Files
aimsight-labeling/auto/analyses/NORDEND_2026-08-23T0515_2026-09-04T0615/analysis.md
T

155 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, 2026-08-23T05:15Z → 2026-09-04T06:15Z (UTC)
Request: verify the 37 existing labels of the range (`work/existing.jsonl`, all reviewed by the user on 2026-11-17), keep / correct / remove, and add what they miss. DETECTORS: no. Result: 32 kept, 5 corrected (one of them split in two), 0 removed, 4 added; 42 labels in `labels.jsonl`.
Context read first:
- **Weather**: no `data/NORDEND/weather/` folder: no MeteoSwiss weather for this station (no Swiss position). The images are the only witness of the weather.
- **Image quality**: none (old camera, `quality.py`: "No quality metrics"); the occultations are read on `error_value`, the rejections and the images.
- **Housing temperature**: 26.5 °C (24/08 05:00) to 46.9 °C (27/08 12:00); pan r = −0.98, scale r = +0.97, tilt r = +0.74 against it over the range (`metrics.py`, hourly): the daily camera cycle is thermal. BusVoltage 13.7 to 14.3 V all along: no energy problem.
- **Inventory**: 1133 instants, 15 min; 7 gaps (1, 1, 1, 16, 1, 2, 2 missing steps); one instant with 0 retained references (28/08 12:45, CTR04 rejected, CTR02 and CTR03 with empty inlier); PR04 rejected 141 times (spanning Target issue NORDEND-24), PR07 135, CTR01 55, everything else 9 or fewer (`work/inventory.txt`).
- **Method of the check**: `scripts/scan.py` lists, at each instant, the targets whose `corr_x`/`corr_y` leave a rolling median (±4 measurements, self excluded) by more than 5 (then 4) own noises, the rejected sensors, and the non-thermal existing labels that cover the instant (`work/scan.txt`); every edge of a short label was then read on `summary.py` at 15 min; the thermal cycles on the pose at 15 min (morning turning points) and `scripts/thermal_days.py` (targets that dip each day); the images with `thumbnails.py` (`work/img/`).
## Verdict per old label
| Old id | Verdict | Evidence |
|---|---|---|
| NORDEND-51 | kept | pose cycle pan 0.039°, tilt 0.058°; corr_x PR18 −0.12, PR16, PR15, PR14/PR17 dip at 07h, PR18 corr_y up; start at the range edge |
| NORDEND-52 | kept | turning point 05:15 (tilt leaves −0.132 at 05:15, pan at 05:30); PR18, PR16, PR15, PR14 dip (PR17 as much as PR14), PR18 corr_y up |
| NORDEND-53 | kept | PR07 rejected 24/08 11:15 → 25/08 12:15; 22:00 thumbnail: pale blurred band over PR07, weak spot by day (10:00, 14:00, 09:00) |
| NORDEND-54 | kept | 25/08 01:30 missing for all; image 01:30 exists, last_sync at 01:30, voltage normal; no other station has data that day |
| NORDEND-55 | kept | turning point 05:15; PR18, PR16, PR15, PR17 dip |
| NORDEND-56 | kept | turning point 06:00; PR18 and PR15 dip (PR16 as much as PR15, not named: incomplete, not wrong) |
| NORDEND-57 | kept | PR07 rejected from 07:15, CTR01 until 09:45, PR16, PR09; jumps of PR08, PR13, PR11, PR06, PR16 at 08:00-08:30; CTR01 still off on corr_y at 10:15 → end 10:30 right |
| NORDEND-58 | kept | 27/08 01:30 missing, no image, no sync at 01:30 (syncs 01:00, 01:15, 01:45) |
| NORDEND-59 | **corrected** | end 06:15 → 07:00: the next morning's turning point is 07:00 (section 1) |
| NORDEND-60 | **corrected** | end 12:00 → 07:45 (last measurement before the gap); description: the drops come right after the spider, then PR08 again around midnight |
| NORDEND-61 | kept | 21:15 image: the spider's lit body bottom right, spots faded; 13 targets rejected at 21:15, 10 at 21:30; PR09 rejected again at 21:45, but the 21:45 image shows no spider and the cobweb (NORDEND-60) explains it |
