Files
aimsight-labeling/auto/analyses/CLARIDEN_2026-10-11T0000_2026-10-20T0400/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

148 lines
20 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: CLARIDEN, 2026-10-11T00:00Z → 2026-10-20T04:00Z (UTC)
The task was to verify the 8 existing labels of the range (`work/existing.jsonl`). The 2 labels that span the range (`work/spanning.jsonl`) are context only. Station: a tunnel under construction, cadence 10 min, five references (D-302-C, PR-291, PR-293, PR-295, PR-296). PR-296 is retained by the compensation at every instant, although `stations.md` lists four references. There are no image-quality metrics (inventory: export peak to snr empty), so `quality.py` does not apply. Images: one per hour, always dark (night-like infrared view).
**Weather**: none. There is no `data/CLARIDEN/weather/` folder and no CLARIDEN line in `~/aimsight-data/extraction/stations.txt`: the station has no Swiss position. In any case it is in a tunnel.
**Requalification of 2026-11-17** (new rule, `labeling-rules.md` §1: only what stands out from the noise of the corrected curves, at least one target beyond 5 of its own noises). The work below is the verification already done; the ids are now the current ones (old id → current: 15 → 49, 16 → 52, 17 → 17, 18 → 55, 19 → 56, 20 → 57, 21 → 58, 22 → 59; the added labels became 50, 51, 53, 54). Result: 10 kept as they are, 1 rewritten (CLARIDEN-49), 1 removed (CLARIDEN-51). The values are in section 2 and in "## Requalification".
## Verdict per old id
| id | verdict | evidence |
|---|---|---|
| CLARIDEN-15 | **corrected** (end, description) | 3 h medians: D-304-C corr_y goes from about −0.48 (14th–15th) to −0.494 at the first steps after the gap of the 16th (10:30), then stays at −0.49…−0.50 until the 17th. D-303-C goes from −0.13 to −0.14…−0.147 and stays there. D-301-C corr_x has been flat around −0.08…−0.09 since the 14th. The return is over at 2026-10-16T12:00Z; afterwards it is a plateau, and a plateau is not a return (labeling-rules §2). The further dip of D-303-C and D-304-C on 18–19 follows the camera tilt and comes back at the jump of the 20th 00:30, so it belongs to the new tilt label. The description said "No camera movement during these days", but the tilt rises from about −2 to +0.3 thousandths of a degree on the 11th–13th and falls to −4.6 on the 19th. |
| CLARIDEN-16 | **corrected** (start 22:45 → 22:50) | Last measurement before the gap: 22:50 (inventory: "between 22:50 and 10:30: 69 missing"). 22:45 is not a measurement instant at a 10 min cadence. The description is kept. |
| CLARIDEN-17 | **kept** | Single missing step 14:30; bounds 14:20 → 14:40 are measurement instants; metrics normal (BusVoltage ~14.36 V). |
| CLARIDEN-18 | **corrected** (start 09:00 → 08:50, description) | At 09:00 the row A targets are already up in corr_y: D-300-A +0.075, D-302_A +0.095, D-303-A +0.098, D-304-A +0.004 (usual about −0.065), D-301-A −0.056 (usual −0.12) and P-298-A +0.053, each about +3 to +3.5 noises, together. D-302-B is at 0.094. At 09:10–09:20, error_value rises on 9 to 11 sensors (PR-293, PR-295, D-300-C, D-301-C, D-302-B, D-303-B, D-303-C, D-304-B…). D-302-B/PR_BG corr_y reaches 0.154 (+8 noises), D-301-C corr_x −0.21 (−9 noises), P-301-B corr_x +0.05. Everything is back at 09:30. The 09:00 image shows nothing in front of the targets (looked at; same as 08:00). Family Visibility kept: sharp edges and an error_value rising on most sensors, like the reviewed CLARIDEN-12. The old text named neither the row A targets nor the 09:00 image. |
