Files
aimsight-labeling/auto/analyses/MICHIGAN_2026-10-23T2000_2026-11-01T0000/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

85 lines
9.6 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: MICHIGAN, 2026-10-23T20:00Z → 2026-11-01T00:00Z (UTC)
Request: requalify the 7 labels of this analysis (`work/requalify_input.jsonl`) under the new rule (`labeling/labeling-rules.md` section 1): only what stands out from the noise of the corrected curves, at least one target beyond 5 of its own noises. This is a requalification, not a new reading; the previous analysis is in `work/analysis.before-requalify2.md`. DETECTORS: no.
Station context (unchanged from the first analysis): indoors, same building as BREITHORN and NEBRASKA; one measurement every 4 h (50 instants, none missing); REFPOU22 rejected at every measurement, so the compensation runs on REF42 and REFG43N alone; no image-quality metrics; no `weather/` folder for MICHIGAN, so no MeteoSwiss station or distance to give (indoors anyway). Targets from 9 m (COLH18', POUH18', 3.6 mm/px) to 51 m (REFH22', 20.2 mm/px); the far targets (COLH21, COLH21', COLH21'', POUH21'' near 41 m, REFH22') carry the most mm.
Result: **4 kept as they are** (MICHIGAN-11, -13, -14, -15, with their review field), **0 rewritten**, **3 removed** (MICHIGAN-20, -12, -21).
## 1. My labels
### 2026-10-25T16:00Z → 2026-10-26T16:00Z (MICHIGAN-11, kept as is)
- **Phenomenon of the catalogue**: The camera drifts away when the temperature of its housing drops, without a daily cycle; A compensation on two references.
- **Precedents**: MICHIGAN-9 (validated, same signature on the same station).
- **Family**: Thermal displacement: the pose and the curves follow the housing temperature and come back with it.
- **Measured**: at 26T08, 7 targets beyond 5 noises together in corr_x (COLH20'' −7, COLH21' −7, COLH21 −6, REFH22' −6, COLH20' −5, COLH21'' −5, POUH21'' −5) with the level rolling over ±1 day and the noise computed outside the 28–30 Oct cold episode (`--outside 2026-10-28T00:00Z/2026-10-30T16:00Z`); with the default noise (computed over the whole batch, episode and 29T12 step included) the largest is REFH22' −4.7. I take the first: the 28–30 Oct episode and the step of 29T12 are themselves labelled anomalies and inflate the noise of corr_x (COLH21' noise 0.020 px outside, 0.033 px over the whole batch); the guide gives `--outside` for this case. corr_y: largest −3.8 noises (POUH18', 26T04), under the threshold. In mm: COLH21' about 5.0 → 2.6 mm, REFH22' about 5.0 → 1.8 mm in x; near targets under 0.3 mm.
- **Images**: 26T04 (first analysis): nothing in front.
- **Confidence**: medium: beyond the threshold only with the noise read outside the cold episode; the pose, the housing temperature and the 7 targets together make it coherent.
### 2026-10-28T00:00Z → 2026-10-28T16:00Z (MICHIGAN-13, kept as is)
- **Phenomenon of the catalogue**: idem.
- **Precedents**: MICHIGAN-9.
- **Family**: Thermal displacement.
- **Measured**: 28T08: 9 targets beyond 5 noises in corr_x with the default settings and `--rolling 1d` (REFH22' −7, COLH21' −6, COLH21 −6, POUH21'' −6, COLH20' −6, POUH20' −6…); 12 targets, up to −10 (COLH21'), with the noise outside the episode. corr_y under the threshold (−3.3). In mm: far targets about −4 to −5 mm in x (REFH22' 5.3 → 0.1 mm, COLH21 4.5 → −0.6 mm).
- **Images**: 28T08: nothing in front.
- **Confidence**: high.
### 2026-10-28T16:00Z → 2026-10-29T16:00Z (MICHIGAN-14, kept as is)
- **Phenomenon of the catalogue**: idem.