| NORDEND-62 | kept | 28/08 01:45 missing, no image, no sync at 01:45 |
| NORDEND-63 | **corrected** | start 06:15 → 07:00 (section 1) |
| NORDEND-64 | kept | 28/08 08:00 → 11:45 missing (16 steps), no image; metrics hourly all along, voltage 14.1-14.3 V |
| NORDEND-65 | kept | 12:45: CTR04 rejected, 0 retained references (CTR02, CTR03 inlier empty), 14 targets jump up to 65 noises; PR06, PR08, PR15, PR16, PR18 rejected; back at 13:00 |
| NORDEND-66 | kept | CTR01 rejected 14:30; at the same measurement the error_value of all the sensors rises and PR18, PR17, PR09 move a little on corr_x (not described: incomplete, not wrong) |
| NORDEND-67 | kept | rejections 16:00-16:15 of PR09, PR11, PR12, PR14, PR15, PR16, PR07; PR15 corr_y −0.048 at 17:00, −0.026 at 17:15 (halfway), −0.013 at 17:30 → end 17:30 right |
| NORDEND-68 | kept | turning point 06:00; PR18, PR16, PR15, PR14 dip, PR18 corr_y up |
| NORDEND-69 | kept | PR07 rejected 16:30-16:45, PR07 corr_x jumps until 18:45 → end 19:00 right |
| NORDEND-70 | kept | PR07 rejected 22:30; PR16, PR15 jump 22:30-23:00 |
| NORDEND-71 | kept | PR07, PR09, PR11, PR12, PR14 rejected 00:30-00:45; PR14 corr_x +21 noises at 00:45, PR12 corr_y +11 at 01:00, back 01:15 |
| NORDEND-72 | kept | turning point 05:45-06:00 by pan and tilt, 06:15 within one or two steps of a gradual onset; PR14, PR15 dip a little, PR18 corr_y up |
| NORDEND-73 | kept | PR07 rejected 00:00; PR16 corr_y −22 noises, PR08, PR15 at 00:00, PR14 at 00:15 |
| NORDEND-74 | kept | turning point 07:00; PR18, PR14 dip a little |
| NORDEND-75 | kept | PR07 rejected 12:45-13:15; PR14 corr_x −21 noises at 13:00, 15 sensors at 13:15; back 13:30 |
| NORDEND-76 | **corrected (split)** | CTR01 rejected at 14:45 and at 16:00 only, four clean measurements between (15:15 thumbnail: rock less lit) → two labels |
| NORDEND-77 | kept | 01/09 01:00 missing, no image; JUNGFRAU has the same gap (00:45 → 01:15). WILDSTRUBEL measures at that minute (not mentioned; the label's "possible" stays right) |
| NORDEND-78 | kept | turning point 06:45; PR18, PR16 dip a little, PR18 corr_y up |
| NORDEND-79 | kept | CTR01 rejected 15:00-15:15 |
| NORDEND-80 | kept | pan step 0.006° at 15:45 (among the largest daytime steps of the range); 11 targets off by 4-8 noises at 15:45 only |
| NORDEND-81 | kept | turning point 05:45-06:00; PR18, PR16, PR17, PR15 dip |
| NORDEND-82 | kept | 03/09 01:30, 01:45 missing, no image, syncs 01:00, 01:15 only; JUNGFRAU same minutes (WILDSTRUBEL measures) |
| NORDEND-83 | kept | turning point 06:30; PR15 dips a little; ends at the range end |
| NORDEND-84 | **corrected** | end 13:45 → 14:00: at 13:45 PR18 corr_x +0.04 (+7 noises) and PR07 below its level; back at 14:00 |
| NORDEND-85 | kept | PR07 error_value up, PR14 +13, PR18 +11, PR07 +9 noises on corr_x at 16:45, back 17:00 |
| NORDEND-86 | kept | PR07 rejected 23:00-23:30; PR16 corr_x +0.10 at 23:00, PR15 down to 00:15; 22:45 and 00:30 already back (PR14 0.112, PR15 0.134) |
| NORDEND-87 | kept | 04/09 01:15, 01:30 missing, no image, syncs 01:00 and 01:45; JUNGFRAU same minutes (WILDSTRUBEL measures) |
Spanning **NORDEND-24** (PR04, Target issue, 17/08 → 01/10): holds within the range: PR04 rejected 50 times in a row on 23/08 and most of 24-25/08, its corr noise 0.012 (twice that of its neighbours). Its "rejected almost every day" is less true here: no rejection of its own on 26/08, 29/08, 01/09 and 02/09 (only shared ones on 27/08 21:15 and 28/08 13:00). Not wrong enough to call for a change.
## 1. My labels
Only the changed and the new labels are detailed; the kept ones are in the table above.