| CLARIDEN-19 | **corrected** (start 13:45 → 13:50) | The last measurement is 13:50. Metrics: none from 15:00 to 21:00 (metrics.py), BusVoltage 13.26 V at the restart then rising: the description holds. |
| CLARIDEN-20 | **corrected** (end 10:15 → 10:10) | Present instants: 09:00, 09:30, 09:50, 10:10. The first measurement after the last gap is 10:10. |
| CLARIDEN-21 | **corrected** (start 20:45 → 20:40) | Present: 20:40, 21:10, 21:30; the gap starts after 20:40. |
| CLARIDEN-22 | **corrected** (bounds 00:15→00:45 → 00:10→00:40, description) | Raw y of the targets: −0.0325 at 00:10, +0.005 at 00:20 (already moving), +0.165 at 00:30, then a plateau. The tilt goes from −0.0027 to −0.0012 to +0.0023. The new level is reached at 00:30, so the end is 00:40. In corr_y, D-306-B goes 0.034 → 0.105 (+7 noises) and stays, beyond its level of the 16th–17th (≈0.05). D-303-C goes −0.157 → −0.124, D-304-C −0.518 → −0.477, D-304-B −0.075 → −0.058, P-301-B and D-303-B also rise: these jumps bring them back to their level of before the 17th, so the jump erases the tilt remainder. D-302-B/PR_BG jump in corr_x (2 h medians 0.064 → 0.090 → 0.101) and do not jump in corr_y. The old description ("D-303-C, D-302-B and PR_BG a smaller jump that remains") was wrong on the nature of most of the jumps. Images of 00:00 and 01:00: work lights at the far end of the tunnel, nothing in front of the targets. |
**Spanning labels** (not changed):
- **CLARIDEN-1** (P-300-B Target issue): holds in the range. P-300-B is rejected at all 1187 instants (inventory), with error_value about 0.93.
- **CLARIDEN-14** (drift of the 6th–11th, ends 11th 06:00): its in-range part (11th 00:00–06:00) is a plateau, which is consistent. But its sentence "No camera movement" is contradicted just before the range: the tilt goes from +0.8 (9th 00h) to −2.0 (11th) thousandths of a degree (6 h table, `work/tilt_vs_targets_6h.txt`), and D-306-B corr_y follows it down (0.045 → 0.022). This is outside my range (9th–11th); it is to be checked by whoever verifies the range of the 6th–11th.
## 1. My labels
### 2026-10-11T06:00Z → 2026-10-16T12:00Z (CLARIDEN-49, ex-15; rewritten at the requalification)
- **Phenomenon of the catalogue**: Targets move slowly, alone, then come back (return part).
- **Precedents**: CLARIDEN-14 (its outward), reviewed.
- **Family**: Displacement. No cause is established, and the movement does not follow the temperature (housing at 40.0–40.4 °C).
- **Measured**: 3 h medians (above). D-301-C corr_x −0.042 (11th) → −0.08/−0.09 (14th–15th). D-303-C corr_y −0.12 → −0.145, and D-304-C −0.455 → −0.495, reached at 16th 09:00–12:00. These movements do not follow the tilt: they continue on the 13th–15th while the tilt is flat around 0.
- **Images**: 13th 12:00, a whole-zone thumbnail: nothing in front of the targets.
- **Confidence**: medium. The end is read on medians of about 0.01 px steps, close to the noise (MAD 0.011 per measurement).
### 2026-10-15T22:50Z → 2026-10-16T10:30Z (CLARIDEN-52, ex-16)
- **Phenomenon**: A gap common to several stations, plus a station restart on its battery (as described).
- **Family**: No data. **Measured**: 69 missing steps; no metrics 00:00–09:00; BusVoltage 13.3 V at the restart, then 14.36 V. **Images**: none from 22:00 to 10:30. **Confidence**: high.
### 2026-10-16T14:20Z → 2026-10-16T14:40Z (CLARIDEN-17)
No data, one step; high confidence.
