# Labels: BREITHORN, 2026-10-03T00:00Z → 2026-10-13T00:00Z (UTC) Task: requalify the 12 labels of this analysis under the new rules (`labeling-rules.md` §1): only what stands out from the noise of the corrected curves is labelled, by more than 5 of the sensor's own noises. Not a new reading: the bounds of the kept labels do not move, and no new event was looked for. The verification analysis before this one is kept in git; its reading (cadence 4 h, 61 instants, old camera without image-quality metrics, vehicles at the bottom of the scene, the camera's cold-morning signature) still holds. No weather: `data/BREITHORN/weather/` does not exist (no Swiss position for the station). The station is indoors, in a factory (`stations.md`). Distances and mm per px (`inventory.py`): COLJ19 8.9 m (3.27 mm/px), POUI19 9.4 m (3.55), COLI20, POUI20, COLH20, COLH20''' 20–23 m (7.6–8.6), COLI21, POUH21 30 m (11.35), COLI22, POUI22 40 m (15.0–15.2), COLI24, POUI24 60 m (22.5–22.8). The four references have no mm factor: their corr are in px only. Result: 7 kept as they are, 1 rewritten (BREITHORN-1), 4 removed (BREITHORN-28, -31, -4, -8). ## 1. My labels ### 2026-10-03T00:00Z → 2026-10-04T20:00Z (BREITHORN-1, rewritten) - **Phenomenon of the catalogue**: a reference hidden by an obstacle; the compensation runs on the other references. - **Precedents**: none new. The label was reviewed, and its substance holds. - **Family**: Obstruction. The cause is a parked vehicle in front of the bottom of the scene (images 03–04 Oct). - **Measured**: REFPOU22, COLH20''' and POUI22 are rejected from 03T04 to 04T16. Under the old description, these missing curves were the label. They are no longer labelled for themselves. What remains is the effect on the other curves. `ranges.py --k 5 --outside 2026-10-03T00:00Z/2026-10-04T20:00Z` (3 → 7 Oct) gives: - the corr_x of the reference REFH22 is out of its level at every measurement from 03T04 to 04T12, at −6.3 to −10.9 noises (−0.032 to −0.055 px; level +0.009, noise 0.005), except 04T08 (−0.013 px, under 5); - it is back at 04T20 (+0.004); - no target passes 5 noises in that window, apart from 04T16 (BREITHORN-3). The rolling run (±6 measurements) does not show it. That window spans 48 h and sits mostly inside the 40 h plateau, so it hides the shift. The level measured outside the episode is the right one here (GUIDE, ranges `--outside`). - **Images**: from the verification (04T08: the vehicle at the bottom of the scene, the three spots hidden). - **Confidence**: medium. The shift is clear and lasts the whole parking. But it is on a reference curve, not on a target, and no target passes the threshold. I kept the label because the rule keeps an episode of missing points "that moves the other curves beyond the noise". I rewrote the description to rest on that effect. It loses its review field. ### Kept as they are (same bounds, family, description and review field) | id | bounds | family | deviation that keeps it | |---|---|---|---| | BREITHORN-3 | 04T12 → 04T20 | Obstruction | 04T16: every sensor out of level. REFH22 corr_y +44 noises, COLI22 corr_y −26 (−3.5 mm), POUH21 −14.5 (−1.3 mm), the others −8 to −12 (rolling ±6) | | BREITHORN-29 | 05T00 → 05T12 | Thermal displacement | against the night level (evening of 4 Oct): REFPOU22 corr_y +6.2 noises at 05T04, REFG44N −5.4, POUI19 corr_x +5.3, COLH20 corr_y +5.0 (+0.062 px, +0.5 mm) at 05T08; rolling: REFPOU22 +5.5, REFG44N −5.3 | | BREITHORN-30 | 05T20 → 06T12 | Thermal displacement | against the night level (evening of 5 Oct): REFPOU22 corr_y +5.8 noises at 06T08 (rolling 5.0). The targets reach 4.6 at most (POUI24 corr_x) | | BREITHORN-5 | 07T08 → 07T16 | Obstruction | 07T12: COLJ19 corr_y +9.4 noises (+0.38 mm), POUI19 +9.2 (+0.36 mm), COLI22 corr_x +8.4 (+0.9 mm), POUI24 corr_y +7.3 (+1.2 mm), COLI20 +6.5 | | BREITHORN-6 | 07T16 → 08T04 | Obstruction | POUI22 corr_y +18.4 noises at 07T20 (+3.9 mm), POUI24 corr_y +59 at 08T00 (+9.6 mm) | | BREITHORN-7 | 08T20 → 09T08 | Obstruction | POUI24 corr_y +40.5 / +51.6 noises at 09T00 / 09T04 (+6.6 / +8.4 mm), corr_x +14.5 (+2.9 mm) | | BREITHORN-9 | 10T00 → 10T08 | Obstruction | 10T04: REFH22 corr_y +70 noises, POUH21 −14 (−1.3 mm), COLH20''' −14 (−1.0 mm), COLI20 −11.6, the others −8 to −10 | BREITHORN-30 is the closest to the threshold. It holds on a reference curve (REFPOU22) against the night level, which is the measure the rule gives for a thermal label. Confidence medium. ## 2. What I left without a label Removed labels: - **BREITHORN-28**, 2026-10-04T00:00Z → 2026-10-04T12:00Z, Thermal displacement. **Under the threshold**. Against the night level (evening of 3 Oct), the largest deviation is 4.1 noises (REFH22 corr_x at 04T08), then COLH20 corr_y 4.0 (+0.049 px, +0.4 mm, 04T04). The rolling run gives nothing over 5 noises in the window. The camera rotation deviated on that cold morning, but the compensation absorbed most of it. - **BREITHORN-31**, 