Files
aimsight-labeling/auto/analyses/NEBRASKA_2026-10-03T0000_2026-10-13T0000/analysis.md
T

72 lines
11 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: NEBRASKA, 2026-10-03T00:00Z → 2026-10-13T00:00Z (UTC)
Request: verify the existing labels of the range (`work/existing.jsonl`: NEBRASKA-1), check the spanning one (`work/spanning.jsonl`: NEBRASKA-2) inside the range, look for what they miss. DETECTORS: no.
## Verdict per old id
- **NEBRASKA-1** (Displacement, 03/10 00:00 → 12/10 12:00): **kept**. The slow corr_x rise of COLJ20, POUJ20, POUJ19, COLJ19, then POUH19 and COLH19 is there (COLJ20 about 0.21 → 0.30 before the 9 October jump, 0.41 → 0.47 after; POUJ20 0.19 → 0.29, then 0.37 → 0.42; COLJ23, COLJ24, POUH21 flat apart from the jump). It reaches a plateau from 12/10 12:00 (COLJ20 0.473, 0.457, 0.475, 0.475, 0.477 up to 13/10 04:00; recedes from 13/10 08:00, the label NEBRASKA-3). Tilt 0.0129 → 0.0297 deg (peak 12/10 04:00), rotation −0.0048 → −0.0201, scale 0.99988 → 0.99968, clear step 07/10 12:00 → 08/10 00:00 (tilt 0.0178 → 0.0224) without a step in corr_x. Housing temperature 34.6 → 33.0-33.5 °C over the phase. Family, bounds and wording consistent with NEBRASKA-6 and BREITHORN-2, MICHIGAN-19.
- **NEBRASKA-2** (spanning, Target issue, 09/10 08:00 → 01/11): **holds within the range** (not modified). REFCOLH22 rejected at all 22 measurements from 09/10 12:00 to 13/10 00:00 (error_value up by about 0.96); fixed-frame thumbnails: spot sharp at 09/10 08:00, gone at 09/10 12:00 and 12/10 12:00, REFH22 intact. Jump at 12:00: corr_x COLJ20 +0.13, POUJ19 +0.115, POUJ20 +0.08, POUH19 +0.08; corr_y COLH19 +0.084, COLJ23 +0.058, COLH20 +0.055, COLJ24 +0.046, as written. corr of REFH22 and REFP''21 at 0 (1e-13) since then.
## 1. My labels
### 2026-10-03T00:00Z → 2026-10-12T12:00Z (NEBRASKA-1, kept as it is)
- **Phenomenon of the catalogue**: The camera tilts little by little (and, after 09/10, A compensation on two references)
- **Precedents**: NEBRASKA-6 and NEBRASKA-3 (same station, reviewed: the same movement later in the month, wording of the scene deforming with the tilt); BREITHORN-2 (reviewed), MICHIGAN-19 (not reviewed): the same tilt on the same dates in the same building.
- **Family**: Displacement: a drift of several days that the housing temperature does not explain alone (it drops by about one degree while the tilt keeps rising, no return at equal temperature); not Thermal displacement.
- **Measured**: see the verdict. `overview.py --rigid 03/10 00:00`: overall translation ty up to 0.56 px, rotation −0.013 deg; residual median of targets 0.007 → 0.05 px, COLJ20 dx up to +0.09, POUJ20 dx up to +0.12 px by 12/10, growing with the tilt (steps on 08/10 with the camera step): the scene deforms with the camera on the same dates (reading rule NEBRASKA-6/5), not a separate event. corr_y also drifts a little over the phase (night levels 03/10 → 08/10: COLH20 +0.044, COLH19 −0.044, COLJ20 −0.063, about 4-5 noises; POUH20 rises after 09/10): the same deformation; the label names the corr_x, which carries most of it. A small bump of the corr_y of the top targets 07/10 12:00 → 08/10 08:00 (2-4 noises per target) at the camera step: within this event.
- **Images**: whole images 04/10, 05/10, 06/10 08:00; thumbnails 05-06/10, 09/10, 12/10: nothing in front of the targets.
- **Confidence**: high for the drift and its bounds; the cause stays as written (not established).
### 2026-10-04T00:00Z → 2026-10-04T16:00Z (added)
- **Phenomenon of the catalogue**: The camera moves through a thermal effect
- **Precedents**: NEBRASKA-7, NEBRASKA-8, NEBRASKA-9 (reviewed): cool mornings, pose nodding forward at the housing temperature trough, top targets' corr out and back; BREITHORN-28 (not reviewed): same morning on BREITHORN.