### 2026-10-18T08:50Z → 2026-10-18T09:30Z (CLARIDEN-55, ex-18)
- **Phenomenon**: Visibility drops over the whole scene, without an obstacle (by its signature); like CLARIDEN-12.
- **Precedents**: CLARIDEN-12 (reviewed Visibility, same group D-301-C, D-302-B/PR_BG, D-304-B).
- **Family**: Visibility rather than Displacement. The error_value rises on most sensors, including references, with sharp edges (families.md boundary "Thermal or Visibility"). The pose jump comes with it.
- **Measured**: values above; the camera tilt goes −0.0012 → −0.0028 → −0.0038 → −0.0025 → −0.0008 (08:50 → 09:30). The housing temperature drops to 39.6 °C on the 08:00–10:00 step.
- **Images**: 08:00 and 09:00, whole image: identical, nothing in front.
- **Confidence**: medium. The bounds are sure; the cause is deduced, not seen.
### 2026-10-18T13:50Z → 2026-10-18T23:30Z (CLARIDEN-56, ex-19)
No data; signature of a power cut (low voltage at the restart). High confidence.
### 2026-10-19T09:00Z → 2026-10-19T10:10Z (CLARIDEN-57, ex-20) and 2026-10-19T20:40Z → 2026-10-19T21:30Z (CLARIDEN-58, ex-21)
No data; the common gaps with other stations are as described (not re-read on the other stations: the descriptions were validated). High confidence on the bounds.
### 2026-10-20T00:10Z → 2026-10-20T00:40Z (CLARIDEN-59, ex-22)
- **Phenomenon**: A step of the camera that the compensation does not erase.
- **Precedents**: WILDSTRUBEL-200 (a camera tilt not caught by the compensation, Displacement).
- **Family**: Displacement. The cause of the camera step is not established, and remainders stay on D-306-B and D-302-B.
- **Measured**: above (native table `summary.py --column corr_y --step 10min`, 19th 23:30 → 20th 01:30).
- **Images**: 00:00 and 01:00 (zone thumbnails): work lights at the far end of the tunnel.
- **Confidence**: high on the jump; medium on what stays.
### 2026-10-11T08:50Z → 2026-10-11T09:10Z (CLARIDEN-50)
- **Phenomenon**: none in the catalogue exactly. It is close to "The base changes without a reference lost", for a single measurement.
- **Precedents**: none useful.
- **Family**: Displacement: a jump then the return, without an established cause. The error_value is flat, so it is not Visibility.
- **Measured**: `scripts/group_flips.py`: 10 targets beyond 3 noises downward in corr_y at 09:00 (D-304-A −5.7, D-301-A −5.5, D-303-A −4.5, D-301-C −4.2, D-302_A −3.9, D-302-B/PR_BG −3.8, P-299-B −3.7, P-301-B −3.3, P-299-A −3.2). Raw values: D-301-A −0.212 against −0.11 around; D-304-A −0.176 against −0.05. Rigid fit (base 08:00): ty +0.063 and rotation −0.00114 at 09:00. D-302-C dy −0.010 against +0.04…+0.09 for the other references. The camera rotation hits its window minimum (−0.0054°) at 09:00. All is back at 09:10.
- **Images**: 09:00 (zone thumbnail): nothing in front.
- **Confidence**: medium. It is clearly a common one-measurement jump, but small (about 0.1 px).
### 2026-10-11T14:00Z → 2026-10-13T02:00Z (CLARIDEN-51): removed at the requalification
The camera straightens a little and the remainder on D-306-B, D-304-B, D-303-B and D-302-B/PR_BG stays under 5 noises: see section 2. Its old reading (2 h medians, `work/tilt_2h_11-13.txt`) already put it at 2 to 3 noises per measurement.
### 2026-10-17T16:00Z → 2026-10-20T00:30Z (CLARIDEN-53)
- **Phenomenon**: The camera tilts little by little; incomplete compensation.
- **Precedents**: BREITHORN-2 and BREITHORN-11 (the compensation corrects some targets better than others), SNCF01-7.
