 5 references (PWD_REF08 to 12, on the road above the tracks), 30 targets, of which e03, e04, e05, e07, e08, e11, e12 always rejected (spanning Target issue). No image-quality metrics (export empty). Camera pose present at every instant. Housing temperature (`sensor_*_deg_c`) present every hour. **No weather**: the station has no `weather/` folder (no Swiss position: the station is in North America, its night falls around 23:00-11:00 UTC); the weather is read on the images only.
Method: `inventory.py` (window), `overview.py --step 1h` (pose, references, residuals), `days.py --night 3-11 --bounds`, `stations.py` (gaps against 11 stations), metrics at the gap hours (my script reading `data.load_metrics`), `scripts/instants.py` (per instant: pose, temperature, rejected sensors outside the permanent ones, one-measurement spikes of `corr` against the mean of the two neighbours, in own noises), `scripts/dev.py` (deviation of each sensor from a reference window), `summary.py` per sensor on each moment, thumbnails and whole images, precedents.
## Verdict per old id
| Old id | Verdict | Evidence |
|---|---|---|
| NIEDERHORN-1 | kept | 01:00 missing for all sensors; voltage 14.1-14.4 V around, CPU temperature present at 01:00; no image at 01:00 either; the other stations measure then (only NIEDERHORN missing). |
| NIEDERHORN-12 | corrected (end 16:00 → 17:00, description) | At 16:00 PWD_IMP_B07_018_C09 (rejected 14:00-15:00) corr_y still clearly off (+0.13 px against its evening level), PWD_RMP_B17_026_C09 off in corr_x and corr_y, e09 in corr_x; 17:00 back. Rejected-target list rewritten from the inventory (PWD_IMP_B07_018_C09, the most rejected, was missing); 07:00 image shows faint out-of-focus drops and a noisy background. |
| NIEDERHORN-13 | corrected (description) | Train at 12:00 holds; but the under-threshold jumpers named were wrong: PWD_IMP_B07_016_C09 does not move at 12:00, PWD_RMP_B12_018_C09 (corr_y +15 noises) was missing. |
| NIEDERHORN-14 | kept | 13:00 missing for all; metrics present, voltage 14.0-14.4 V; no image at 13:00; NIEDERHORN alone. |
| NIEDERHORN-15 | kept | 21:00: e02, e06, e09, PWD_RMP_B17_026_C09 rejected (and e01, not named), nothing else moves. |
| NIEDERHORN-16 | corrected (start 02:00 → 01:00, end 14:00 → 15:00, description) | At 02:00 e14 corr_y drops (−0.27 px, −10 noises), its corr_x and PWD_IMP_B07_018_C09 corr_x drop, error_value of e14 and PWD_RMP_B16_026_C09 rises: the rain is already there; at 14:00 PWD_RMP_B15_020_C09 corr_x still −0.37 px with its error_value at three times its level, e09 error still high; 15:00 mostly back. 07:00 image: large drops on the glass and rain streaks under the floodlight. |
| NIEDERHORN-17 | corrected (start 07:00 → 04:00, description) | PWD_IMP_B07_016_C09 already jumps at 05:00 (corr_x +0.21, corr_y −0.30, error_value down to 0.11, the plant signature) and 06:00; 03:00-04:00 clean; 05:00 thumbnail: leaves over the spot. The single hit at 02:00 is a separate label (two clean measurements between). |
| NIEDERHORN-18 | kept | 02/10 12:00 missing for all; metrics present, voltage 14.0-14.2 V; no image at 12:00; NIEDERHORN alone. |
| NIEDERHORN-19 | kept | Pose bounds: pan, tilt first beyond 5 noises at 14:00, rotation 14:00-15:00, temperature 26.5 → 36.5 °C; curves at the level of the next night around 23:00. |
| NIEDERHORN-20 | kept | 21:00: e02, e06, e09, PWD_IMP_B07_016_C09 rejected (e01 in its daytime regime), PWD_RMP_B17_026_C09 jumps in corr_y. |
| NIEDERHORN-21 | kept | 03/10 07:00: PWD_IMP_B07_016_C09 alone rejected. |
| NIEDERHORN-22 | corrected (description) | At 10:00 e01 is rejected too (error_value 0.96), at night when e01 is normally measured; on the 10:00 thumbnails both spots (and a small one next to e01) are gone. "e02 alone" was false. |
| NIEDERHORN-23 | corrected (description) | Sun at 13:00 seen on the 03/10 13:00 image; but "e06 and e09 often rejected" is false (both measured at 13:00 on 03/10), and the image cited was that of 21.10. Rewritten with that day's references and targets. |
| NIEDERHORN-24 | kept | Pose 13:00-16:00 tilt up / rotation down, pan to 17:00, scale to 19:00, temperature 23.5 → 45 °C; bounds 12:00 → 23:00 consistent with `days.py --bounds`. |
| NIEDERHORN-25 | corrected (description) | Same as 23 for 04/10 (e06, e09 not rejected; image of 21.10). |