### 2026-08-27T05:45Z → 2026-08-28T07:00Z (NORDEND-59, end corrected)
- **Phenomenon of the catalogue**: The camera moves through a thermal effect
- **Precedents**: NORDEND-41 (a cycle whose morning falls in a gap) for the reading of 28/08
- **Family**: Thermal displacement, unchanged
- **Measured**: 28/08, pan −0.0013 to −0.0019° from 03:00 to 06:15, then −0.0022 (06:30), −0.0024 (06:45), −0.0030 (07:00), −0.0046 (07:45); tilt −0.1129 to −0.1152° all night, one point at −0.1110 at 06:30 then back to −0.1146 at 06:45, −0.1126 at 07:00, out at 07:15 (−0.1075). The later of the two axes leaves at 07:15: the cycle of 27/08 ends, and that of 28/08 starts, at 07:00; 06:15 was three steps early.
- **Images**: none needed (pose)
- **Confidence**: medium: a weak morning, and the gap at 08:00 hides the fast part.
### 2026-08-28T07:00Z → 2026-08-29T06:00Z (NORDEND-63, start corrected)
- Same reading as above; the rest of the label holds (weak cycle, the jump at 12:45 and the afternoon disturbances labelled apart).
### 2026-08-27T19:30Z → 2026-08-28T07:45Z (NORDEND-60, end and description corrected)
- **Phenomenon of the catalogue**: A spider web in front of the lens (Visibility when it veils the image)
- **Precedents**: none useful beyond the label itself
- **Family**: Visibility, unchanged (the web veils the bottom right of the scene, several targets)
- **Measured**: CTR01 and PR07 rejected almost every measurement from 21:00 to 07:45 (PR07 from 19:45); after the spider, corr_x of PR08 −0.106 (21:45), PR03, PR06, PR16, PR15, PR05 down at 21:45-22:15, back at 22:30; PR08 down again from 23:00 to 01:00 (−0.004 at 01:00 against about 0.05), back by 02:00. The old end, 12:00, ran over the four-hour gap (NORDEND-64): the last touched measurement before the gap is 07:45, and at 12:00 neither CTR01 nor PR07 is rejected any more (labelling rules, 2: at the edge of a gap, the last measurement before). The old description put the drops "a few hours after the spider's passage": they come right after it, and only PR08 drops again around midnight.
- **Images**: 21:00, 21:45, 22:15, 22:30: wide blurred pale bands across the bottom of the scene, over the targets at the bottom right; 05:00 the same; 07:45 (day) still a faint pale band
- **Confidence**: high
### 2026-08-31T14:30Z → 15:00Z and 15:45Z → 16:15Z (NORDEND-76, split)
- **Phenomenon of the catalogue**: The sun in the field or behind a spot
- **Precedents**: NORDEND-66, NORDEND-79 (the same CTR01 on sunny afternoons); JUNGFRAU-45/55 (a lit background behind a spot)
- **Family**: Visibility, unchanged
- **Measured**: CTR01 rejected at 14:45 and at 16:00 only; at 15:00, 15:15, 15:30, 15:45 it is measured and nothing deviates (scan at 4 noises). An intermittent episode is cut at a pause of several measurements, and the isolated rejections of a target each have their label (labelling rules, 1 and 3).
- **Images**: CTR01 at 14:30, 14:45, 15:15, 16:00: the rock behind it in full sun at 14:45 and at 16:00, less lit at 15:15
- **Confidence**: high
### 2026-09-03T13:15Z → 14:00Z (NORDEND-84, end corrected)
- **Phenomenon of the catalogue**: A faint spot lost when the light is missing
- **Precedents**: NORDEND-85, NORDEND-75 (PR07 the same days)
- **Family**: Visibility, unchanged (consistency with the other PR07 moments of the station)
- **Measured**: 13:30 PR07 rejected, PR15 corr_x +0.05, PR16, PR12, PR01 corr_y up; 13:45 PR18 corr_x 0.260 against 0.22 (+7 noises), PR07 0.046 against 0.07; 14:00 all back. A measurement still touched at 13:45: the end is 14:00.
