11 KiB
Labels: NEBRASKA, 2026-10-22T00:00Z → 2026-11-01T00:00Z (UTC)
Task: verify the 4 existing labels of the range (work/existing.jsonl) and look for what they miss. DETECTORS: no. The station measures every 4 h (61 instants in the range, none missing, a pose at each). No weather/ folder for NEBRASKA: no MeteoSwiss weather is available for this station (indoors, in the same building as BREITHORN and MICHIGAN according to the expert). No image-quality columns (older camera): quality.py not applicable.
Verdict per existing label
| Old id | Verdict | Evidence |
|---|---|---|
| NEBRASKA-8 (26T00→26T20, Thermal) | kept | Housing temperature a little lower 26T02→26T14 (30–30.5 °C vs 31–31.5). Pose: tilt 0.0362 (00) → 0.0375 → 0.0399 → 0.0410 (12) → 0.0383 (16) → 0.0394 (20), rotation −0.0229 → −0.0283 (08) → −0.0243 (20); pan last at rest at 00:00. corr_y COLJ20 −0.090 (00) → −0.009 → +0.029 (08) → −0.029 (20); POUJ19 −0.016 → +0.093 (08) → +0.009 (20); POUJ20, COLJ19 the same; corr_x nearly flat. 26T16 still touched (POUJ19 +0.046), 26T20 back. Bounds, family, description hold. |
| NEBRASKA-9 (27T20→28T16, Thermal) | kept | Temperature 31.5 → 30.25–30.75 from 28T01 to 28T10. Pan 0.0009 (27T20) → 0.0066 (04), rotation −0.0256 → −0.0194 (04), tilt 0.0440 and scale 0.999456 at 08, all back at 16 (tilt 0.0387, scale 0.999577). corr_y down from 28T00: COLH19 +0.065 → −0.079 (04), POUH19 +0.068 → −0.018, POUJ20 +0.015 → −0.062, COLJ20 −0.016 → −0.125, COLH20 0.109 → 0.035; corr_x peak at 08 (COLJ20 +0.064, POUJ19 +0.058, POUH21 +0.057, COLJ19, POUH20 about +0.04). Back at 28T16. Holds. |
| NEBRASKA-10 (28T20→29T12, Thermal) | kept | Coldest night: 27.25 °C at 29T07. Tilt 0.0387 (28T16) → 0.0422 (20) → 0.0480 (04); pan last at rest at 28T20 (0.0026) then 0.0075 (00), 0.0081 (08): the later of the two axes rests until 28T20, so the start is right (labeling-rules 2, camera cycle start; one step of difference stays within the uncertainty anyway). corr_x up 29T00–08 (COLJ20 0.773/0.837/0.806, POUJ20 0.717 at 04); corr_y jumps up at 04 (COLJ20 −0.117 → +0.065, POUJ19 −0.061 → +0.113, POUH19, COLH19, COLH20), lower at 00 and 08. error_value at 04: REFH22 0.155 → 0.194, COLH19 0.163, COLH20 0.184, far below 0.7. 29T12 back at the new level of 29T16 (corr_x COLJ20 0.778/0.773, tilt 0.0421/0.0406). Holds. |
| NEBRASKA-11 (29T20→30T16, Thermal) | kept | Temperature 31.75 (29T19) → 27.75 (30T06). Pan 0.0037 → 0.0057 at 29T20, tilt last at rest at 29T20 (0.0416) then 0.0449, 0.0481 (04); scale 0.999397 at 04; back at 16. corr_y jumps up at 04 (COLJ20 −0.027 → +0.077, POUJ19 +0.027 → +0.106, POUH20 −0.013 → +0.065, POUJ20 −0.024 → +0.048), down at 08; corr_x higher 00–12. error_value at 04: REFH22 0.207, COLH20 0.216, COLH19 0.163. Back at 16. Holds. |
Spanning labels (not changed, not in labels.jsonl): NEBRASKA-2 (REFCOLH22 lost) holds in the range: REFCOLH22 rejected at all 61 instants, compensation on REFH22 and REFP''21 only (inventory). NEBRASKA-6 (long drift) holds in the range: corr_x of all targets up (COLJ20 0.67 → 0.78, POUJ20 0.58 → 0.62), corr_y of COLH19, POUH19, COLH20, COLJ19, POUJ20 down, tilt 0.036 → 0.042, slight backward movement on 31 October. Its sentence "the corr_y of POUH20 and COLH21 rises" is not visible inside this range (both flat within noise), but it describes the whole label since 16 October: not wrong.