| NIEDERHORN-26 | kept | Tilt −0.068° at 19:00 back at 20:00, rotation, pan to 17:00; bounds consistent. |
| NIEDERHORN-27 | corrected (description) | 05/10 13:00: sun seen on the image, reference error_value up (PWD_REF10_C09 0.14 → 0.34), but the targets hardly move (all spikes ≤ 4 noises); "many targets jump" and "e06, e09 often rejected" were false. Kept as a label (references disturbed, small effect), described as small. |
| NIEDERHORN-28 | kept | Thermal 05/10 consistent. |
| NIEDERHORN-29 | kept | PWD_IMP_B07_016_C09 rejected 14, 15, 17, 21:00, back 22:00. |
| NIEDERHORN-30 | kept | 06/10 04:00 rejected alone (a small raised level of its corr_x until 11:00 stays under 3 noises). |
| NIEDERHORN-31 | corrected (description) | Train at 12:00 holds; PWD_RMP_B12_019_C09 (corr_y +30 noises, the largest jump) was missing from the under-threshold list. Its rejection at 11:00 is a separate added label. |
| NIEDERHORN-32 | kept | 06/10 13:00 sun, references PWD_REF08 to 10 up; image of that day cited. |
| NIEDERHORN-33 | kept | Thermal 06/10 consistent. |
| NIEDERHORN-34 | kept | 22:00 list matches the inventory; under-threshold PWD_RMP_B12_019_C09, PWD_RMP_B12_020_C09 (+39 to +43 noises), PWD_RMP_B12_018_C09, e13. |
| NIEDERHORN-35 | kept | PWD_IMP_B07_016_C09 rejected 22:00-23:00. |
| NIEDERHORN-36 | kept | 07/10 12:00 list matches; PWD_RMP_B15_019_C09 +39 noises under the threshold named. |
| NIEDERHORN-37 | corrected (end 18:00 → 23:00, description) | After the 17:00 rejection, the plant hits again: 19:00 corr_y −0.30 px against 18:00, 20:00 corr_y +0.37 with error_value 0.09 (its lowest), 22:00 corr_x +0.24, corr_y −0.45, error_value 0.11; 23:00 back. Thumbnails 19:00, 20:00, 22:00: leaves across the spot. Pause of one measurement only: same episode. |
| NIEDERHORN-38 | kept | Thermal 07/10 consistent (pose first beyond 5 noises at 14:00; the night stays warm). |
| NIEDERHORN-39 | kept | Rain night 08/10: error_value up from 01:00 to 04:00, back 05:00; bounds hold. The description omits that PWD_IMP_B07_018_C09 is rejected 02:00-04:00 and PWD_RMP_B12_019_C09 at 01:00, but what it says is true. |
| NIEDERHORN-40 | kept | 08/10 02:00 corr_x +0.50, 05:00 rejected, 06:00 corr_y −0.59; 07:00 back. |
| NIEDERHORN-41 | corrected (description) | Same as 23 for 08/10. |
| NIEDERHORN-42 | kept | Thermal 08/10 consistent. |
| NIEDERHORN-43 | kept | 08/10 17:00: PWD_RMP_B12_017_C09 and PWD_IMP_B07_016_C09 rejected. |
| NIEDERHORN-44 | kept | 20:00 missing for all, the 20:00 image exists, metrics present, NIEDERHORN alone. |
| NIEDERHORN-45 | corrected (description) | Same as 23 for 09/10. |
| NIEDERHORN-46 | kept | Thermal 09/10 consistent. |
| NIEDERHORN-47 | kept | 16:00: PWD_REF08_C09 rejected (error_value 0.97, raw jump), 4 references retained, all targets jump; 16:00 image: dark blurred mass at the top left. |
| NIEDERHORN-48 | kept | 20:00: e02, e06, e09, PWD_RMP_B17_026_C09 rejected (e01 in its regime), nothing else. |
| NIEDERHORN-49 | corrected (end 01:00 → 02:00, description) | At 01:00 its corr_y is still away (+0.24 against its night level) with error_value 0.08, its lowest of the range; leaves on the 00:00-02:00 thumbnails; 02:00 back within 3 noises. |
| NIEDERHORN-50 | kept | Thermal 10/10 consistent. |
| NIEDERHORN-51 | kept | 11/10 00:00 missing for all; metrics present, voltage 14.0-14.4 V; no image at 00:00; NIEDERHORN alone. |
| NIEDERHORN-52 | kept | 03:00 train; list true (e01 and PWD_IMP_B08_024_C09, also rejected, not named). |
Spanning labels (NIEDERHORN-2 to 11), within the range: all hold. e03, e04, e05, e11, e12 rejected at all 241 instants, e07 at 239, e08 at 240. PWD_RMP_B15_017_C09 rejected every night (about 23:00 → 13:00), as NIEDERHORN-9 says. e01 rejected by day on every day except the overcast 01/10 ("almost every day" holds). PWD_IMP_B08_024_C09 rejected most days from late morning to evening, as NIEDERHORN-11 says. Nothing to report.
## 1. My labels
> **Requalification of 2026-11-17** (new rules, `labeling-rules.md` 1): 10 of the labels below are removed, NIEDERHORN-155 is rewritten; see section 2 and "## Requalification". The subsections of the removed labels (01/10 23:00 → 02/10 01:00, 03/10 09:00 → 11:00, the 05/10 sunrise in the grouped sun subsection, 06/10 10:00 → 12:00) stay as a record of the reading, but those labels are no longer in `labels.jsonl`.