- **Images**: the 13:30 thumbnail already cited (weak PR07)
- **Confidence**: high on the bounds
### 2026-08-28T12:00Z → 12:30Z (added)
- **Phenomenon of the catalogue**: none exactly (closest: visibility drops over the whole scene, without an obstacle, but brief)
- **Precedents**: NORDEND-65 (the next measurement, same signature with CTR04 lost); NORDEND-173 (error up on almost all sensors)
- **Family**: Visibility: the whole scene degraded (error_value up on every sensor together), no object, no camera step; same family as NORDEND-65
- **Measured**: error_value of all the sensors about three times higher at 12:00 and 12:15 (CTR04 0.20, 0.24; targets 0.15-0.21; 0.04-0.07 at 12:30), no rejection; corr_x at 12:15: PR03 0.045 (0.028 around), PR09 +0.03, PR12 +0.025, PR11 +0.02, PR01, PR10 corr_y down (7 targets at 5-6 noises); 12:00 already PR03 −0.02, PR17 corr_y −7 noises; 12:30 back. 12:00 is the first measurement after the gap: the event starts there (labelling rules, 2). 12:30 is clean, hence a label apart from NORDEND-65.
- **Images**: zone of the targets at 12:00, 12:15, 12:30, 12:45: quarry in the sun, nothing in front, no visible difference at that scale
- **Confidence**: medium (the cause is not seen)
### 2026-08-29T11:45Z → 12:15Z (added)
- **Phenomenon of the catalogue**: A step of the camera that the compensation does not erase
- **Precedents**: NORDEND-80 (same station, same signature, reviewed); LENZSPITZE-84 (a one-measurement jolt of the camera during the thermal day, Displacement); SNCF01-42
- **Family**: Displacement, like NORDEND-80: a camera step with a remainder, cause not established
- **Measured**: pan −0.0090 → −0.0003 → −0.0033°, rotation −0.0127 → −0.0031 → −0.0090° (a pan step of 0.0087°, the 4th largest daytime pan step of the range; median night step 0.0005°); rigid fit from 11:30: translation 0.37, −0.39 px at 12:00, still 0.28, −0.33 px at 12:15 (the camera keeps most of it); residuals PR15 +0.032, PR16 +0.030 px, CTR04 −0.025 px at 12:00 only; corr_x PR15 0.168 (0.12 around), PR16 0.170 (0.12-0.13), PR14 +0.015; error_value at its level, no rejection.
- **Images**: PR15, PR16, CTR04 at 11:45, 12:00, 12:15 (fixed frame): spots intact, nothing in front; background darker at 12:00
- **Confidence**: medium
### 2026-08-31T11:30Z → 12:00Z (added)
- **Phenomenon of the catalogue**: none exactly (closest: the base changes without a reference lost)
- **Precedents**: JUNGFRAU-45/55/72 (group jumps of one measurement when a reference shifts, there with the sun behind it); none of this exact shape at NORDEND
- **Family**: Displacement: a jump then the return without an established cause; no nuisance seen that would make it Visibility
- **Measured**: corr_x PR05 0.072 → 0.009 → 0.072 (−13 noises), PR08 0.052 → 0.015 → 0.046, PR12 0.059 → 0.025 → 0.060; error_value lower, not higher (CTR04 0.025 against 0.064); all three references retained; rigid residuals at 11:45: PR05 −0.045 px, CTR04 +0.036 px, median of the targets 0.015. Pose agitated (rotation −0.0189° at 11:30, −0.0109° at 11:45).
- **Images**: PR05, PR08, PR12, PR06, CTR04, CTR02 at 11:30, 11:45, 12:00: spots intact; image background 23 → 5 grey levels from 11:30 to 11:45 (whole scene darker)
- **Confidence**: medium (a real but small, one-measurement event; cause open)
### 2026-09-03T08:45Z → 09:15Z (added)
- **Phenomenon of the catalogue**: A faint spot lost when the light is missing (here kept, not rejected)
- **Precedents**: NORDEND-85 (PR07, error up for one measurement without rejection, Visibility)
- **Family**: Visibility, consistent with the other PR07 moments
- **Measured**: PR07 error_value 0.194 (0.06-0.07 around, threshold 0.4), corr_y 0.141 against 0.087-0.091, back at 09:15; PR06 and PR08 unchanged; pose without step.
- **Images**: PR07, PR06 at 08:45, 09:00, 09:15 (fixed frame): PR07's spot dimmer at 09:00, nothing in front
- **Confidence**: high on the event, cause not established
## 2. What I left without a label
- **Morning agitation of the thermal cycles** (23/08 08:30, 24/08 08:00-09:00, 25/08 08:15-09:30, 30/08 09:00, 01/09 08:15-08:30, 02/09 08:30-09:45): 3-9 targets at 5-9 noises during the fast morning movement of the camera (tilt steps up to 0.026° by day against 0.0007° at night): the remainder of the compensation, described in the thermal labels ("the compensation does not erase everything").
