- 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>
17 KiB
Labels: WILDSTRUBEL, 2026-09-01T00:00Z → 2026-09-13T00:00Z (UTC)
Task: verify the 67 existing labels of the range (work/existing.jsonl, all reviewed by the user) against the rules of 2026-11-17 (labeling-rules.md, section 1: a label only where a target deviates by more than 5 of its own noises on corr_x/corr_y; a thermal cycle only when the corrected curves keep it; a missing point not for itself). Result: 50 labels in labels.jsonl: 41 kept as they are (with their review), 9 corrected, 17 removed, 0 added. The 3 spanning labels (WILDSTRUBEL-46, -69, -70) are not in labels.jsonl; checked within the range, nothing wrong found.
Data over the range: measurements from 01/09 00:00 to 09/09 21:45 (15 min), then nothing until the end of the range (station out of energy, spanning No data WILDSTRUBEL-69). Two single missing instants (05/09 02:15, 06/09 02:00). No image-quality metrics (old camera: peak…snr empty), so no quality.py. No weather: data/WILDSTRUBEL/weather/ does not exist (no MeteoSwiss station attached). Own noises at the native step (window, accepted measurements): corr_x / corr_y in px: PA2_2 0.019/0.018, PA3_2 0.019/0.018, PA_9 0.030/0.026, PB1_2 0.023/0.017, PB2_2 0.020/0.015, PB4 0.017/0.015, PB_5 0.016/0.015. So 5 noises ≈ 0.08–0.15 px, i.e. about 0.5 mm (PB1_2, 13.8 m) to 2 mm (PA_9, 39.1 m).
Method of the verdict: ranges.py --columns corr_x,corr_y --k 5 with --rolling 4 and --rolling 8 on the targets (work/ranges_roll4.txt, work/ranges_roll8t.txt, list of instants --instants 1): a short label is kept if at least one target is beyond 5 noises at one of its instants with either rolling window; my script scripts/check_labels.py (work/check_existing.txt) gives, for each label, the jump against the mean of the measurement before and after, and the deviation from the preceding night level. The thermal labels were checked with scripts/thermal_check.py (work/thermal_check.txt): hourly median of each target's corr minus its night level, in noises, next to the tilt and the housing temperature.
1. My labels
One line per old id, then the detail by group.
| Old id | Verdict | Evidence |
|---|---|---|
| WILDSTRUBEL-1 | kept | R3 rejected every day from about 04:00 to 15:30 (inventory: 28+18, 6+10+27, 8+38, 43+4, 44+2, 43 rejected measurements by day, almost none at night); families.md: a reference rejected every day is a Target issue even without effect on the targets |
| WILDSTRUBEL-2 | removed | thermal camera cycle absorbed: hourly corr of the targets other than PB1_2 at most +5/+6 n at 05h (PA2_2, PA3_2), not following the tilt (still +3 n at 11h when the tilt is at its trough); the label itself says "only barely more than the noise" |
| WILDSTRUBEL-3 | corrected | day agitation real (instants beyond 5 n by day: 01/09 12:00 PB2_2 −5, 12:45 PA_9 −5, 13:30 PA2_2 +7, PB1_2 +5, PB_5 +5; none at night); description pointed to the removed thermal label |
| WILDSTRUBEL-4 | kept | PB1_2 corr_x hourly −6 to −10 n from 06h to 14h, back at 16h |
| WILDSTRUBEL-5 | removed | 11:00, 11:15: no target beyond 5 n (rolling 4 and 8); largest jump PA2_2 corr_y +5.1 n against its neighbours only, under the threshold with a local level; rejections only |
| WILDSTRUBEL-6 | kept | PA3_2 14:45: −439 n x, −1967 n y (false point, ~0.1 m in mm, absurd) |
| WILDSTRUBEL-7 | kept | 16:30: PB2_2 x +15 n, y −7 n; PB4 y −14 n; PB_5 y −6 n |
| WILDSTRUBEL-8 | removed | thermal absorbed: hourly corr ≤ +4/+5 n outside PB1_2 (PB1_2 under its own label) |