The 29 kept labels are not repeated here (their evidence is in the table above). Corrected and added labels:
### 2026-10-01T02:00Z → 2026-10-01T03:00Z (added)
- **Phenomenon of the catalogue**: A plant shaken by the wind in front of a target
- **Precedents**: NIEDERHORN-29, -30, -35, -40 (same target, same plant; excluded from `precedents.py`, read in `existing.jsonl`)
- **Family**: Obstruction (one target, an object in front)
- **Measured**: PWD_IMP_B07_016_C09 corr_x −0.36, corr_y −0.43 px against its level (noise 0.05 / 0.07), error_value 0.18 against 0.24; 03:00 back. First measurement after the 01:00 gap: start on that measurement (labeling-rules 2, edge of a gap).
- **Images**: thumbnail 02:00 (fixed frame, with PWD_IMP_B07_017_C09 as witness): a leaf tip touches the spot; 00:00, 04:00 clean.
- **Confidence**: high
### 2026-10-01T04:00Z → 2026-10-01T17:00Z (NIEDERHORN-12, corrected)
- **Phenomenon of the catalogue**: Rain
- **Precedents**: NIEDERHORN-16, -39 (same station)
- **Family**: Visibility (whole scene veiled)
- **Measured**: from 05:00, 5 to 15 targets rejected per measurement (inventory); corr_x dips of the left targets: PWD_RMP_B15_020_C09 −22, e15 −20, e10 −20, e14 −15, e13 −13, PWD_IMP_B08_025_C09 −17 noises; references calm. Last touched: 16:00 (PWD_IMP_B07_018_C09 corr_y 0.10 against −0.03 the evening; PWD_RMP_B17_026_C09 corr_x 0.25 against 0.10; e09 corr_x 0.35 against 0.22); 17:00 back.
- **Images**: 07:00: grey noisy background, faint out-of-focus drops, weaker spots; 04/10 07:00 for comparison: clean.
- **Confidence**: high
### 2026-10-01T05:00Z → 2026-10-01T16:00Z (added)
- **Phenomenon of the catalogue**: A plant shaken by the wind in front of a target
- **Precedents**: NIEDERHORN-17 (the same target labelled apart from the rain of 02/10)
- **Family**: Obstruction
- **Measured**: PWD_IMP_B07_016_C09 corr_x +0.36 (06:00), +0.30, +0.25, −0.21 (09:00), +0.46 (11:00), −0.40 (14:00), +0.40 (15:00), error_value dropping to 0.09-0.12 when it jumps (the plant signature on this target); 05:00 and 16:00 clean. Its neighbours PWD_IMP_B07_017_C09 and PWD_IMP_B07_018_C09 are rejected by the rain, not moved like this.
- **Images**: thumbnail 14:00: a leaf touches the spot; 11:00: no leaf in the frame at that instant.
- **Confidence**: medium (inside the rain; the plant is seen at 14:00 only)
### 2026-10-01T11:00Z → 2026-10-01T14:00Z (NIEDERHORN-13, corrected)
- **Phenomenon of the catalogue**: A pass-by hides the targets for the length of one measurement
- **Precedents**: NIEDERHORN-31, -34, -36, -52
- **Family**: Obstruction
- **Measured**: 12:00: 15 targets rejected; under the threshold corr_y PWD_RMP_B12_019_C09 +23, PWD_RMP_B12_020_C09 +22, PWD_RMP_B12_018_C09 +15, PWD_RMP_B12_017_C09 +8 noises; PWD_IMP_B07_016_C09 not moved at 12:00.
- **Images**: 12:00 image cited by the original label (freight train); not re-read.
- **Confidence**: high
### 2026-10-01T19:00Z → 2026-10-01T21:00Z (added)
- **Phenomenon of the catalogue**: none (Displacement "jump then return, cause not established")
- **Precedents**: LENZSPITZE-98 (a single target jumps for one measurement, cause not established, auto not reviewed); WYOMING-17 (reviewed, one-measurement jump, cause not established)
- **Family**: Displacement (no established cause)
- **Measured**: e10 corr_y −0.07 → +0.32 → −0.03 (13 noises), raw y +0.4 px, error_value 0.065 → 0.088; PWD_IMP_B08_025_C09 corr_x 0.23 → 0.11 → 0.33 (6 noises), error_value 0.134 → 0.065. Pose and references still.
- **Images**: 20:00 whole image: overcast daylight, nothing in front; thumbnails 19:00, 20:00, 21:00 of both: spots intact.
- **Confidence**: medium
### 2026-10-01T23:00Z → 2026-10-02T01:00Z (added)
- **Phenomenon of the catalogue**: none
- **Precedents**: as above
- **Family**: Displacement
- **Measured**: PWD_IMP_B08_025_C09 corr_x 0.355 → 0.176 → 0.283 (about 5 noises against the mean of its neighbours), error_value 0.161 → 0.081 → 0.159; others still.
- **Images**: thumbnail 00:00: spot intact.
- **Confidence**: low (at the threshold of 5 noises; kept because the error_value drops as at 20:00)
### 2026-10-02T01:00Z → 2026-10-02T03:00Z (added)
- **Phenomenon of the catalogue**: A plant shaken by the wind in front of a target
- **Family**: Obstruction
- **Measured**: PWD_IMP_B07_016_C09 corr_x +0.23, corr_y +0.50 (7 noises), error_value 0.14 against 0.22; 03:00-04:00 clean (two measurements: a separate episode from NIEDERHORN-17).
