Files
aimsight-labeling/auto/analyses/NORDEND_2026-09-04T0615_2026-09-17T0545/analysis.md
T

152 lines
19 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: NORDEND, 2026-09-04T06:15Z → 2026-09-17T05:45Z (UTC)
Request: verify the 38 existing labels of the range (`work/existing.jsonl`, all validated by the user on 2026-11-17) and the spanning label NORDEND-24; keep, correct, remove, add. DETECTORS: no.
Result: 39 labels in `labels.jsonl`: **31 kept** unchanged (with their review field), **7 corrected** (NORDEND-101, 102, 106, 111, 113, 114, 122), **0 removed**, **1 added** (PR07 at night, 12/09 22:45 → 13/09 00:00).
Weather: **no weather** for NORDEND: the folder has no `weather/` (the inventory lists no weather column): no MeteoSwiss station, no distance. The weather is read on the images only.
Image quality: the six quality columns are empty (old camera; `quality.py`: "No quality metrics"): occultations and veils are read on the thumbnails.
## Verdict per old id
| Id | Verdict | Evidence |
|---|---|---|
| NORDEND-88 | kept | Pose cycle 04/09 (pan dropping, tilt rising in jerks 10:00–11:00, tilt −0.086…−0.103 from one measurement to the next), the targets follow the jerks a little; return in the evening and night; start = range start. |
| NORDEND-89 | kept | CTR04 rejected alone at 11:15, 0 references retained (inventory), 17 sensors jump together, all back at 11:30; edges clean. |
| NORDEND-90 | kept | Same at 13:45 (CTR04 rejected, 29 deviations ≥3.5 noises), back at 14:00. |
| NORDEND-91 | kept | PR07 rejected 16:30–18:00, CTR01 17:00–19:00, 19 targets deviating at 16:30, last deviation 19:00; edges 16:15 and 19:15 clean. |
| NORDEND-92 | kept | Steps 01:45 and 02:00 absent for all sensors, images absent too (01:15→02:15 gap in the image runs). |
| NORDEND-93 | kept | Pose leaves the night level at 06:30 (tilt −0.1216 → −0.1181): start 06:15 right; ends at the 06/09 start. |
| NORDEND-94 | kept | PR17 corr_y step at 12:15, stays (residual plot: plateau to the end of the range); edges clean. |
| NORDEND-95 | kept | CTR01 rejected 14:30–15:15 and 16:00–16:15; bounds 14:15/16:30 right. |
| NORDEND-96 | kept | Steps 01:30 and 01:45 absent, images absent (01:15→02:00 image gap). |
| NORDEND-97 | kept | Pose leaves at 05:45 (tilt −0.1321 → −0.1296, pan 0.0223 → 0.0206); big jump at 06:45–07:00. |
| NORDEND-98 | kept | Step 01:15 absent for all sensors, image present (image runs continuous). |
| NORDEND-99 | kept | Pan leaves at 05:30 (0.0083 → 0.0063), tilt from 05:15; later of the two 05:30. |
| NORDEND-100 | kept | Step 01:00 absent for all sensors, image present. |
| NORDEND-101 | **corrected** (end 02:30 → 02:45, description) | At 02:30 PR16 corr_x is still off (0.102 against ~0.13, about five noises) and its corr_y stays low until 02:30, back at 02:45; PR15 corr_x low at 02:00–02:15. The old text named only PR07. |
| NORDEND-102 | **corrected** (05:30 → 08:15, description) | PR07 corr_x leaves its night level (~0.055) at 05:45 (0.030) and 06:00 (0.039), its error_value doubles from 05:45 (0.048 → 0.072, 0.088), corr_y up at 06:00; main dip 06:30–07:45 (0.004–0.013); at 08:00 still 0.027 (more than half way), back at 08:15 (0.051) and 08:30 (0.068). |
| NORDEND-103 | kept | Pose leaves at 06:45 (tilt from 06:30, pan from 06:45). The evening pose jerk at 19:30 (tilt −0.111 → −0.134, sigma0 up, 15 sensors at 4–11 noises for one measurement) is written in its description. |
| NORDEND-104 | kept | PR07 rejected at 17:15 alone. |
| NORDEND-105 | kept | Pose leaves at 06:00 (tilt −0.1291 → −0.1263). |
| NORDEND-106 | **corrected** (end 15:30 → 16:00, description) | At 15:15 the targets that move most are PR18 (corr_y −0.04, about seven noises), PR15 (corr_y −0.03), PR14 and PR05 (corr_x +0.035, +0.03), PR08 (corr_y −0.027); the old text named PR05, PR14, PR01 only. PR08 and PR15 corr_y stay low at 15:30–15:45 (edges: −6, −8 local noises), back at 16:00. |