- **Family**: Displacement, as above. The temperature drops on the 18th (39.6 °C at 08:00, 39.55 °C at the restart), but the tilt keeps going down until 19th 04:30 while the housing is back at 40.1 °C. A gap does not cut it: the movement is visible on both sides of CLARIDEN-19.
- **Measured** (2 h medians, `work/tilt_2h_17-20.txt`): tilt −0.01 (17th 16h) → −1.0 (18th 00h) → −4.6 (19th 04h) → −2.4…−2.9 (19th afternoon) → +1.35 (20th 00h). Over the same time, corr_y moves as follows: D-302-B 0.041 → 0.098 (19th 04h), back to 0.073–0.085; D-303-B 0.062 → 0.026 (19th 08h); D-304-B −0.069 → −0.086; P-301-B −0.009 → −0.033; D-306-B 0.058 → 0.022; D-303-C −0.139 → −0.182; D-304-C −0.487 → −0.540. D-301-C corr_x −0.09 → −0.064. All but D-306-B and D-302-B come back at once at the jump of 00:30. `overview.py --others`: the same shapes remain after removing the median of the others, so they are specific to these targets. Rigid fit: the target |residual| median rises only slightly. Error_value: at its level (the 19th 01:50 instant with 5 sensors is isolated).
- **Images**: 19th 04:00 (zone thumbnail): nothing in front. 18th 07:00: work lights at the far end.
- **Confidence**: medium-high on the event and its bounds (start ±2 h, a slow onset). The family is uncertain between Displacement and a lagging thermal effect of the camera.
### 2026-10-18T06:30Z → 2026-10-18T06:50Z (CLARIDEN-54)
- **Phenomenon**: as on the 11th 09:00, in corr_x.
- **Family**: Displacement (error_value flat).
- **Measured**: group_flips: 7 targets at +3.2…+5.0 noises in corr_x at 06:40 (D-304-A 5.0, D-302_A 4.9, D-303-A 4.3, D-300-A 4.2, D-302-B/PR_BG 3.8, P-298-A 3.2). Rigid fit (base 06:00): tx ≈ 0, with PR-291 dx −0.06, PR-293 −0.03, PR-295 −0.03. The camera pan hits its minimum and the scale its maximum at 06:40 (overview, 16th–20th window). Back at 06:50.
- **Images**: no image at 06:40; the 06:00 and 07:00 images exist, and 07:00 shows work lights at the far end.
- **Confidence**: medium.
## 2. What I left without a label
**Removed at the requalification (2026-11-17):**
- **CLARIDEN-51**, 2026-10-11T14:00Z → 2026-10-13T02:00Z, Displacement (remainder of the slight straightening of the camera on D-306-B, D-304-B, D-303-B, D-302-B/PR_BG). `ranges.py --k 5`, corr_x and corr_y: over its own range, **no point of any sensor** beyond 5 noises (0 ranges on the 29 sensors). Against the level before (11th 06:00–14:00, `--outside` the label): corr_y none; corr_x a single isolated point of D-304-B at 13th 00:30, +5.4 noises, which is a one-step spike, not the vertical movement of the label. Against the level after (13th 02:00–14th 12:00): none. The step of D-306-B in corr_y on 2 h medians (from about 0.03 to 0.064 px, noise 0.0099–0.0122 px) is about 3 noises; D-304-B and D-303-B are smaller. It is a camera movement that the compensation absorbs almost entirely: under the threshold, ordinary noise of the station.
**Kept near the threshold** (stated so the reader knows the margin):
- CLARIDEN-50 (11th 09:00): D-304-A corr_y −5.3 noises (one measurement, `--rolling 6h`); D-301-A −6.7 and D-304-A −7.4 for the jump of 08:50→09:00.
- CLARIDEN-54 (18th 06:40): D-302_A and D-304-A corr_x +5.4 noises each, at 06:40.