- **Images**: thumbnail 02:00: leaves touch the spot.
- **Confidence**: high
### 2026-10-02T01:00Z → 2026-10-02T15:00Z (NIEDERHORN-16, corrected)
- **Phenomenon of the catalogue**: Rain
- **Family**: Visibility
- **Measured**: 02:00: e14 corr_x 0.25 → 0.13, corr_y 0.01 → −0.26, PWD_IMP_B07_018_C09 corr_x 0.16 → −0.01, error_value of e14 0.037 → 0.065, of PWD_RMP_B16_026_C09 0.062 → 0.083 (01:00 still clean). 03:00-11:00: 3 to 12 targets rejected per measurement. 14:00 (after the 12:00 gap): PWD_RMP_B15_020_C09 corr_x −0.37 (level about 0.02), error_value 0.14 (level 0.045), e09 error_value 0.19 (level 0.11); 15:00: PWD_RMP_B15_020_C09 −0.08, error 0.078, others back.
- **Images**: 07:00: large out-of-focus drops (two big ellipses), streaks of rain at the top, weaker spots.
- **Confidence**: high
### 2026-10-02T04:00Z → 2026-10-03T02:00Z (NIEDERHORN-17, corrected)
- **Phenomenon of the catalogue**: A plant shaken by the wind in front of a target
- **Family**: Obstruction
- **Measured**: PWD_IMP_B07_016_C09 05:00 corr_x +0.21, corr_y −0.30, error 0.11; 06:00 +0.11/−0.17, error 0.12; 07:00 +0.58; then as described by the original label; 03:00 and 04:00 clean.
- **Images**: thumbnail 05:00: leaves over the spot.
- **Confidence**: high
### 2026-10-03T09:00Z → 2026-10-03T11:00Z (NIEDERHORN-22, corrected)
- **Phenomenon of the catalogue**: A pass-by hides the targets for the length of one measurement (cause not seen)
- **Family**: Obstruction (kept: the spots are hidden, an object deduced)
- **Measured**: 10:00: e01 error_value 0.96 (0.25 around), e02 0.97 (0.21 around), both rejected; raw positions of both jump (latched elsewhere). e01 is normally measured at night.
- **Images**: thumbnails 09:00 and 10:00 of e01, e02: both spots gone at 10:00, a small spot next to e01 gone too.
- **Confidence**: high on the facts, cause not established
### 2026-10-03, -04, -05, -08, -09, each 12:00 → 14:00 (NIEDERHORN-23, -25, -27, -41, -45, corrected)
- **Phenomenon of the catalogue**: The sun in the field or behind a spot
- **Precedents**: NIEDERHORN-32 (same station, the image of its own day)
- **Family**: Visibility
- **Measured** (error_value of the references at 13:00, against 11:00; spikes against the neighbours at 13:00, in noises): 03/10 PWD_REF10_C09 0.14 → 0.40, PWD_REF11_C09 0.10 → 0.18, PWD_REF08_C09 0.23 → 0.30; e02 corr_x +7, PWD_IMP_B07_018_C09 −6, PWD_IMP_B07_016_C09 −5, e13 corr_y −5. 04/10 PWD_REF10_C09 0.42, PWD_REF11_C09 0.20; PWD_IMP_B07_018_C09 −12, PWD_RMP_B12_019_C09 −7, PWD_RMP_B12_018_C09 −6, PWD_IMP_B07_017_C09 −6, e10 corr_y −7, e13, e15 −6, PWD_RMP_B15_020_C09 −5. 05/10 PWD_REF10_C09 0.34, PWD_REF08_C09 0.30; no target beyond 4 noises. 08/10 PWD_REF10_C09 0.40, PWD_REF08_C09 0.31, PWD_REF11_C09 0.20; e02 +7, e10 +6, PWD_IMP_B07_017_C09 −6, PWD_REF11_C09 / PWD_REF12_C09 corr_x −7 / +6. 09/10 PWD_REF10_C09 0.39, PWD_REF08_C09 0.33, PWD_REF09_C09 0.26, PWD_REF11_C09 0.19; e02 +11, e06 +8, e10, e13, e15 +6 to +7, e14 +5, PWD_IMP_B07_017_C09, PWD_IMP_B07_018_C09 −6, PWD_REF11_C09 / PWD_REF12_C09 −6 / +8. At 14:00 the alternation of signs is only the return of the 13:00 point: the bounds 12:00 → 14:00 hold. e06 and e09 are measured at 13:00 on all five days (rejected only at the train instants of 01, 06, 07/10).
- **Images**: 13:00 of each day (`sun_sunrise_13utc.png`): sun at the top of the field, rays, halo and flares on 03, 04, 05, 08, 09/10; overcast without sun on 07/10.
- **Confidence**: high (05/10: medium on the effect)
### 2026-10-04T17:00Z → 2026-10-04T19:00Z (added)
- **Phenomenon of the catalogue**: A plant shaken by the wind in front of a target
- **Family**: Obstruction
- **Measured**: PWD_IMP_B07_016_C09 at 18:00: corr_x +0.28 (5.5 noises), corr_y +0.29 → −0.02 → +0.24, error_value 0.15 against 0.21; neighbours only follow the thermal cycle.
