Files
aimsight-labeling/auto/analyses/NORDEND_2026-08-23T0515_2026-09-04T0615/analysis.md
T
maximeandClaude Opus 5.5 ddb17c2189 auto: labels requalified under the 5-noise rules (only what stands out from the noise): 32 analyses refined, the 6 paused periods verified, the long labels requalified with the fixes the period analysts found; 1153 -> 961 labels
- 32 analyses: each label kept, rewritten or removed by its analyst (section Requalification of the report; mm magnitudes where the targets have them)
- SNCF01 x3, WILDSTRUBEL 2026-09-01 -> 2026-10-06 x3: verified from their existing labels under the new rules
- long labels (no analysis): auto/analyses/long-labels/<STATION>/; NORDEND-169 and -195 run to 07/11, so NORDEND-283 and -284 (their continuation) are removed
- reading-rules.md, stations.md: the citations of removed or rewritten labels point to their successor, or to [expert]

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-17 23:58:00 +01:00

125 lines
16 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: requalify the 42 labels of this analysis (`work/requalify_input.jsonl`) under the new rules (`labeling-rules.md`, section 1): only what stands out from the noise of the corrected curves, at least one target beyond 5 of its own noises; no label for missing points alone; a thermal label only if the corrected curves keep it; No data only when the station delivers nothing. Not a new reading: the previous verification (kept in `work/rq/analysis.before-requalify.md`) stands. DETECTORS: no.
Result: **35 kept as they are** (id and review field unchanged), **3 rewritten** (NORDEND-66, NORDEND-252, NORDEND-79: right in substance, their description rested on CTR01's missing curve), **4 removed** (NORDEND-53, NORDEND-248, NORDEND-253, NORDEND-83); 38 labels in `labels.jsonl`.
Context (unchanged from the verification): no `data/NORDEND/weather/` folder, so no MeteoSwiss weather (station without a Swiss position): the images are the only witness of the weather. No image quality metrics (old camera). Housing temperature 26.5 to 46.9 °C, pan r = −0.98, scale r = +0.97 against it: the daily camera cycle is thermal. Targets 84 m (PR08, PR16) to 151 m (PR01, PR09) from the camera, 31 to 60 mm per px; the corr noise of a target is 0.004 to 0.007 px, i.e. about 0.15 to 0.35 mm (PR04 0.012 px, Target issue NORDEND-24).
## 1. My labels
How each label was measured (section "Requalification" below): the short labels against a rolling level (±8 measurements, ±2 h, self excluded), at k = 5, in the target's own noise over the batch; the thermal labels per target, hourly median of corr minus the level of the night before the cycle (3 h before its start), in the same noise, the instants of the other short labels left out. References CTR02, CTR03, CTR04 are not counted (they are not targets); PR04 is left out of the thermal check (Target issue NORDEND-24).
### The 35 labels kept as they are
| id | Family | Largest deviation of a target in the label (noises) | mm |
|---|---|---|---|
| NORDEND-51 | Thermal | PR18 corr_x −17.8 (23/08 07h) | ≈ 5 |
| NORDEND-52 | Thermal | PR18 corr_x −23.0 (24/08 08h), PR16 −14, PR18 corr_y +12.8 | ≈ 7 |
| NORDEND-54 | No data | the station delivers nothing at 01:30 | – |
| NORDEND-55 | Thermal | PR18 corr_x −13.4 (25/08 08h) | ≈ 4 |
| NORDEND-56 | Thermal | PR18 corr_x −11.1 (26/08 11h); −9.5 against the following night | ≈ 3.4 |
| NORDEND-57 | Visibility | PR16 corr_x +18.6 (08:30), PR08 +14, PR14 +14, PR06 +13.5 | PR01 y 3.8, PR16 x 3.3 |
| NORDEND-58 | No data | nothing at 01:30 | – |
| NORDEND-246 | Thermal | PR18 corr_x −13.9 (27/08 10h), PR16 −10 | ≈ 4 |
| NORDEND-247 | Visibility | PR08 corr_x −24.5 (21:45), PR06 −15, PR05 −13; PR08 −16 (hourly) around midnight | ≈ 3 |