| NORDEND-107 | kept | CTR02, CTR03 rejected 05:00–05:15, no pose at these two instants, all targets jump (up to 96 noises), CTR01 rejected until 07:45; bounds right. |
| NORDEND-108 | kept | Pose cycle 10/09 with the big jump at 09:00 (tilt −0.113 → −0.080). |
| NORDEND-109 | kept | CTR01 rejected 14:45–15:00 and 15:30–16:15. A small swing of PR10, PR01, PR09 at 16:30–16:45 (4–7 noises) and a group jump at 14:45 (PR06, PR17, PR07 beside PR14, PR18) were weighed: at 16:45 the deviations stay at or under five noises and the evening pose return starts at 17:00; not enough to move the bound. |
| NORDEND-110 | kept | Pose leaves at 06:15 (tilt −0.1302 → −0.1261). |
| NORDEND-111 | **corrected** (end 07:00 → 05:45 on 13/09) | The cycle of 12/09 ends at the next morning turning point: on 13/09, after the step at 04:30, pan and tilt rest at 04:30–05:30 (pan 0.0111–0.0121, tilt −0.1201…−0.1218), the tilt leaves from 05:30, the pan from 05:45 (0.0107, then 0.0095, 0.0095, 0.0090, 0.0085, 0.0072 without coming back); the housing temperature is at its trough at 05:00 and rises from 06:00. 07:00 was the fast part of the movement (raw y jumps at 07:15). |
| NORDEND-112 | kept | 04:15: CTR02, CTR03, CTR04 at 8–18 noises, PR17, PR18, PR11 too, back at 04:30; PR07 rejected 04:15–04:45 and still off at 05:00 (−6): end 05:15 right. |
| NORDEND-113 | **corrected** (start 07:00 → 05:45) | Same reason as NORDEND-111: the morning turning point is 05:45. |
| NORDEND-114 | **corrected** (end 15:30 → 16:00, description) | PR07, rejected until 15:15, comes back at 15:30 far below its level (corr_x 0.002 against ~0.05–0.065, about ten noises) and still low at 15:45 (0.041) with its error_value still high (0.126, 0.109 against ~0.06); back at 16:00. Now overlaps NORDEND-115 (CTR01 in the sun, another event). |
| NORDEND-115 | kept | CTR01 rejected at 15:45 alone. |
| NORDEND-116 | kept | Pose 14/09: tilt flat until 07:00, rises from 07:15; pan drifts slowly; temperature rises from 07:00. |
| NORDEND-117 | kept | PR07 rejected at 08:30 alone. |
| NORDEND-118 | kept | CTR01 corr_x −10 noises at 14:30 (before the first rejection 14:45), rejections until 16:15. |
| NORDEND-119 | kept | Pose leaves at 06:00; mid-morning step 08:00–08:45. |
| NORDEND-120 | kept | CTR01 rejected 15:00–16:00. |
| NORDEND-121 | kept | PR07 rejected 01:45–03:00, error_value already up at 01:15–01:30, corr_y +0.02 at 01:30 (about three noises, within its night spread): start 01:30 holds. |
| NORDEND-122 | **corrected** (end 08:30 → 08:45, description) | PR13 corr_y at 08:00, 08:15, 08:30: 0.096, 0.110, 0.099 against 0.058–0.064 (seven noises still at 08:30), back at 08:45 (0.061); PR16 corr_y +0.02–0.03 at the same three measurements. Rigid residual from 07:45: PR13 dy +0.027, +0.047, +0.036, PR16 dx −0.028, −0.054, while the median of the targets stays at 0.01–0.02: they move on their own in the image. At the same time the housing cools (34.5 °C at 07:00, 33.9 at 08:00) and the pose turns back (tilt −0.1136 → −0.1213, pan 0.0088 → 0.0159). The old text said "for one measurement". |
| NORDEND-123 | kept | Pose false start 06:00–07:30, return at 08:00–08:30 with the cooling, leaves for good at 08:45; ends at range end. |
| NORDEND-124 | kept | CTR01 rejected 10:00–11:15. |
| NORDEND-125 | kept | CTR01 rejected 14:30, 15:30, 16:15. |
| NORDEND-24 (spanning) | not changed, holds within the range | PR04 rejected 475 times in the range, by day and by night, a good part of the time 06–08/09, and again on 10/09 and 15–16/09; its error_value far above its neighbours'. Nothing wrong within the range. |
## 1. My labels
The 31 kept labels are the existing ones, verified as in the table above; their fields are unchanged. Below, the corrected and added labels in full, then the kept ones by kind.