- **Images**: thumbnail 18:00: blurred leaves across the spot; 16:00: a leaf beside it.
- **Confidence**: high
### 2026-10-06T10:00Z → 2026-10-06T12:00Z (added)
- **Phenomenon of the catalogue**: A faint spot lost when the light is missing
- **Precedents**: none useful (`precedents.py` returns Target issue cases of permanent threshold targets)
- **Family**: Visibility (light change at dawn, nothing in front); not Target issue: its only other rejections of the range come with the rain
- **Measured**: PWD_RMP_B12_019_C09 error_value 0.285, 0.287, 0.299 at 08-10:00, 0.301 at 11:00 (threshold 0.3), 0.265 at 12:00; raw position without a jump.
- **Images**: thumbnails 10:00, 11:00: spot intact, smaller at 11:00 in a darker image, nothing in front.
- **Confidence**: medium
### 2026-10-06T11:00Z → 2026-10-06T13:00Z (NIEDERHORN-31, corrected)
- **Phenomenon of the catalogue**: A pass-by hides the targets for the length of one measurement
- **Family**: Obstruction
- **Measured**: 12:00 corr_y PWD_RMP_B12_019_C09 +30, PWD_RMP_B12_020_C09 +23, PWD_RMP_B12_018_C09 +19, e13 +11, PWD_RMP_B12_017_C09 +9, PWD_IMP_B07_017_C09 corr_x +6 noises.
- **Images**: as cited by the original label.
- **Confidence**: high
### 2026-10-07T13:00Z → 2026-10-07T23:00Z (NIEDERHORN-37, corrected)
- **Phenomenon of the catalogue**: A plant shaken by the wind in front of a target
- **Family**: Obstruction
- **Measured**: rejections 14:00, 16:00-17:00; then corr_y +0.25 (18:00, thermal level of the day) → −0.07 (19:00) → +0.37 with error 0.09 (20:00) → +0.02 (21:00) → −0.45 with corr_x +0.24, error 0.11 (22:00) → −0.02 (23:00).
- **Images**: thumbnails 19:00, 20:00, 22:00: leaves across the spot; GIF `images_20261007T1300Z_20261007T2300Z.gif`.
- **Confidence**: high
### 2026-10-08T11:00Z → 2026-10-08T13:00Z (added)
- **Phenomenon of the catalogue**: none
- **Family**: Displacement (no established cause)
- **Measured**: PWD_RMP_B16_026_C09 corr_y 1.468 → 1.687 → 1.366 (8 noises), corr_x and error_value unchanged (0.103-0.104); nothing else at 12:00 (the sun and thermal labels start there).
- **Images**: thumbnail 12:00: very dark image, spot smaller but present, nothing in front.
- **Confidence**: medium
### 2026-10-08T21:00Z → 2026-10-08T22:00Z (added)
- **Phenomenon of the catalogue**: A plant shaken by the wind in front of a target
- **Family**: Obstruction
- **Measured**: PWD_IMP_B07_016_C09 corr_y +0.52 (level +0.02 at 19:00, +0.05 at 22:00), first measurement after the 20:00 gap; nothing else moves except PWD_REF11_C09 a little.
- **Images**: thumbnail 21:00: a leaf on the spot.
- **Confidence**: high
### 2026-10-09T22:00Z → 2026-10-10T02:00Z (NIEDERHORN-49, corrected)
- **Phenomenon of the catalogue**: A plant shaken by the wind in front of a target
- **Family**: Obstruction
- **Measured**: 23:00 rejected; 00:00 corr_y −0.54; 01:00 corr_y +0.20 (level −0.04), error 0.08; 02:00 corr_x +0.18 (2.4 noises), error 0.16: back.
- **Images**: thumbnails 00:00, 01:00, 02:00: leaves touching the spot.
- **Confidence**: medium (01:00 is 3.4 noises, the error_value at its lowest)
### 2026-10-10T12:00Z → 2026-10-10T14:00Z (added)
- **Phenomenon of the catalogue**: The sun in the field or behind a spot
- **Family**: Visibility
- **Measured**: PWD_REF09_C09 error_value 0.15 → 0.37 (12:00) → 0.24; PWD_REF10_C09 0.17 → 0.25, PWD_REF11_C09 0.11 → 0.16 (13:00); PWD_IMP_B08_025_C09 corr_x 0.28 → 0.43 → 0.24, e10 0.40 → 0.51 → 0.39 (13:00, +8 noises each against the neighbours).
- **Images**: 12:00: sun glow behind the buildings at the left; 13:00: glare in the haze at the top, scene darkened.
- **Confidence**: medium
## 2. What I left without a label
### Removed at the requalification (2026-11-17)
Deviations at the touched measurements, all targets, corr_x and corr_y, in each target's own noise: "global" against the median of the range (as `ranges.py --k 5`), "rolling" against the median of the 6 measurements on each side (`--rolling 6`), the reference under a thermal cycle or a slow level. "Rejected": filtered by the station (missing point).
- **NIEDERHORN-15**, 2026-10-01T20:00Z → 22:00Z, Obstruction (train, 21:00): missing points only (e01, e02, e06, e09, PWD_RMP_B17_026_C09 rejected). Global: e10 corr_x −8.0, PWD_RMP_B16_026_C09 corr_x −6.7, but both are the slow post-rain evening level (e10 already −10.9 at 20:00 and −11.2 at 23:00); rolling max 3.8 (PWD_RMP_B16_026_C09 corr_y). Nothing jumps with the train.