| NORDEND-61 | Obstruction | PR18 corr_y −13.6 (21:15), PR13 corr_y +13.4 (21:30), PR08 corr_x −24.5 (21:45) | ≈ 3.4 |
| NORDEND-62 | No data | nothing at 01:45 | – |
| NORDEND-64 | No data | nothing 08:00 → 11:45 | – |
| NORDEND-249 | Visibility | PR12 corr_x +7.1, PR03 +5.8, PR11 +5.6 (12:15) | ≈ 1.5 |
| NORDEND-65 | Visibility | PR14 corr_x −64, PR07 −63, PR05 −58, PR13 −56 (12:45) | ≈ 12 |
| NORDEND-67 | Visibility | PR08 corr_x −15.8 (16:15), PR07 −14, PR15 +12.7 | ≈ 2.5 |
| NORDEND-68 | Thermal | PR18 corr_x −12.2 (29/08 07h), PR18 corr_y +10.1 | ≈ 3.7 |
| NORDEND-250 | Displacement | PR15 corr_x +8.6, PR16 +7.7 (12:00) | ≈ 1.5 |
| NORDEND-69 | Visibility | PR07 corr_x +10.0 (18:45), −9.4 (16:15), PR16 +9.4 | ≈ 1.8 |
| NORDEND-70 | Visibility | PR16 corr_x +11.2, PR15 corr_y +10.1 (22:30) | ≈ 2 |
| NORDEND-71 | Visibility | PR14 corr_x +20.5, PR15 +15.8 (00:45) | ≈ 4 |
| NORDEND-72 | Thermal | PR14 corr_x −5.9 (30/08 18h), −5.4 (08h); PR18 corr_y +5.5 (09h) | ≈ 1.2 (marginal) |
| NORDEND-73 | Visibility | PR15 corr_x −12.1 (00:00), PR14 +10.7 (00:15) | ≈ 2.2 |
| NORDEND-74 | Thermal | PR18 corr_y −5.1, PR01 corr_y +5.3 (31/08 14h); PR18 corr_x −6.1 (10h) against the following night | ≈ 1 (marginal) |
| NORDEND-251 | Displacement | PR05 corr_x −12.0 (11:45), PR08 −8.4, PR12 −6 | ≈ 2.4 |
| NORDEND-75 | Visibility | PR14 corr_x −21.2 (13:00), corr_y −15 (13:15), PR16 corr_y +12.7 | ≈ 4 |
| NORDEND-77 | No data | nothing at 01:00 | – |
| NORDEND-78 | Thermal | PR18 corr_x −8.0 (01/09 08h); −10.1 against the following night | ≈ 2.4 |
| NORDEND-80 | Displacement | PR11 corr_x −7.1, PR17 −6.6, PR13 corr_y −6.3, PR08 −6 (15:45) | ≈ 1.7 |
| NORDEND-81 | Thermal | PR18 corr_x −12.9 (02/09 10h), PR16 −9 | ≈ 3.9 |
| NORDEND-82 | No data | nothing at 01:30, 01:45 | – |
| NORDEND-254 | Visibility | PR07 corr_y +8.3 (09:00) | ≈ 1.7 |
| NORDEND-255 | Visibility | PR15 corr_x +10.4, PR16 corr_y +9.8, PR12 +8.8 (13:30); PR18 corr_x +5.7 (13:45) | ≈ 2–3 |
| NORDEND-85 | Visibility | PR14 corr_x +11.3, PR07 +9.6, PR18 +9.1 (16:45) | ≈ 2.8 |
| NORDEND-86 | Obstruction | PR16 corr_x +18.0 (23:00), corr_y −17.4 (23:15), PR15 corr_x −16.6 (23:30) | ≈ 3.2 |
| NORDEND-87 | No data | nothing at 01:15, 01:30 | – |
The seven No data labels all mark instants where no sensor has a measurement (inventory: steps absent for all sensors): they hold under the new rule whatever their length. NORDEND-72 and NORDEND-74 are just above the threshold (hourly medians at 5.3 to 6.1 noises, depending on the reference night); they stay, as the rule asks, but they are the weakest labels of the batch.
### 2026-08-28T14:15Z → 2026-08-28T14:45Z (NORDEND-66, rewritten)
- **Phenomenon of the catalogue**: The sun in the field or behind a spot (CTR01), with a whole-scene error_value rise of one measurement
- **Precedents**: NORDEND-249 (same day, a whole-scene error_value rise of two measurements with small jumps, Visibility); the old description of this label
- **Family**: Visibility, unchanged: the whole image degraded for one measurement in strong sun, no object, no camera step
- **Measured**: at 14:30 the error_value of 21 sensors, the three references included, rises together (+20 to +74 noises; CTR04 0.28, CTR02 0.17, CTR03 0.20, inlier 1 for all), CTR01 rejected; PR18 corr_x +7 noises (+2.1 mm), PR09 −5 (−1.8 mm), PR14 +5 (+1.0 mm); 14:45 back. The old description spoke only of CTR01's missing curve, which is no longer labelled: rewritten on the effect on the other targets. PR12 corr_x is at −6 noises already at 14:15, the opening measurement: strictly the start would be 14:00; the bounds of a label that holds are not moved here.