### 2026-09-08T01:45Z → 2026-09-08T02:45Z (NORDEND-101, corrected)
- **Phenomenon of the catalogue**: A faint spot lost when the light is missing (noisy night image).
- **Precedents**: NORDEND-70 (validated: PR07 rejected at night in a noisy image, PR15 and PR16 jump with it) and NORDEND-73: same reading, neighbours named.
- **Family**: Visibility, as the existing label and the precedents: the whole image is noisy, nothing hides PR07.
- **Measured**: PR07 rejected 02:00–02:15 (error_value ~0.98). PR16 corr_x 0.135 (01:45) → 0.150, 0.124, 0.102 → 0.149 (02:45), 0.132 (03:00); PR16 corr_y −0.009 → −0.022, −0.023, −0.022 → −0.012 (02:45). PR15 corr_x 0.129 → 0.112, 0.116 → 0.124 → 0.137. PR14 corr_y lower 02:00–03:15 (0.012–0.018 against 0.03), within its night wander (0.018 at 04:15–04:45, 0.019 at 23:15): not named.
- **Images**: not looked at again; the reviewed reading of the 02:00 image is kept.
- **Confidence**: medium (end on a deviation of about five noises).
### 2026-09-08T05:30Z → 2026-09-08T08:15Z (NORDEND-102, corrected)
- **Phenomenon of the catalogue**: Targets move slowly, alone, then come back (one target, a few hours).
- **Precedents**: none new; the reviewed reading (weak spot of PR07, cause not established) is kept.
- **Family**: Displacement, unchanged: a corr_x dip of PR07 alone, without rejection, without a seen cause.
- **Measured**: PR07 corr_x, night level 00:00–05:30 0.049–0.063: 05:45 0.030, 06:00 0.039, 06:15 0.050, 06:30–07:45 0.004–0.013, 08:00 0.027, 08:15 0.051, 08:30 0.068. Error_value 0.042–0.048 until 05:30, 0.072 and 0.088 at 05:45–06:00, 0.139 at 06:30, back to 0.049 at 07:45. Corr_y 0.075 → 0.095 (06:00), 0.104 (06:30), back at 07:15. PR06, PR09, PR16 flat.
- **Images**: the reviewed 07:00 reading is kept.
- **Confidence**: medium.
### 2026-09-09T14:45Z → 2026-09-09T16:00Z (NORDEND-106, corrected)
- **Phenomenon of the catalogue**: The sun in the field or behind a spot (CTR01).
- **Precedents**: NORDEND-95, 109, 118, 120 in the range (same hour, same signature).
- **Family**: Visibility, unchanged.
- **Measured**: CTR01 rejected 15:00–15:15 (0.46, 0.48). At 15:15 error_value ×5–6 on PR08, PR15, PR14; corr_y PR18 −0.237 (level −0.195), PR15 −0.005 (0.025), PR08 −0.024 (0.003); corr_x PR14 0.155 (0.12), PR05 0.091 (0.06), PR01 0.024 (0). At 15:30–15:45 PR08 corr_y −0.007, −0.009, PR15 0.015, 0.009; back at 16:00 (0.0005, 0.021).
- **Images**: none (cause deduced, as in the reviewed label).
- **Confidence**: medium (end on deviations of a few local noises).
### 2026-09-12T06:00Z → 2026-09-13T05:45Z (NORDEND-111) and 2026-09-13T05:45Z → 2026-09-14T07:00Z (NORDEND-113), corrected
- **Phenomenon of the catalogue**: The camera moves through a thermal effect.
- **Precedents**: the other cycles of the range.
- **Family**: Thermal displacement, unchanged.
- **Measured**: 13/09 pose, 15 min: 04:30–05:30 pan 0.0121, 0.0117, 0.0116, 0.0111, 0.0116, tilt −0.1218, −0.1213, −0.1217, −0.1204, −0.1201; then pan 0.0107 (05:45), 0.0095, 0.0095, 0.0090, 0.0085, 0.0072 (07:00), tilt −0.1188, −0.1177, −0.1169, −0.1170, −0.1159, −0.1152. Housing temperature (hourly median) 33.6 °C at 05:00, 33.75 at 06:00, 34.5 at 07:00. Raw y of the targets −0.31 (05:30) → −0.18 (06:00) → +0.04 (07:15). Descriptions unchanged: they still hold (weak day for NORDEND-113).
- **Images**: none.