- **NIEDERHORN-145**, 2026-10-01T23:00Z → 02T01:00Z, Displacement (PWD_IMP_B08_025_C09 dip, 00:00): PWD_IMP_B08_025_C09 corr_x −2.4 noises rolling, under 5 global; max of all targets rolling 3.7 (e10 corr_y); global e10 corr_x −9.5 is the slow level of that night. Under the noise.
- **NIEDERHORN-21**, 2026-10-03T06:00Z → 08:00Z, Obstruction (plant, 07:00): missing point only (PWD_IMP_B07_016_C09 rejected alone; a target, no effect on the compensation). Rolling max 1.9; global 5.2 on PWD_RMP_B12_019_C09 corr_x, its ordinary night level, not a jump.
- **NIEDERHORN-149**, 2026-10-03T09:00Z → 11:00Z, Obstruction (e01, e02 hidden, 10:00): missing points only. Rolling max 3.8 (e06 corr_y); global 6.7 / 6.2 on PWD_RMP_B15_018_C09 corr_y and PWD_RMP_B12_019_C09 corr_x, the night level, not a jump.
- **NIEDERHORN-153**, 2026-10-05T12:00Z → 14:00Z, Visibility (sun at sunrise, 13:00): at the start of the thermal cycle (NIEDERHORN-28), rolling max 4.7 (PWD_RMP_B15_018_C09 corr_y); the global 6.4 on PWD_IMP_B07_018_C09 corr_x is the onset of the thermal drift, already in NIEDERHORN-28. The label itself said "the other targets hardly". Under the noise.
- **NIEDERHORN-30**, 2026-10-06T03:00Z → 05:00Z, Obstruction (plant, 04:00): missing point only (PWD_IMP_B07_016_C09 rejected). Rolling max 1.7; global 6.5 on PWD_RMP_B12_019_C09 corr_x, its night level.
- **NIEDERHORN-154**, 2026-10-06T10:00Z → 12:00Z, Visibility (PWD_RMP_B12_019_C09 rejected at dawn, 11:00): missing point only. At 11:00, rolling +7.5 on PWD_IMP_B07_018_C09 corr_x and +5.5 on e09 corr_y, global +8.6 on e09 corr_y; but the raw values at 11:00 continue those of 09:00-10:00 (PWD_IMP_B07_018_C09 0.438, 0.467, 0.456 px; e09 −0.025, −0.062, −0.013 px): no jump; the rolling deviation comes from the thermal drop of the following hours inside its window, the global one from the slow pre-dawn level. Not an effect of the rejection.
- **NIEDERHORN-35**, 2026-10-06T22:00Z → 07T00:00Z, Obstruction (plant, 23:00): missing point only (PWD_IMP_B07_016_C09 rejected, a target). At 23:00 PWD_RMP_B15_018_C09 and PWD_RMP_B15_019_C09 corr_x make a one-measurement spike (rolling +7.6 and +6.5) while PWD_IMP_B07_018_C09 regains its night level in one step: the return of the thermal cycle, inside NIEDERHORN-33 (→ 07/10 00:00), not caused by the plant.
- **NIEDERHORN-43**, 2026-10-08T16:00Z → 18:00Z, Obstruction (plant, 17:00): missing points only (PWD_RMP_B12_017_C09 and PWD_IMP_B07_016_C09 rejected, both targets). The deviations at 17:00 (global −10.3 PWD_IMP_B07_017_C09, rolling +6.8 e06) are the thermal cycle and its hot-hour agitation, inside NIEDERHORN-42.
- **NIEDERHORN-48**, 2026-10-09T19:00Z → 21:00Z, Obstruction (train, 20:00): missing points only (e01, e02, e06, e09, PWD_RMP_B17_026_C09 rejected). Rolling max 4.6 (e13 corr_y); global 6.4 on PWD_RMP_B12_018_C09 corr_x, the evening thermal level (NIEDERHORN-46).
### Earlier notes (first analysis)
- **01/10, no thermal label** (as the existing set): the housing temperature stays between 33 and 38 °C all day (overcast); pan, tilt, rotation never beyond 5 noises from the night; the scale rises a little (5.5 noises at 19:00-20:00 only); corr_x median within 3 noises. Below the rule of labeling-rules 1 ("no thermal label when neither the pose nor the corr deviate by more than 5 noises"), the scale being a secondary axis.
- **Night levels that differ from 01-02/10 to the following nights** (corr_x of the left targets, corr_y of PWD_REF11_C09 / PWD_REF12_C09, sigma0): the 01-02/10 nights are warm (33-36 °C) and overcast, the following nights cold (23-28 °C) after sunny days; sigma0 follows the temperature from 02/10 on. Not a step with its own date: the night level follows the housing temperature, part of the thermal regime; nothing on the curves within a night beyond the thermal labels.
- **PWD_IMP_B07_016_C09, small moves under 4 noises**: 05/10 04:00 (corr_y −0.16), 05/10 10:00 (corr_y +0.24, 3.4 noises, error 0.13), 06/10 01:00-02:00 and a slightly raised corr_x 06/10 05:00-11:00, 07/10 07:00, 08/10 23:00: the plant probably, below the threshold, without a rejection.