- **Precedents**: MICHIGAN-9.
- **Family**: Thermal displacement.
- **Measured**: corr_x beyond 5 noises at 29T00, 29T04, 29T08 on up to 14 targets with the batch level (REFH22' −11.9, COLH21' −10.4, COLH20' −10.1 at 29T08); with `--rolling 1d`, 10 targets at 29T08, REFH22' −8.4. corr_y under the threshold (−4.6 batch level, −3.0 rolling). In mm: REFH22' 5.3 → −4.4 mm, COLH21 about −10 mm from its level before; near targets about −1 mm.
- **Images**: 29T08: nothing in front.
- **Confidence**: high.
### 2026-10-29T16:00Z → 2026-10-30T12:00Z (MICHIGAN-15, kept as is)
- **Phenomenon of the catalogue**: idem.
- **Precedents**: MICHIGAN-14.
- **Family**: Thermal displacement.
- **Measured**: corr_x beyond 5 noises at 30T00, 30T04, 30T08 (batch level: REFH22' −11.6, COLH20' −10.4, COLH21' −10.1 at 30T04); `--rolling 1d`: 11 targets at 30T04, REFH22' −9.4. corr_y −4.7 at most. In mm: REFH22' about −4.2 mm, COLH21 about −5.2 mm at 30T04, from a level near +2 to +3 mm.
- **Images**: 30T04: nothing in front.
- **Confidence**: high.
## 2. What I left without a label
Removed labels (not submitted again):
- **MICHIGAN-20**, 2026-10-23T20:00Z → 2026-10-25T00:00Z, Thermal displacement: largest deviation inside the label 2.7 noises (COLH20' corr_x, 25T00) with `--rolling 1d`, 3.6 with the batch level (REFH22' corr_x), 4.0 with the noise outside the cold episode (COLH21' corr_x, 25T00); corr_y 2.5 to 3.1. The camera does move with the housing temperature, but the corrected curves stay within the noise: a camera movement largely absorbed by the compensation. In mm: about 1 to 1.5 mm on the far targets.
- **MICHIGAN-12**, 2026-10-26T20:00Z → 2026-10-27T20:00Z, Displacement: corr_y rise of all the targets at 27T04, largest 3.2 noises (COLH18S''') with `--rolling 1d`, 3.3 with the noise outside the cold episode; with the batch level 5.0 exactly (COLH18S'''), not beyond, and that level is pulled down by the low corr_y plateau of the end of the month (spanning MICHIGAN-7); ranges.py at `--k 5` gives no corr_y range over the whole batch. corr_x under 2 noises. In mm: about +2 mm on the far targets. It was validated on 2026-11-17; it falls under the new threshold.
- **MICHIGAN-21**, 2026-10-31T00:00Z → 2026-10-31T16:00Z, Thermal displacement: under 2 noises in both columns with `--rolling 1d` (and with the noise outside the cold episode); 2.7 noises with the batch level (COLH20' corr_x, 31T12). A small camera nod, absorbed: about 1 mm.
Unchanged from the first analysis: 25T12 (a single measurement, pose wobble), 01T00 (start of the cold event of 1 November, next range), the spanning MICHIGAN-3 (REFPOU22 rejected) and MICHIGAN-7 (slow camera tilt) not touched by this requalification.
## Requalification
Commands (from the repository root, outputs in `work/requalify/`):
- `python3 tools/scripts/ranges.py data/MICHIGAN --columns corr_x,corr_y --k 5 --start 2026-10-23T20:00Z --end 2026-11-01T00:00Z` (`batch_k5.txt`): corr_x ranges only on 28T20–30T08 instants (29T00/04/08, 30T00/04/08), up to −11.9 noises (REFH22', 29T08); corr_y: none.
- same with `--rolling 1d --instants 2` (`batch_k5_roll1d.txt`): 28T08 (9 targets), 29T04 (9), 29T08 (10), 30T04 (11), all in corr_x; corr_y: none.