2026-10-11T20:00Z → 2026-10-12T16:00Z, Thermal displacement. **Under the threshold against the night level**. The largest deviation is 4.9 noises (POUI19 corr_y −0.055 px, −0.19 mm, 12T08), then POUI24 corr_y 4.4. REFPOU22 and COLH20 are near 3. The rolling run puts POUI24 at −5.9 (12T04) and POUI19 at −5.2 (12T08). That comes from the later neighbours, not from the night: POUI24 rises afterwards from −0.04 to −0.009 px, which pulls its rolling median up. Against its evening level, POUI24 is −3.8 noises at 12T04. The previous analysis already called it the weakest case. - **BREITHORN-4**, 2026-10-06T16:00Z → 2026-10-07T08:00Z, Obstruction. **Missing points only**. COLI22, COLI24 and POUI22 are rejected at night. No sensor passes 5 noises between 06T16 and 07T08, and the references hold. - **BREITHORN-8**, 2026-10-09T16:00Z → 2026-10-10T00:00Z, Obstruction. **Missing points only**. COLI22, COLI24 and POUI22 are rejected at 09T20. No sensor passes 5 noises there. Unchanged from the verification, still not labelled: the morning of 3 Oct and the morning of 11 Oct (thermal signature under 5 noises), and the identical camera pose at 12T16 and 12T20. The slow drift of corr_x is BREITHORN-2, a spanning label outside this batch. ## 3. The documentation - `labeling-rules.md` §1 sets the measure: ranges at k 5, rolling under a drift or a thermal cycle, the night level for a thermal label. On a 4 h cadence the two measures can disagree. The rolling window (±6 measurements, ±24 h) mixes days and later drifts. Against the night level, BREITHORN-31 falls to 4.9 but BREITHORN-30 rises to 5.8. For the thermal labels I used the night level, as the rule says. - The rolling level hides a plateau longer than its window (BREITHORN-1). `--outside` around the episode shows it. The GUIDE says so, and it decided this case. - The rule says "at least one target". It does not say whether the corr of a reference counts. The tools treat the references like the targets, and the reviewed BREITHORN thermal labels rest on REFPOU22. I counted them. Two labels hold only on a reference curve (BREITHORN-1 on REFH22, BREITHORN-30 on REFPOU22). If the expert counts only the targets, these two would be removed. A sentence in §1 would settle it. ## Requalification Commands (from the repository root, output in `work/`): - `python3 tools/scripts/ranges.py data/BREITHORN --columns corr_x,corr_y --k 5 --rolling 6 --start 2026-10-03T00:00Z --end 2026-10-13T00:00Z` → `work/requalify_ranges_roll6.txt`; the same on `corr_x_mm,corr_y_mm` → `work/requalify_ranges_roll6_mm.txt`. - `python3 tools/scripts/ranges.py data/BREITHORN --columns corr_x,corr_y --k 5 --outside 2026-10-03T00:00Z/2026-10-04T20:00Z --start 2026-10-03T00:00Z --end 2026-10-07T00:00Z` → `work/requalify_ranges_B1_outside.txt`. - `python3 scripts/requalify_thermal.py data/BREITHORN work/requalify_ranges_roll6.txt work/requalify_input.jsonl` → `work/requalify_thermal.txt`. For each thermal label and each sensor, it gives the largest deviation from the median of the previous evening (12, 16, 20 UTC), in the sensor's noise from ranges.py. - `summary.py --column corr_x|corr_y --step 4h` on 2–5 Oct (REFH22) and 11–13 Oct (POUI24, POUI19). | id | family | verdict | max deviation (noises) | |---|---|---|---| | BREITHORN-1 | Obstruction | rewritten | REFH22 corr_x −10.9 (outside level), 9 measurements over 5 | | BREITHORN-28 | Thermal displacement | removed | 4.1 (night level) | | BREITHORN-3 | Obstruction | kept | REFH22 corr_y +44, COLI22 −26 | | BREITHORN-29 | Thermal displacement | kept | REFPOU22 corr_y +6.2 (night level) | | BREITHORN-30 | Thermal displacement | kept | REFPOU22 corr_y +5.8 (night level) | | BREITHORN-4 | Obstruction | removed | missing points only, nothing over 5 | | BREITHORN-5 | Obstruction | kept | COLJ19 corr_y +9.4 | | BREITHORN-6 | Obstruction | kept | POUI24 corr_y +59 | | BREITHORN-7 | Obstruction | kept | POUI24 corr_y +51.6 | | BREITHORN-8 | Obstruction | removed | missing points only, nothing over 5 | | BREITHORN-9 | Obstruction | kept | REFH22 corr_y +70 | | BREITHORN-31 | Thermal displacement | removed | 4.9 (night level); rolling 5.9 from the neighbours | `validate.py`: 0 errors. One overlap remains (BREITHORN-1 and BREITHORN-3). It was already there, it was reviewed, and the bounds are not to be moved. ## Difficulties - The data now start on 1 Sep (pumped since). BREITHORN-1 starts at 03T00 as the batch edge. REFH22 corr_x already drops a little between 02T20 and 03T00, but the bounds of a kept label are not to be moved. ## Sources read `instructions/analyst.md`, `labeling/labeling-rules.md` (§1–3), `labeling/families.md` (via grep), `tools/GUIDE.md` (ranges, conventions), `tools/DATA.md` (columns), the previous `analysis.md`, `report.md`, `work/requalify_input.jsonl`, `work/ranges_roll.txt`, `scripts/thermal_table.py`. Data: `inventory.py`, `ranges.py` (above), `summary.py`, `scripts/requalify_thermal.py`.