- **Thermal agitation** at the hot hours (spikes of 5-8 noises on e02, e06, e09, e13, e15, PWD_IMP_B07_017_C09, PWD_IMP_B08_025_C09, the references PWD_REF11_C09 / PWD_REF12_C09 in opposite directions): written in the thermal labels ("agitated from one measurement to the next").
- **Neighbour artefacts of my spike script** (06/10 11:00, 07/10 11:00 and 13:00, 11/10 02:00 and 04:00, 03/10 and 04/10 12:00, 08/10 00:00, 09/10 12:00, 15:00, 17:00): a point flagged only because its neighbour is the spike (train, sun, bird); the values themselves are at their level.
- **08/10 23:00 PWD_IMP_B08_024_C09 corr_x** (5 noises against its neighbours, just after the 22:00 point, at the end of the thermal label): at the edge of the thermal cycle, not a separate event.
- **Omissions in kept labels** (true but incomplete, left to keep the user's validation): NIEDERHORN-1, -14, -18, -51 do not say that the image of the missing hour is missing too and that NIEDERHORN alone misses it (-44 says it); NIEDERHORN-15 and -52 do not name e01 (and, in -52, PWD_IMP_B08_024_C09) among the rejected; NIEDERHORN-39 does not name PWD_IMP_B07_018_C09 (rejected 02:00-04:00) nor PWD_RMP_B12_019_C09 (01:00).
## 3. The documentation
- `labeling-rules.md` 2 (bounds: "a measurement that has come back halfway is still touched", "at the edge of a gap") decided NIEDERHORN-12, -16, -49 and the two added labels after a gap; 3 ("an intermittent episode is cut at a pause of several measurements") is not quantified: I took two clean measurements as a pause (01/10 02:00 apart from 05:00; 02/10 02:00 apart from NIEDERHORN-17) and one as no pause (NIEDERHORN-37 extended over 19:00-22:00). A number would help.
- `labeling-rules.md` 5 ("a sentence taken from a model is reread on the range of each label", "the image cited is that of the day") is exactly what the sun labels broke: the same text on five days, citing the image of 21.10 and e06/e09 rejections that do not happen on these days.
- `families.md`: the plant (Obstruction), rain, sun, pass-by phenomena fitted. The one-measurement jumps of one or two targets without cause (01/10 20:00, 08/10 12:00) have no phenomenon of their own; Displacement by default, as the precedents WYOMING-17, LENZSPITZE-98.
- A signature of the plant on this station, not written anywhere: the error_value of PWD_IMP_B07_016_C09 **drops** (to 0.08-0.15 against 0.20-0.24) when a leaf touches its spot without hiding it, while its corr jumps; it helped find the forgotten hits. Worth a line in `reading-rules.md` or `stations.md` (NIEDERHORN: the foreground plant in front of PWD_IMP_B07_016_C09 and PWD_RMP_B12_017_C09).
- `stations.md` (NIEDERHORN: references on the road above, floodlight at night) helped; it could add that the station is outside Switzerland (no weather) and that its night is about 23:00-11:00 UTC, which `days.py --night` needs.
- `tools/GUIDE.md`: no tool gives the one-measurement spike against the two neighbours per instant for all sensors; my `scripts/instants.py` did, with its neighbour artefacts. `thumbnails.py --fixed` with a witness target was decisive for the plant.
- Precedents: the station's own labels are excluded from `precedents.py` (by `--exclude` on the range); those of other stations brought little except the "cause not established" wording.
## Requalification
Rules of 2026-11-17 (`labeling-rules.md` 1): a movement is kept if at least one target deviates by more than 5 of its own noises over the label's range; a thermal label if the corrected curves keep it; a label of missing points is removed unless it moves the other curves; No data only when nothing is delivered.
Commands (outputs in `work/requalify/`):
- `python3 tools/scripts/ranges.py data/NIEDERHORN/ --columns corr_x,corr_y --start 2026-10-01T00:00Z --end 2026-10-11T05:00Z` → `ranges_k5.txt`
- `python3 tools/scripts/ranges.py data/NIEDERHORN/ --columns corr_x,corr_y --rolling 6 --start 2026-09-30T12:00Z --end 2026-10-11T05:00Z` → `ranges_k5_rolling6.txt`
- `python3 scripts/requalify.py work/requalify_input.jsonl` + both outputs → `per_label.txt`: the k5 ranges of the targets overlapping the touched measurements of each label
- `python3 scripts/maxdev.py data/NIEDERHORN work/requalify_input.jsonl` + ids → `maxdev.txt`: the largest deviation in noises at the touched measurements, all targets, global and rolling (same level and noise as `ranges.py`)
- No data: number of measurements of all sensors at 01/10 01:00, 01/10 13:00, 02/10 12:00, 08/10 20:00, 11/10 00:00: 0 each (35 at the neighbouring hours).
Result: 52 in, 41 kept as they are, 1 rewritten, 10 removed (section 2), 42 out.
Values of the kept labels (largest target deviation at the touched measurements, global / rolling, in noises):
- No data NIEDERHORN-1, -14, -18, -44, -51: no measurement of any sensor; kept.