- same with `--rolling 1d --outside 2026-10-28T00:00Z/2026-10-30T16:00Z --instants 2` (`batch_k5_roll1d_outside.txt`): adds 26T08 (7 targets, up to −7) and more targets on 28–30 Oct (up to −13, COLH21', 30T04).
- `--k 2` runs (batch level, `--rolling 1d`, `--rolling 3`, `--rolling 1d --outside …`) and `scripts/requalify_max.py` for the largest deviation inside each label (`max_per_label.txt`):
| id | bounds | batch level x / y | rolling 1d x / y | rolling 1d, noise outside 28–30 Oct x / y | verdict |
|---|---|---|---|---|---|
| MICHIGAN-20 | 23T20 → 25T00 | 3.6 / 3.1 | 2.7 / 2.5 | 4.0 / 2.5 | removed |
| MICHIGAN-11 | 25T16 → 26T16 | 4.5 / 2.2 | 4.7 / 3.6 | **7.1** / 3.8 | kept |
| MICHIGAN-12 | 26T20 → 27T20 | 4.0 / 5.0 | <2 / 3.2 | <2 / 3.3 | removed |
| MICHIGAN-13 | 28T00 → 28T16 | 4.6 / 2.1 | **6.7** / 3.2 | **10.1** / 3.2 | kept |
| MICHIGAN-14 | 28T16 → 29T16 | **11.9** / 4.6 | **8.4** / 3.0 | **11.8** / 3.3 | kept |
| MICHIGAN-15 | 29T16 → 30T12 | **11.6** / 4.7 | **9.4** / 3.9 | **13.2** / 4.0 | kept |
| MICHIGAN-21 | 31T00 → 31T16 | 2.7 / <2 | <2 / <2 | <2 / <2 | removed |
(absolute values, in noises; the batch level is skewed by the slow drift of MICHIGAN-7, so the rolling columns decide.)
- `summary.py --column corr_x_mm --instants …` for the magnitudes in mm; `plot.py --columns corr_x_mm,corr_y_mm` (`curves_mm.png`).
- `python3 labeling/scripts/validate.py labels.jsonl`: 4 labels, 0 errors.
## 3. The documentation
- `labeling-rules.md` section 1 (the 5-noise threshold, also for the thermal) decided; the request's rule ("with `--rolling` under a drift or a thermal cycle") was needed: the slow tilt of the month (MICHIGAN-7) skews the batch level.
- What is not said: which noise to use when the batch contains a strong labelled episode. Here the noise of corr_x over the whole batch is inflated by the 28–30 Oct cold mornings and the 29T12 step, and MICHIGAN-11 lies on either side of the threshold depending on it (4.7 / 7.1). I used `--outside` on the episode, as `tools/GUIDE.md` (ranges) allows; a sentence in `labeling-rules.md` on this choice would remove the ambiguity.
- The first analysis noted that its labels 20 and 12 did not reach 5 noises from the night level and kept them after the validated precedents; the new rule settles it: removed. MICHIGAN-12 was validated by the user on 2026-11-17; it falls under the threshold on every variant.
- No precedent consulted for this requalification (the decisions are measurements, not classifications).
## Difficulties
- `plot.py` takes `--zone` once per zone, not several values after one flag (first call failed).
## Sources read
`instructions/analyst.md`, `work/requalify_input.jsonl`, the previous `analysis.md` and `report.md` (kept in `work/`), `labeling/labeling-rules.md` (sections 1–3), `tools/GUIDE.md` (ranges, plot), `work/inventory.txt` and `inventory.py` (distances, mm/px). Tools on `data/MICHIGAN`: `ranges.py` (k 5 and k 2, batch level, `--rolling 1d`, `--rolling 3`, `--outside`, `--instants 2`), `summary.py --column corr_x_mm --instants`, `plot.py`. Scripts: `scripts/requalify_max.py`.