- **Images**: 14:30 thumbnail (from the verification): rock in full sun around CTR01
- **Confidence**: high on the event (just above threshold), cause of the whole-scene rise not established
### 2026-08-31T14:30Z → 2026-08-31T15:00Z (NORDEND-252, rewritten)
- **Phenomenon of the catalogue**: same as NORDEND-66
- **Precedents**: NORDEND-66 (same signature three days before)
- **Family**: Visibility, unchanged
- **Measured**: at 14:30 (the opening measurement), error_value of 20 sensors up together (+12 to +62 noises), PR18 corr_x +5, PR03 corr_y +7 noises (+2.0 mm); at 14:45 CTR01 rejected, PR18 corr_x +6 noises (+1.8 mm); 15:00 back. Rewritten: the old description rested on CTR01's missing curve and pointed to NORDEND-253, now removed. As for NORDEND-66, the effect starts at the start bound (14:30): strictly the start would be 14:15; left as it is (bounds not moved in a requalification), to be fixed by the user if wanted.
- **Images**: 14:45 thumbnail (from the verification): rock in full sun around CTR01
- **Confidence**: medium: the jumps are just above the threshold
### 2026-09-01T14:45Z → 2026-09-01T15:30Z (NORDEND-79, rewritten)
- **Phenomenon of the catalogue**: The sun in the field or behind a spot (CTR01)
- **Precedents**: NORDEND-66, NORDEND-252
- **Family**: Visibility, unchanged (same episode as the other sunny afternoons of CTR01)
- **Measured**: CTR01 rejected 15:00, 15:15; at 15:00 PR01 corr_y +7 noises (+3.0 mm), at 15:15 PR05 corr_x +7 (+1.3 mm), PR14 +5 (+1.1 mm), PR01 corr_y +6; the error_value of the other sensors stays at its level (only CTR01's rises), all references retained; 15:30 back. Rewritten on the effect on the other targets; the old description cited dates outside the range for its deduced cause.
- **Images**: none (cause deduced, as in the old label)
- **Confidence**: medium: just above threshold, the link between CTR01's sun and the jumps of the others is not established
## 2. What I left without a label
### The removed labels
| id | Bounds | Family | Why |
|---|---|---|---|
| NORDEND-53 | 2026-08-24T11:00Z → 2026-08-25T12:30Z | Obstruction | Missing points only: PR07 rejected almost without interruption under a cobweb thread. PR07 is not a reference, its loss moves nothing else. The deviations inside the range belong to other events: PR01 corr_y +8.9 and PR14 corr_x +6.8 noises at 24/08 15:15 (afternoon agitation of the thermal day, under NORDEND-52), PR18 corr_x −7.6 at 25/08 09:30 and −7 at 08:45 (morning of the thermal cycle NORDEND-55). PR07 is not rejected almost always over the batch (135 rejections out of 1133 measurements, by episodes): not a Target issue. |
| NORDEND-248 | 2026-08-28T07:00Z → 2026-08-29T06:00Z | Thermal displacement | Compensated thermal: the night of 28/08 before the cycle is under the cobweb (NORDEND-247), so the level was taken on the night of 27/08 and on that of 29/08. Max 5.9 noises against 27/08 (PR15 corr_y at 18h, the tail of the afternoon episode NORDEND-67; PR14 corr_x +5.8 at night, a difference of night level, not a cycle), 5.1 against 29/08 (PR15 corr_y 18h again); the daytime remainder of the camera cycle itself is at most 4.4 noises (PR18 corr_y 14h). The label itself said "the corrected curves hardly depart from it". |
| NORDEND-253 | 2026-08-31T15:45Z → 2026-08-31T16:15Z | Visibility | Missing point only: CTR01 rejected at 16:00; no target beyond 5 noises (the error_value of all the sensors rises at 16:00 and 16:15, without effect on corr). |
| NORDEND-83 | 2026-09-03T06:30Z → 2026-09-04T06:15Z | Thermal displacement | Compensated thermal: outside the cobweb thread of NORDEND-86 (kept, it carries the 23:00 deviations of PR15, PR14, PR16 up to 16–21 noises), the largest deviation from the night level is 4.6 noises (CTR01 corr_y) against the night of 03/09, 3.8 against the night of 04/09; PR15's "drop a little" stays under the threshold. |
Consequences on the kept labels: NORDEND-246 and NORDEND-68 now touch a day without a thermal label (28/08), which is right (camera cycle absorbed that day). The descriptions of NORDEND-247 and NORDEND-249 cite only labels that remain (gap NORDEND-64, spider NORDEND-61, CTR04 NORDEND-65). NORDEND-252 cited NORDEND-253 ("next label"): rewritten.
### Unchanged from the verification
- **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: inside the thermal labels, described there.