- Plant on PWD_IMP_B07_016_C09: -140 corr_x −7.1 / −7.5; -142 corr_x +8.9 global (11:00), −9.6 rolling (14:00); -146 corr_y +7.0 / +7.2; -148 corr_x +20.5 / +19.2 (02/10 08:00); -152 corr_x +5.5 / +6.7; -29 corr_y +7.8 / +7.2 (16:00); -156 corr_y −6.3 / −6.2 (22:00); -40 corr_x +9.8 / +9.9 (02:00), corr_y −8.2 / −7.9 (06:00); -159 corr_y +7.4 / +7.0; -161 corr_y −7.6 / −8.0 (00:00).
- Rain: -141 e10 corr_x −31.4 / −23.1; -147 e15 corr_x −32.8 / −29.8; -39 PWD_RMP_B15_019_C09 corr_y +26.1 / +27.3.
- Trains: -143 PWD_RMP_B12_019_C09 corr_y +26.1 / +26.4; -20 PWD_RMP_B17_026_C09 corr_x +7.9 global, corr_y +5.6 rolling; -155 PWD_RMP_B12_019_C09 corr_y +31.5 / +29.9; -34 +38.3 / +40.7; -36 PWD_RMP_B15_019_C09 corr_y +40.7 / +41.4; -52 PWD_RMP_B12_020_C09 corr_y +22.0 / +21.2.
- Sun at sunrise: -150 PWD_IMP_B07_018_C09 corr_x −9.8 global, e02 +7.9 rolling; -151 PWD_IMP_B07_018_C09 −12.5 / −13.6; -32 PWD_IMP_B08_024_C09 corr_x −15.7 / −16.8, e10 +7.8 rolling; -158 e02 +9.4 / +7.0; -160 e02 +10.8 / +10.2; -162 PWD_IMP_B08_025_C09 corr_x +4.7 / +5.5 (kept: beyond 5 against its local level, a spike against the start of the thermal drift; the weakest kept label).
- One-measurement jumps without cause: -144 e10 corr_y +14.1 / +11.8; -157 PWD_RMP_B16_026_C09 corr_y +9.8 / +9.0.
- Bird in front of PWD_REF08_C09: -47 PWD_IMP_B08_025_C09 corr_y +21.4 / +20.1, almost all targets beyond 5.
- Thermal (corrected curves; the global level is close to the night level, the rolling one at 6 h): -19 PWD_IMP_B07_018_C09 corr_x −12.4 / −9.0; -24 PWD_IMP_B07_017_C09 −19.5 global, e01 +18.2 rolling; -26 PWD_IMP_B07_018_C09 −12.5 / −13.6; -28 PWD_IMP_B07_017_C09 −16.6 / −11.6; -33 e01 +21.9 / +23.2; -38 e01 +12.3 global, e06 +9.8 rolling; -42 PWD_IMP_B07_018_C09 −15.2 / −9.4; -46 PWD_IMP_B07_018_C09 corr_x and many others beyond 5 outside the bird hour; -50 PWD_IMP_B07_018_C09 −9.2 global, e06 +7.0 rolling. All keep the cycle on the corrected curves.
Rewritten: NIEDERHORN-155 (train at dawn, 06/10 12:00; PWD_RMP_B12_019_C09 corr_y +31.5): its parenthesis "the rejection of PWD_RMP_B12_019_C09 at the measurement before is labelled apart" pointed to NIEDERHORN-154, removed; dropped, the rest unchanged (it had no review field).
Kept with descriptions that also list rejected targets (trains, rain, plant): their substance is the movement measured above; the lists of rejections stay as context and the user's validation is kept.
## Difficulties
- `tools/scripts/data.py` returns measurement times with seconds (`17:00:07`) and camera / temperature times on the minute: my first script matched nothing until I rounded to the minute.
- One Bash call was refused once by the auto-mode classifier (transient), then passed as is.
## Sources read
- requalification: `instructions/analyst.md`, `labeling/labeling-rules.md` 1-3, `tools/GUIDE.md` (ranges), `work/requalify_input.jsonl`, this analysis, `report.md`, `tools/scripts/inventory.py` (distances, mm/px)
- `agents/agents/analyst.md`, `instructions/analyst.md`, the request (`work/prompt.txt`, `request.json`), `work/existing.jsonl`, `work/spanning.jsonl`
- `tools/GUIDE.md`, `tools/DATA.md`, `tools/scripts/data.py` (to write my scripts)
- `labeling/families.md`, `labeling/labeling-rules.md`, `labeling/reading-rules.md`, `labeling/stations.md`
- `labeling/auto/NIEDERHORN.jsonl` (labels at the edges: NIEDERHORN-53 at 11/10 05:00 → 07:00, outside the range)
- precedents: WYOMING-17, WYOMING-13, LENZSPITZE-98, LENZSPITZE-65, BREITHORN-3, SNCF01-1, SNCF01-5, NIEDERHORN-79
- data: `data/NIEDERHORN/` (advanced exports, camera, metrics, images), `stations.py` on ALASKA, CLARIDEN, FLORIDA, JUNGFRAU, NORDEND, SNCF01, SUSTENHORN, WILDSTRUBEL, WYOMING, NADELHORN, LENZSPITZE (gaps only)
- outputs in `work/`: `inventory.txt`, `overview.txt` and its PNGs, `days.txt`, `stations.txt`, `instants.txt`, thumbnails