- CLARIDEN-49 (return of row C): over its own range the median amplitude of the return is under 5 noises (12 h medians: D-301-C corr_x −0.043 → −0.085 px, about 3.6 noises; D-304-C corr_y −0.454 → −0.496, about 3.8; D-303-C −0.119 → −0.147, about 3); against its level after the return (16th 12:00–17th 16:00, `--outside` the label), D-301-C corr_x stays beyond 5 noises for 14 measurements on the 11th 12:00–18:10, extreme +6.1. It holds by the rule, at the margin; medium confidence.
**Before the requalification** (unchanged):
- One-measurement group flips of 5–7 targets at 3–4 noises (group_flips): 12th 01:00, 13th 21:30, 15th 09:00, 15th 13:30, 17th 03:40, 18th 06:00, 18th 09:50 (just after CLARIDEN-18), 19th 13:00. None reaches 5 noises on any target. They are the ordinary jitter of the pose (the rotation scatters by about ±0.002° from one measurement to the next). The 19th 04:20 instant (8 targets, max 4.1) falls inside the tilt label.
- Isolated single-sensor spikes at ≈5 noises (P-299-B corr_x 15th 03:20, D-300-C corr_y 11th 02:00, P-299-A corr_y 15th 05:10, PR-284 corr_x 18th 11:20): on noisy targets (noise 0.01–0.03 px), single points within what each sensor does elsewhere.
- The housing temperature dip of the 18th 08:00–10:00 is inside CLARIDEN-18. On its own, it moves only the pose already labeled.
## 3. The documentation
- Helpful: `families.md` "The camera tilts little by little" and `reading-rules.md` §5 ("A drift specific to a few targets that follows the camera, on the same dates, is the scene deforming with it"). Also labeling-rules §2 ("bounds are measurement instants", "a plateau is not a return", "a lasting jump is bounded on the jump") and §3 ("A deviation specific to targets, mixed with a camera movement, is the same event if it follows it in both directions").
- The precedents BREITHORN-2 and BREITHORN-11 gave the family and the wording.
- Reading rule "A jump is judged against the neighbours / the past of the target" was decisive for CLARIDEN-22: most of its jumps are returns.
- The validator did not flag the off-cadence bounds (22:45, 13:45, 10:15, 20:45, 00:15, 00:45): perhaps it could check against the station cadence.
- Contradicted: CLARIDEN-14 and CLARIDEN-15 said "No camera movement", but the camera tilt does move by a few thousandths of a degree over days. This is small next to "a few hundredths" (DATA.md orders of magnitude), but enough to leave remainders of several noises on some targets. `stations.md` lists four references for CLARIDEN, while PR-296 is a fifth global reference retained at every instant.
- The catalogue entry "The camera drifts away when the temperature of its housing drops, without a daily cycle" was tempting. The test "comes back at equal temperature" rejected it here.
## Requalification
Rule applied: a movement of the corr is kept if at least one target deviates by more than 5 of its own noises over the label's range (`tools/scripts/ranges.py --k 5`); for a slow drift the level is also taken outside the label (`--outside`, before and/or after), since its own median sits in the middle of the drift; No data is kept only when no sensor is measured. Outputs in `work/rq_*.txt`.