- **Confidence**: high on the direction, medium on the exact step (slow pose).
### 2026-09-12T22:45Z → 2026-09-13T00:00Z (added)
- **Phenomenon of the catalogue**: A faint spot lost when the light is missing / part of the scene goes dark at night; cause not established.
- **Precedents**: NORDEND-150 (validated: clear PR07 corr_x jump at night, spot smaller, nothing in front: Visibility, cause not established), NORDEND-73 (lit cobweb thread on PR07 at night).
- **Family**: Visibility, as NORDEND-150 and the other night losses of PR07 in this range (101, 104, 117): the spot weakens, no object seen; not Obstruction, no thread seen.
- **Measured**: PR07 corr_x 0.072, 0.065 (22:30–22:45), 0.044 (23:00), 0.105, 0.122 (23:15–23:30), 0.080 (23:45), 0.070 (00:00); noise 0.006: +6 and +9 noises against the local median. Corr_y 0.086 → 0.107, 0.106, 0.108 → 0.093 (00:00). Error_value 0.042 → 0.113 at 23:15 → 0.044. PR04 rejected 23:15–23:30 (0.93, 0.96). PR06, PR08, PR09, PR03 flat. Not covered by any existing label (`work/brief.txt`).
- **Images**: fixed thumbnails 23:00, 23:15, 23:30, 23:45 (`pr07_night_1209.png`): PR07 spot smaller at 23:15 and 23:30, PR04 spot almost gone, PR06 unchanged; zone sheet with contrast: nothing visible in front.
- **Confidence**: medium (deviation clear, cause not established).
### 2026-09-13T11:30Z → 2026-09-13T16:00Z (NORDEND-114, corrected)
- **Phenomenon of the catalogue**: Atmospheric turbulence (possible), as the reviewed label says.
- **Precedents**: none new.
- **Family**: Visibility, unchanged.
- **Measured**: PR07 rejected 14:00–15:15; corr_x 0.002 (15:30), 0.041 (15:45), 0.047 (16:00), 0.053 (16:15), morning level 0.055–0.065; error_value 0.126, 0.109, then 0.062. PR18 corr_y −0.176, −0.157 at 15:00–15:15. Overlap with NORDEND-115 (CTR01 at 15:45, the sun): two events.
- **Images**: the reviewed reading (12:00 to 14:45) kept.
- **Confidence**: medium.
### 2026-09-16T07:45Z → 2026-09-16T08:45Z (NORDEND-122, corrected)
- **Phenomenon of the catalogue**: none exactly (faint PR07 by day; PR13, PR16 moving alone in the image for three measurements).
- **Precedents**: none useful.
- **Family**: Visibility, unchanged: the expert's reading of PR07, one event at the same instants; the PR13/PR16 deviation is their own in the image (rigid residual), not a remainder of the compensation.
- **Measured**: see the verdict table; `overview.py --rigid 2026-09-16T07:45Z --follow PR13,PR16,PR11,PR07`.
- **Images**: the reviewed 08:00 reading kept.
- **Confidence**: medium.
### The kept labels
- **Thermal displacement** (NORDEND-88, 93, 97, 99, 103, 105, 108, 110, 116, 119, 123): each start checked on the 15 min pose (table above); each end is the next start (the camera keeps coming back all night: no rest). Housing temperature, morning trough → afternoon (hourly median, °C): 37.4→40.5 (04/09), 31.4→39.6, 26.8→45.3, 32.9→48.6, 35.6→43.8, 28.5→37.0, 28.7→37.6, 28.2→38.3, 31.2→39.5, 33.6→35.7, 36.3→43.8, 32.0→48.2, 33.3→37.9 (16/09). The corr_x of PR18, PR16, PR15, PR14 dip every morning with the pose and come back (`thermal_corr_x.png`).
- **No data** (92, 96, 98, 100): the four absent steps of the inventory, exactly bounded.
- **CTR01 in the sun** (95, 109, 115, 118, 120, 125) and **CTR01 by morning** (124): every CTR01 rejection of the range is inside one of them, or inside 89, 91, 107, 114.
- **PR07 at night or in poor light** (104, 117, 121): every PR07 rejection of the range is inside a label.
- **Reference losses** (89, 90, 107) and **night jump** (112): checked on the inventory (retained references) and the per-instant deviations.
- **PR17 jump** (94): plateau to the end of the range on the corr_y residual.
- **PR07, cobweb at dusk** (91): bounds clean.