1. My labels
2026-10-23T20:00Z → 2026-10-25T04:00Z (added)
- 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-20 (23T20→25T00, same building, not reviewed) and BREITHORN-17 (23T12→24T20): the neighbouring cameras have a thermal label on this same cooling; MICHIGAN-9, -11, -13 (reviewed) for the wording of a cool-morning thermal on a two-reference basis; NEBRASKA-8 (reviewed) as the weakest validated thermal of this station, against which this day is compared.
- Family: Thermal displacement: the pose follows the housing temperature and comes back with it (scale 0.999584 → 0.999514 → 0.999569), not a trend; no image nuisance, error_value at its level. Not Displacement: the return at equal temperature is there; the part that does not come back (corr_x) belongs to NEBRASKA-6.
- Measured: housing temperature (hourly) 31–31.5 °C until 23T18, 30.25–30.75 from 23T19, 29–29.5 from 24T05 to 24T23, back to 30.5–31 at 25T00–01. Pose: tilt 0.0340 (23T16) → 0.0360 → 0.0373 (24T00) → 0.0383 (08) … 0.0383 (20) → 0.0368 (25T00) → 0.0346 (25T04); pan 0.0024 (23T20), 0.0029 (00), 0.0045 (04), 0.0043–0.0045 (12–16), 0.0027 (25T00); scale 0.999556 → 0.999514 (04, 12, 16) → 0.999544 (25T00) → 0.999569 (25T04). Pan last at rest at 24T00, tilt at 23T20 within its 22–23 range: start 23T20.
overview.py --rigid 2026-10-23T16:00Z: the targets translate together (tx ≈ ty ≈ 0.10 px at 24T00–20, back to 0.025/0.011 at 25T04), their residual stays about 0.01–0.02 px (no deformation); REFP''21 departs from the targets' movement (dx −0.044 at 24T00, dy +0.05 at 08): the two-reference basis renders the movement badly. corr_y: COLJ23 −0.021 → −0.052/−0.050 (00/04), COLJ24 +0.028 → −0.013, POUH21 +0.039 → +0.007, COLH19 +0.034 → +0.010/−0.004, again at 12 (COLH19 +0.006, POUJ20 −0.039), back in the evening; corr_x COLH19 0.216 → 0.242 (04, 12), POUH21 0.063 → 0.08–0.10 then 0.072 at 25T04, COLJ24 a little up. error_value COLH19 0.125 → 0.133, COLH20 0.129 → 0.140 at 04 (tiny).days.py --bounds: scale −2.7 noises, tilt +1.6, pan +2.1 (the night levels are themselves inside cycles at this cadence; the validated 26 October shows tilt +4.8, scale −2.8, corr_y +2.2). - Images: thumbnails 23T16, 24T04, 24T16 of COLJ20, POUJ19, COLH19, COLJ23, REFH22, REFP''21: spots unchanged; whole image 24T04: nothing in front of the targets.
- Confidence: medium: pose and temperature clear and neighbours concordant; the effect on the corrected curves is small and mixed with the drift.
2026-10-31T00:00Z → 2026-10-31T16:00Z (added)
- Phenomenon of the catalogue: same as above.
- Precedents: MICHIGAN-21 (31T00→31T16, same building, not reviewed) on the same day; NEBRASKA-9 for the shape (corr_y down at 04 on most targets).
- Family: Thermal displacement (pose and corr return with the temperature at 16:00; no nuisance on the images; error_value flat).
- Measured: temperature 30.5–31.25 on 30T18–23, 29.75–30.5 from 31T00 to 31T11, 30.5 then 31–32 from 31T16. Tilt 0.0422 (00) → 0.0433 → 0.0437 (08) → 0.0432 → 0.0420 (16); pan 0.0030 → 0.0046 (04, 08) → 0.0031 (16); scale 0.999506 → 0.999465/0.999464 → 0.999495 (16). Rigid (base 31T00): targets translate 0.01–0.03 px, residual ≤ 0.01 px; REFH22 dy +0.028 at 04. corr_y at 04 vs 00: COLJ20 −0.032 → −0.106, POUJ20 −0.011 → −0.076, COLH19 −0.015 → −0.059, POUJ19 −0.003 → −0.059, POUH19 +0.013 → −0.040, COLJ23 −0.034 → −0.062; 08 and 12 still low; 16 back (COLJ20 −0.059, close to −0.067 at 20). corr_x at 04: COLH19 0.225 → 0.270, COLJ23 0.109 → 0.126, POUH20 0.120 → 0.138. error_value flat.