| id | bounds | family | command (all `python3 tools/scripts/ranges.py data/CLARIDEN --columns corr_x,corr_y`) | value | verdict |
|---|---|---|---|---|---|
| CLARIDEN-49 | 11th 06:00 → 16th 12:00 | Displacement | `--start 2026-10-11T06:00Z --end 2026-10-17T16:00Z --outside 2026-10-11T06:00Z/2026-10-16T12:00Z` | D-301-C corr_x, 14 measurements beyond 5, max +6.1 (11th 13:00); own window: one point 5.4 | kept, **rewritten** (it named CLARIDEN-51 "labeled separately"); loses nothing (it had no review) |
| CLARIDEN-50 | 11th 08:50 → 09:10 | Displacement | `--rolling 6h --start 2026-10-11T00:00Z --end 2026-10-11T20:00Z` | D-304-A corr_y −5.3 at 09:00 | kept |
| CLARIDEN-51 | 11th 14:00 → 13th 02:00 | Displacement | own range; `--outside` before (11th 06:00–14:00) and after (13th 02:00–14th 12:00) | no target beyond 5 in corr_y; one isolated corr_x point (D-304-B 5.4); about 3 noises on medians | **removed** |
| CLARIDEN-52 | 15th 22:50 → 16th 10:30 | No data | `inventory.py --start/--end` | 69 steps absent for all sensors | kept |
| CLARIDEN-17 | 16th 14:20 → 14:40 | No data | idem | 1 step absent for all sensors | kept (review 2026-11-17) |
| CLARIDEN-53 | 17th 16:00 → 20th 00:30 | Displacement | `--start 2026-10-16T12:00Z --end 2026-10-20T04:00Z --outside 2026-10-17T16:00Z/2026-10-20T00:30Z` | D-303-B corr_y runs of 3–7 measurements beyond 5 through the 19th, max −7.4 (19th 12:00); D-304-C −6.8, D-303-C −6.1, D-302-B +5.8 | kept |
| CLARIDEN-54 | 18th 06:30 → 06:50 | Displacement | `--rolling 6h --start 2026-10-18T00:00Z --end 2026-10-18T13:50Z` | D-302_A and D-304-A corr_x +5.4 at 06:40 | kept |
| CLARIDEN-55 | 18th 08:50 → 09:30 | Visibility | idem | D-302-B/PR_BG corr_y +11.8, D-301-C corr_x −8.1, P-301-B corr_x +6.8 | kept |
| CLARIDEN-56 | 18th 13:50 → 23:30 | No data | inventory | 57 steps absent for all sensors | kept |
| CLARIDEN-57 | 19th 09:00 → 10:10 | No data | inventory | 4 steps absent for all sensors, between single measurements | kept |
| CLARIDEN-58 | 19th 20:40 → 21:30 | No data | inventory | 3 steps absent for all sensors | kept |
| CLARIDEN-59 | 20th 00:10 → 00:40 | Displacement | `--start 2026-10-19T18:00Z --end 2026-10-20T04:00Z --outside 2026-10-20T00:10Z/2026-10-20T04:00Z` | D-306-B corr_y 18 measurements beyond 5, max +7.4 | kept |
No label of the range was a missing-points label or a thermal label. The validator reports 0 errors and 3 overlaps (two short events inside the drifts 49 and 53, and the jump 59 at the end of 53): distinct events, kept.
## Difficulties
- The instructions file changed during the work (the weather moved to `data/<STATION>/weather/`); neither source has CLARIDEN.
- The regression part of `scripts/tilt_vs_targets.py` printed nothing: the timestamps of the displacement and camera files differ by seconds. The 2 h tables were enough.
## 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`, `labeling/auto/CLARIDEN.jsonl` (labels 10–25 at the edges), `work/existing.jsonl`, `work/spanning.jsonl`, precedents (WILDSTRUBEL-83/102/71/47/55/200, BREITHORN-2/11, SNCF01-7, NEBRASKA-1). Tools: inventory, overview (range; 16th–20th with --others and --rigid; 19th–20th and 18th at 10 min; 11th and 18th with --rigid), summary, plot, metrics, ranges (--rolling 12), thumbnails (zone); my scripts `scripts/tilt_vs_targets.py`, `scripts/group_flips.py`, `scripts/fig_tilt.py`. Images: 18th 08:00 and 09:00 whole; zone thumbnails 11th 09:00, 13th 12:00, 18th 07:00, 19th 04:00, 20th 00:00, 20th 01:00.
Requalification: `instructions/analyst.md`, `labeling/labeling-rules.md` (in full), `work/requalify_input.jsonl`, this analysis and `report.md` before rewriting; tools `inventory.py`, `ranges.py`, `summary.py` (12 h medians of row C).