- **Family**: Thermal displacement: the pose follows the housing temperature and comes back with it the same day; not Displacement (it returns), not Visibility (error_value flat, image clean).
- **Measured**: temperature 34.0 (00:00), 33.6 (04:00), 33.75 (08:00), 34.75 (12:00). `days.py --night 16-21 --bounds`: tilt +6.3 noises at 08:00 (0.0134 → 0.0174 deg), rotation −5.1 noises, scale −3.7; back at 12:00. corr_x COLJ20 0.219 → 0.259 → 0.273 → 0.242 → 0.241; POUJ20 0.215 → 0.243 → 0.257 → 0.229 → 0.207; POUJ19 0.144 → 0.177 → 0.176 → 0.164 → 0.154; POUH19 0.087 → 0.109 → 0.117 → 0.101 → 0.095; corr_y POUJ20 0.055 → 0.073 → 0.090 → 0.036, POUH19 0.069 → 0.093 → 0.086 → 0.069, COLJ23 −0.033 at 08:00. Rigid residual 08:00: tgt_med 0.021, POUJ20 dx +0.027, COLJ20 dx +0.023. 12:00 halfway (POUJ20 corr_x), 16:00 back.
- **Images**: whole image 08:00: dark indoor scene, all spots present, nothing in front.
- **Confidence**: medium: small amplitude (a few noises per target) but pose above 5 noises and the signature of the reviewed cool mornings.
### 2026-10-04T20:00Z → 2026-10-05T16:00Z (added)
- **Phenomenon of the catalogue**: The camera moves through a thermal effect
- **Precedents**: NEBRASKA-7, -8 (reviewed); MICHIGAN-17 and BREITHORN-29 (not reviewed), same morning.
- **Family**: Thermal displacement, as above.
- **Measured**: temperature 34.6 (04/10 20:00), 33.9 (00:00), 33.4 (04:00), 33.75 (08:00), 34.6 (12:00). Tilt 0.0135 → 0.0148 → 0.0152 → 0.0168 → 0.0156 → 0.0144 (+5.0 noises at 08:00); scale 0.999882 → 0.999828 at 08:00; rotation −0.0072 at 00:00, −0.0078 at 08:00. corr_y at 08:00: POUJ20 0.129 (night 0.03), POUH19 0.134, COLJ19 0.120, POUJ19 0.069, COLJ20 0.050, COLH20 0.096; median of targets +5.4 noises. corr_x at 00:00 COLJ19 0.136 (vs 0.105), at 08:00 COLJ19 0.164, POUH19 0.129, COLH21 +0.014 (vs −0.01). 12:00: COLJ19 0.151, POUJ19 0.180 corr_x, POUJ20 corr_y 0.066: still touched; 16:00 back.
- **Images**: whole image 08:00: same as 04/10, nothing in front.
- **Confidence**: medium.
### 2026-10-05T20:00Z → 2026-10-06T20:00Z (added)
- **Phenomenon of the catalogue**: The camera moves through a thermal effect; plus a deviation of two targets that follows the housing temperature (no exact entry).
- **Precedents**: NEBRASKA-8 (reviewed, top targets' corr_y out until the evening); MICHIGAN-18 (not reviewed): the same day, curves staying low until 16:00 and back at 20:00; BREITHORN-30 (not reviewed).
- **Family**: Thermal displacement: pose and the targets' own deviation follow the housing temperature and come back with it at 20:00 (boundary Thermal / Displacement); not Displacement (it returns at equal temperature).
- **Measured**: temperature 33.75 (00:00), 33.0 (04:00), 33.4, 33.9, 34.25 (16:00), 34.5 (20:00). Tilt 0.0140 → 0.0154 → 0.0166 → 0.0182 (08:00, +6.2 noises) → 0.0167 → 0.0150 (16:00, back); scale 0.999813 at 08:00. corr_y POUJ19 −0.035 (05/10 20:00), −0.074, −0.051, −0.088, −0.082, −0.107 (16:00, −7.8 noises), −0.029 (20:00); COLJ20 −0.044 … −0.078 (16:00) → −0.029; POUJ20 0.054 (04:00) → −0.040 (16:00) → 0.033; COLJ19 0.046 (04:00) → −0.014 → 0.048; POUH19 0.105 (04:00) → 0.021 → 0.049. corr_x COLJ19 0.121 → 0.164 → 0.198 (04:00) → 0.183 → 0.179 → 0.159 → 0.137; COLJ24 0.054 at 04:00. Rigid (base 03/10 20:00): POUJ19 dy −0.069, −0.073, −0.080, −0.063, −0.067 from 00:00 to 16:00, −0.026 at 20:00; COLJ19 dx +0.036 → +0.060 (12:00) → +0.050 → +0.015; tgt_max 0.07-0.08 px vs 0.03 the days around; the three references within the targets' band (sigma0 high at 04:00 only). Start: POUJ19 already slightly low at 05/10 20:00 (2 noises), kept as the bound (one step within the uncertainty).