## 2. What I left without a label
- **PR10 by night 16–17/09**: its error_value rises from about 03:00 on 16/09 (×3 to ×5, up to 0.33) and stays high until about 05:00 on 17/09; PR11, PR01, PR02 rise less from the evening of 16/09. The 23:00 thumbnail shows a broad lit diagonal band (a cobweb thread close to the lens) over PR10, PR12 at its edge. PR10 corr_y minus the median of the other targets: +0.03 to +0.046 on 16/09 19:00–22:00 against +0.016 to +0.028 at the same hours of 15/09: about two noises, within its usual spread; no rejection. Not labelled: the curves do not move clearly. To watch in the next range (after 17/09 05:45).
- **Small group deviations at the morning and evening jerks of the pose** (04/09 10:15–10:45, 06/09 07:00 and 09:00, 10/09 09:00–09:30, 12/09 08:30, 14/09 10:00, 10/09 17:00…): 10–19 sensors at 4–7 noises for one measurement, exactly when the pose jumps (10/09 09:00: tilt −0.113 → −0.080, raw y +1.9 px). Remainder of the camera's thermal movement, inside the thermal labels, as the reviewed labels do (NORDEND-97, 103 write it).
- **Single-sensor deviations of 6–7 noises** (PR18 at 06:45 on 04/09, PR18 and PR10 at 16:30 on 05/09, PR10 at 16:30 on 10/09, PR15 at 10:30 on 12/09, PR13 at 16:30 on 15/09): one sensor, one measurement, of the order of the largest noise; no rejection, no error_value rise.
- **PR10, PR01, PR09 at 10/09 16:30–16:45**, at the edge of NORDEND-109: at most five noises at 16:45.
- **The missing image at 15/09 02:15** (the measurement is there): touches no curve.
- **Error_value rises of many sensors without effect on the corr** (08/09 12:45 and 15:30; 09/09 10:15–11:45; 16/09 12:15): nothing beyond the noise on the corr residuals.
## 3. The documentation
- `labeling-rules.md` section 2 decided most corrections: "a measurement that has come back halfway is still touched" (102, 114, 122), "each edge is checked sensor by sensor" (101, 106), "the camera cycle starts at the morning turning point … not the fast part of the movement" (111, 113). The rule that one step of difference stays within uncertainty helped to keep the other thermal starts.
- The precedents NORDEND-150 and NORDEND-73 gave the family and wording of the added label.
- Several reviewed labels describe the neighbours that jump with a PR07 rejection loosely ("for one measurement", or not at all): reading the per-instant deviations of all sensors (`scripts/brief.py`, `scripts/edges.py`) found them. A tool that prints, for each label, the deviation of every sensor at s−1, s, e, e+1 would make the edge rule cheap to apply.
- `tools/DATA.md`, inventory: no weather and no quality metrics at NORDEND.
- Nothing contradicted by the data. The thermal labels include the one-measurement jerks of the pose in the cycle rather than labelling them apart ("what falls under a long label is labeled apart" could suggest otherwise); followed, as the reviewed labels do.
## Difficulties
- The measurement times carry seconds: my first edge script, keyed by the exact time, found nothing until it truncated to the minute (fixed).
- The Write tool refused `.md` files in this run; `analysis.md` and `report.md` were written with a shell heredoc.
## Sources read
- `agents/agents/analyst.md`, `instructions/analyst.md`, `work/prompt.txt`, `request.json`, `work/existing.jsonl`, `work/spanning.jsonl`.
- `tools/GUIDE.md`, `tools/DATA.md` (in full), `tools/scripts/data.py` (for my scripts).
- `labeling/families.md`, `labeling/labeling-rules.md`, `labeling/reading-rules.md`, `labeling/stations.md` (in full).
- Precedents: `labeling/scripts/precedents.py "at night PR07 left under its threshold jumps on corr_x …" --exclude NORDEND 2026-09-04T06:15Z 2026-09-17T05:45Z --k 6` (NORDEND-73, 69, 70, 150, 71, 84).
- Tools: `inventory.py` (window), `overview.py` (range: table, residuals; `--rigid` on 16/09 07:45), `summary.py` (15 min, many windows), `metrics.py` (temperature), `plot.py`, `thumbnails.py` (12/09 23:00–23:45 PR07/PR04/PR06 fixed and zone; PR10/PR12 15–17/09), `quality.py` (no metrics).
- My scripts: `scripts/brief.py` (brief deviations against a rolling median, with label coverage; `work/brief.txt`, `work/brief35.txt`), `scripts/edges.py` (edge deviations of each label; `work/edges.txt`), `scripts/others.py` (corr minus the other targets), `scripts/write_labels.py` (builds `labels.jsonl`).