- Images: thumbnails 31T00, 04, 08: spots unchanged; whole image 31T08: background lights on below the targets (a lit opening of the building), away from the spots.
- Confidence: medium: the temperature drop is small (about half a degree); the signature and the neighbour say thermal.
2026-10-31T20:00Z → 2026-11-01T00:00Z (added)
- Phenomenon of the catalogue: same.
- Precedents: none useful beyond the above (no label after the range on the three stations yet).
- Family: Thermal displacement, as the cold nights before.
- Measured: temperature 31–32 (31T18–20) → 29.75 at 01T01; tilt 0.0419 → 0.0434, pan 0.0034 → 0.0043, scale 0.999497 → 0.999475 at 01T00, continuing at 01T04 (pan 0.0060, scale 0.999445). corr_y 01T00 vs 31T20: POUH21 −0.041, COLH21 −0.031, POUH19 −0.027, POUJ20 −0.026, COLJ20 −0.024, COLH20 −0.022; corr_x within ±0.015. Ends at the range edge (labeling-rules 2: no neighbouring range yet); to be extended.
- Images: whole image 01T00: nothing in front of the targets.
- Confidence: medium (a single measurement inside the range).
2. What I left without a label
- 22T12–22T16 and 23T12–23T16: corr_y of COLJ20, POUJ19, POUJ20 a little lower (2–3 MAD), tilt a little lower (0.0338–0.0349 vs 0.036), housing temperature a little higher (32–32.5 at 22T15–17 and 23T17). Within the noise of this 4-hourly series and well below the validated thermal days; part of the corr_y fall at 22T12 does not come back and belongs to the long drift NEBRASKA-6.
- 27T12: several corr_y up by 2–2.5 noises for one measurement (COLH20, COLH21, POUH21, POUJ20), pose almost still (rotation −0.0237 → −0.0250). One measurement, under what is labelled here; MICHIGAN-12 (Displacement, 26T20→27T20) is on the neighbour, NEBRASKA shows no comparable rise.
- 25T08–26T00, 27T00–27T16: calm, inside NEBRASKA-6.
- Exact repeated values between some instants (26T20 and 27T00, 30T16 and 31T16): raw positions are quantized to about a hundredth of a pixel, and with only two references the pose can repeat to the digit: not a data fault.
3. The documentation
families.md(thermal/displacement boundary, "camera drifts when the housing temperature drops") andlabeling-rules.md(section 1 "a signature found once is searched over the whole batch"; section 2 camera cycle start, "a deviation of a single rate step stays within the uncertainty") decided the verdicts: the 24 and 31 October cold episodes were the forgotten ones; the 28T20 / 29T20 starts of NEBRASKA-10/-11 are right by the "later of the two axes" rule.labeling-rules.md"No thermal label on a day when neither the pose nor the corr deviate by more than 5 noises": at a 4-hour cadencedays.pynever reaches 5 noises, not even on the validated 26 October (tilt +4.8); the night levels are themselves inside the cycles. I compared with the validated days instead. A note for 4-hourly stations would help.- Precedents: MICHIGAN-20 and -21 (same building, same dates) confirmed the two added episodes; I also read the BREITHORN and MICHIGAN labels of the range directly in their
labeling/auto/files (the same labelsprecedents.pyreturns). days.py --nightwith a 4 h cadence keeps only 00:00 and 20:00 per night: weak reference.
Difficulties
plot.py --zonetakes one range per flag (--zone A --zone B), not several after one flag.- The Write tool refused
.mdfiles in this sub-agent context; analysis.md and report.md were written with a shell heredoc.
Sources read
agents/agents/analyst.md, instructions/analyst.md, work/prompt.txt, work/existing.jsonl, work/spanning.jsonl, request.json, tools/GUIDE.md, tools/DATA.md, labeling/families.md, labeling/labeling-rules.md, labeling/reading-rules.md, labeling/stations.md, labeling/auto/NEBRASKA.jsonl, the BREITHORN and MICHIGAN labels overlapping the range, precedents.py (two queries). Tools: inventory.py, overview.py (window, --rigid on 23–25 and 30–01), metrics.py (4 h and 1 h), summary.py (corr_x, corr_y, error_value, x, y at 4 h), days.py --bounds, stations.py (NEBRASKA, BREITHORN, MICHIGAN), thumbnails.py, plot.py; whole images 24T04, 31T08, 01T00.