- **Images**: thumbnails 05/10 16:00, 06/10 08:00, 16:00, 20:00 of POUJ19, COLJ19, COLJ20 and the witness COLJ23: spots sharp, round, unchanged; whole image 06/10 08:00: nothing in front.
- **Confidence**: medium: clear on POUJ19 and COLJ19, the origin of their own deviation not established.
## 2. What I left without a label
- **3 October morning**: tilt +2.2 noises at 08:00, corr below 5 noises, temperature dip 34.6 → 33.5; no cycle to label (rule of 5 noises); no night before the start of the data.
- **12 October morning**: the coldest housing temperature of the range (31.9 °C at 04:00), tilt +4.7 noises at 04:00, scale −2.7, corr_x COLH21 +3.5 noises, COLJ19 +2; under the 5-noise rule on pose and curves, and it falls at the tilt peak of the slow drift (NEBRASKA-1). Borderline: BREITHORN-31 (not reviewed) labels the same morning there. A reviewer may want a label 12/10 00:00 → 12/10 16:00.
- **7-8 October**: corr_y bump of the top targets at the camera step and the POUH21 corr_x dip (−0.039, −7 noises at 08/10 12:00, back on 09/10) with its corr_y rising: the scene deforming with the camera step, within NEBRASKA-1.
- **9-11 October mornings**: temperature dips of half a degree or less, pose without a cycle beyond 3 noises.
- **error_value**: slow drifts of a few thousandths, far below the thresholds (0.7), no rejection except REFCOLH22: no effect of their own on the curves.
- **12/10 12:00 → 13/10 04:00**: plateau of the drift, under NEBRASKA-2; nothing new.
## 3. The documentation
- `families.md` (boundary Thermal / Displacement: "comes back with it at equal temperature") decided the 6 October case; the catalogue entry "The camera moves through a thermal effect" fits the three mornings.
- `labeling-rules.md`: "what falls under a long label is labeled apart" and "a signature found once is searched over the whole batch" led to the cool mornings, which the long Displacement had left out; "no thermal label under 5 noises" decided 3 and 12 October.
- Reading rule "a drift specific to a few targets that follows the camera, on the same dates, is the scene deforming" (NEBRASKA-6/5) settled the 7-8 October deviations; the 6 October deviation of POUJ19 and COLJ19 has its own dates, following the temperature rather than the pose: no rule says what that is (camera optics or targets), hence "not established".
- At a 4 h cadence, `days.py` nights have two measurements: the noise of the pose (and the 5-noise threshold) rests on very few values; 5 October sits exactly at 5.0, 12 October at 4.7. The threshold rule is fragile at this cadence.
- No image-quality metrics (old camera), no weather folder (no Swiss position): no weather available; the cause of a nuisance can only be read on the images, all dark indoor scenes.
- Precedents: the reviewed NEBRASKA cool mornings (7-11) gave the wording; the neighbours' labels of 4-6 and 12 October (not reviewed) confirmed the dates.
## Difficulties
- `overview.py --rigid --follow`: once a reference is rejected, its empty cells shift the columns of the printed table; read with care.
## 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` (all), `labeling/auto/MICHIGAN.jsonl` and `labeling/auto/BREITHORN.jsonl` (labels before 14/10)
- `precedents.py` ("cool morning … camera nods forward … top targets")
- Tools on `data/NEBRASKA` (outputs in `work/`): `inventory.py`, `overview.py` (4 h step, `--rigid` 03/10 00:00 and 03/10 20:00 with `--follow`), `metrics.py`, `days.py --night 16-21 --bounds`, `summary.py` corr_x / corr_y at 4 h (range and 11-17/10), `thumbnails.py` (05-06/10 top targets; REFCOLH22, REFH22 `--fixed` 09/10, 12/10), whole images 04/10, 05/10, 06/10 08:00, `plot.py`
- Weather: no `weather/` folder for NEBRASKA. Image quality: empty.