- **Afternoon jumps under the pose agitation** (23/08 15:15 and 15:45, 24/08 15:15 and 15:45 with PR01 corr_y +0.05 and PR14 corr_x +0.04, 26/08 14:45 and 16:15): 2-5 targets at 4-9 noises, the pose stepping (tilt −0.107 → −0.115° at 24/08 15:45), error_value not rising: the agitation of the thermal day, which goes into the thermal label (labelling rules, 5), not a Visibility. The 24/08 ones fall under NORDEND-53, whose subject is PR07 alone.
- **Whole-scene error_value rises with nothing on the curves**: 27/08 10:45, 11:45, 19:30; 28/08 13:30; 29/08 07:45, 10:00; 30/08 10:30; 31/08 08:45, 09:00, 11:00, 12:00, 14:00, 14:30, 15:15, 16:15, 16:30; 01/09 11:15 (14-21 sensors above 5 MAD): no target leaves its level at 5 noises at these instants; an error_value without effect on the curves is not labelled.
- **References alone at 5-8 noises** (23/08 08:30, 24/08 14:45 and 17:00, 27/08 10:45 and 14:30, 02/09 14:00, 03/09 14:15): their corr only, the targets do not follow.
- **CTR04 and PR05, PR08, PR15 error_value raised all night 24-25/08** without effect on their corr (the thread of NORDEND-53 nearby).
## 3. The documentation
- `labeling-rules.md`, section 2, settled the three bound corrections: the camera cycle starts at the later of tilt and pan (28/08), at the edge of a gap the bound is the last measurement before (NORDEND-60), a measurement still halfway is touched (NORDEND-84; it also confirmed NORDEND-67 and NORDEND-57 as right). Sections 1 and 3 (isolated rejections, an intermittent episode cut at a pause of several measurements) settled the split of NORDEND-76.
- `families.md`, "A jump then the return ... without an established cause: Displacement", for the 31/08 group jump. The station's own brief PR07 labels (NORDEND-75, 84, 85) are Visibility with "cause not established": I followed them for PR07 (same target, same family), but the boundary text would rather say Displacement for a jump without a seen nuisance; a line in `families.md` on a weak target that flickers in daylight (Visibility, "A faint spot lost when the light is missing", even without rejection) would remove the doubt.
- Missing: the catalogue has no entry for a whole-scene error_value rise of one or two measurements by day, without an obstacle (28/08 12:00-12:45, and the many instants left without a label when the curves stay still). Its cause (exposure, sun, a less sharp image) is not established here.
- `labeling-rules.md`, 3 (a target rejected alone just after a passage is the same event) would extend NORDEND-61 to 22:00 for PR09 at 21:45; the 21:45 image shows the web and no spider, and the web is labelled over that measurement: kept as is.
- The No data labels NORDEND-77, 82, 87 cite JUNGFRAU only; WILDSTRUBEL has measurements at those minutes. The rule ("with two stations only, the common cause is possible") still fits; saying that a third station measures would be more complete.
- Precedents: NORDEND-80 and LENZSPITZE-84 fixed the family and conclusion of the 29/08 camera step; the JUNGFRAU REF_G series pointed to the reference residual for 31/08.
## Difficulties
- No weather for NORDEND (no `weather/` folder), no image quality (old camera): the causes rest on the images alone.
- My scan (`scripts/scan.py`) first printed false gaps (the export times carry seconds); fixed by rounding to the minute.
## Sources read
- `agents/agents/analyst.md`, `instructions/analyst.md`, the request (`work/prompt.txt`, `request.json`), `work/existing.jsonl`, `work/spanning.jsonl`
- `tools/GUIDE.md`, `tools/DATA.md`, `labeling/families.md`, `labeling/labeling-rules.md`, `labeling/reading-rules.md`, `labeling/stations.md` (in full)
- `tools/scripts/data.py` (the loader, for my scripts)
- Tools: `inventory.py`, `overview.py` (range; 27-28/08; `--rigid` on 29/08 and 31/08), `ranges.py --rolling 4 --instants 3`, `summary.py` (15 min tables), `metrics.py`, `stations.py` (NORDEND, JUNGFRAU, WILDSTRUBEL), `quality.py` (no metrics), `thumbnails.py`, `plot.py`, `gif.py`, `precedents.py` (4 queries)
- Raw: `metrics/` last_sync on the gap nights, the images listing around each gap
- My scripts: `scripts/scan.py`, `scripts/thermal_days.py`; outputs in `work/`