- **Afternoon jumps under the pose agitation** (23/08 15:15 and 15:45, 24/08 15:15 and 15:45, 26/08 14:45 and 16:15): 2-5 targets at 4-9 noises, error_value not rising: the agitation of the thermal day, inside the thermal labels.
- **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, 15:15, 16:15, 16:30; 01/09 11:15): no target at 5 noises.
- **References alone at 5-8 noises** (23/08 05:15-05:45, 24/08 08:30, 08:45 and 17:00, 27/08 10:45 and 14:30, 28/08 15:30, 02/09 09:00): their corr only, the targets do not follow.
- **Missing points not labelled** (new rule): CTR01's sunny-afternoon rejections and PR07's rejections are labelled only where they come with jumps of the other targets (the labels above).
## Requalification
Commands (from the repository root, outputs in `work/rq/`):
- `python3 tools/scripts/ranges.py data/NORDEND --columns corr_x,corr_y --k 5 --rolling 8 --start 2026-08-23T05:15Z --end 2026-09-04T06:15Z` → `ranges_roll8.txt` (level ±2 h, noise per target over the batch); with `--columns corr_x,corr_y,error_value --instants 1` → `instants_roll8.txt` (who deviates at each instant, the references).
- `python3 scripts/requalify.py work/requalify_input.jsonl work/rq/ranges_roll8.txt` → `requalify.txt` (short labels: target ranges overlapping the label, max in noises; thermal labels: hourly median against the night before, the instants of the short labels left out → `thermal_excl.txt`).
- `python3 scripts/thermal_nights.py … NORDEND-248 2026-08-27T02:45Z NORDEND-248 2026-08-29T03:00Z NORDEND-72 … NORDEND-74 … NORDEND-83 … NORDEND-78 … NORDEND-56 …` → `thermal_nights.txt` (the weak thermal days against the night before and the night after).
- `python3 tools/scripts/ranges.py data/NORDEND --columns corr_x,corr_y --k 5 --min-measurements 4 --start S --end E` on the weak thermal days → `thermal_MM-DD.txt` (window level; confirms that 28/08 and 03/09 keep nothing of their own outside the short labels).
Values: thermal max (noises, night before / night after): 51 17.8; 52 23.0; 55 13.4; 56 11.1 / 9.5; 246 13.9; **248 5.9 / 5.1, cycle itself ≤ 4.4 → removed**; 68 12.2; 72 5.9 / 4.5; 74 5.3 / 6.7; 78 8.0 / 10.1; 81 12.9; **83 4.6 / 3.8 → removed**. Short labels: max 6.0 (252) to 64 (65); **253: none beyond 5 → removed**; **53: nothing of its own → removed**. `validate.py labels.jsonl`: 38 labels, 0 errors.
## 3. The documentation
- `labeling-rules.md`, section 1 and `families.md` ("What is not labeled", "A gap") gave the three decisions directly: missing points alone (NORDEND-53, 253), compensated thermal (NORDEND-248, 83), the CTR01 labels kept only for their effect on the others.
- The rule for thermal labels says "beyond 5 noises from the night level" without saying which noise: I took the target's own noise at the native step over the batch (the `ranges.py` noise) against the hourly median, which is a little lenient for a slow cycle; at the native step the weak days give the same verdicts. It also does not say what to do when the night before is under another event (28/08): I used the nights on either side.
- Requalifying without moving bounds leaves two known bound errors (NORDEND-66, NORDEND-252: the effect on the others is already at the opening measurement); noted above for the user.
- From the verification (still valid): 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, 14:30; 31/08 14:30); the brief PR07 labels are Visibility by consistency, though the boundary text would say Displacement for a jump without a seen nuisance; NORDEND-77, 82, 87 cite JUNGFRAU only while WILDSTRUBEL measures at those minutes.
## Difficulties
- No weather for NORDEND (no `weather/` folder), no image quality (old camera).
- The thermal threshold depends on the reference night chosen for the two weakest days kept (NORDEND-72, NORDEND-74): both nights are given.
## Sources read
- `instructions/analyst.md`, the request, `work/requalify_input.jsonl`, the previous `analysis.md` and `report.md` (now in `work/rq/`)
- `labeling/labeling-rules.md`, `labeling/families.md` (in full), `tools/GUIDE.md` (`ranges.py`, reading order), `tools/DATA.md` (in full); `labeling/reading-rules.md`, `labeling/stations.md` read in full during the verification
- Tools: `ranges.py` (k 5, `--rolling 8`, `--instants 1`, windows of the weak thermal days), `inventory.py` (distances, mm per px), `validate.py`
- My scripts: `scripts/requalify.py`, `scripts/thermal_nights.py` (new), `scripts/scan.py`, `scripts/thermal_days.py` (verification)