| WILDSTRUBEL-9 | corrected | day agitation (02/09 04:30 PA2_2 +5, 05:00 PA3_2 +5/+6, 06:30 PB2_2 y +6, 10:00 PB2_2 x +7, 13:15/13:30 PA2_2 y +6/−11); thermal pointer removed |
| WILDSTRUBEL-10 | kept | PB1_2 corr_x hourly −6 to −10 n by day |
| WILDSTRUBEL-11 | kept | 07:30: PB4 x +7, y −9; PB_5 x +6; PA3_2 y −5 |
| WILDSTRUBEL-12 | kept | 11:15: PB4 x +9, PB_5 x +6 |
| WILDSTRUBEL-13 | kept | 12:00: PB1_2 corr_y +5.1 n (rolling 4), +6.0 n against its neighbours: at the threshold, kept |
| WILDSTRUBEL-14 | removed | 12:30: nothing beyond 5 n with a local level (PB_5 x −5.6, PA2_2 y +5.2 only against the two neighbours, which are agitated by day) |
| WILDSTRUBEL-15 | kept | PB1_2 corr_x night level drops by −5.1 n on the night of 03/09, stays −2.7 n lower on 04 and 05/09, back above on 06/09 (days.py); day hourly −16 n on 03/09; 06/09 02:45 spike −23 n; at 03:00 it is still at −4 n (under the threshold), back at 03:15: the end at 03:00 holds |
| WILDSTRUBEL-16 | removed | thermal absorbed: outside PB1_2, hourly ≤ +4 n |
| WILDSTRUBEL-17 | corrected | day agitation (03/09 05:00 PA3_2 +8, PB_5 +6; 13:15 PA2_2 −5; 13:30 PA2_2 y −6); thermal pointer removed |
| WILDSTRUBEL-18 | kept | PB1_2 rejected most of the day, corr_x hourly −16/−17 n |
| WILDSTRUBEL-19 | kept | 07:30: PB4 x +10/+13, PB_5 x +7, PB2_2 x +6 |
| WILDSTRUBEL-20 | kept | 08:15: PB4 x +10, PB2_2 x +6, PB_5 x +6 |
| WILDSTRUBEL-21 | removed | PA2_2 rejected alone at 21:45; no other target moves (0 instant beyond 5 n); a missing point only |
| WILDSTRUBEL-22 | removed | thermal absorbed: outside PB1_2, hourly ≤ +6 n at 05h and +10 n at 06h on PA3_2 only, at the hour when the compensation lost all its references (05:45, 06:00), not the tilt wave (+3 n at 11h at the tilt trough) |
| WILDSTRUBEL-23 | corrected | day agitation and sunrise base losses (04/09 05:45 PB_5 +7, 06:00 PA3_2 +7, PA2_2 +6, 06:30 PA3_2 −7, 07:45 PA_9 +5); thermal pointer removed |
| WILDSTRUBEL-24 | kept | PB1_2 corr_x −7 n at 07h, rejected late morning |
| WILDSTRUBEL-25 | kept | PA3_2 06:45: +429 n x, +1683 n y (false point) |
| WILDSTRUBEL-26 | kept | 10:15: PB4 x +8, PB_5 x +6 |
| WILDSTRUBEL-27 | removed | 11:45–12:15: no target beyond 5 n (rolling 4 and 8); rejections only |
| WILDSTRUBEL-28 | removed | 14:00: no target beyond 5 n; the hidden targets are rejected |
| WILDSTRUBEL-29 | removed | PB4 rejected alone at 15:15; nothing beyond 5 n |
| WILDSTRUBEL-30 | kept | 02:15 absent for all sensors, image absent too (inventory: gap 02:00→02:30): No data holds at any rate |
| WILDSTRUBEL-31 | kept | PA2_2 04:30 +1747 n; PB4 06:00 y −30 n, x +14 n; PB2_2 04:00 +8 n |
| WILDSTRUBEL-32 | removed | thermal absorbed (the label says "only barely more than the noise") |
| WILDSTRUBEL-33 | corrected | day agitation; thermal pointer removed, and the 09:15 passage it pointed to (WILDSTRUBEL-35) is removed: stated in the text |
| WILDSTRUBEL-34 | kept | PB1_2 corr_x hourly −6 to −8 n, rejected late morning |
| WILDSTRUBEL-35 | removed | PA_9 rejected alone at 09:15, nothing else beyond 5 n |
| WILDSTRUBEL-36 | kept | 12:00: PA2_2 corr_x +8 n (rolling 4), +6 (rolling 8) |
| WILDSTRUBEL-37 | removed | parked car at night: hidden targets rejected, the compensation and the other targets do not move (0 instant beyond 5 n from 23:15 to 00:15) |
| WILDSTRUBEL-38 | kept | 02:00 absent for all sensors, image absent |
| WILDSTRUBEL-39 | removed | thermal absorbed: hourly ≤ +5/+6 n (PA_9 +6 at 04h, PA3_2 +5 at 08–09h), no tilt shape |
| WILDSTRUBEL-40 | kept | PB1_2 corr_x −6 n at 08h, corr_y +4/+5 n |
| WILDSTRUBEL-41 | corrected | day agitation (06/09 05:15 PA3_2 −6, 07:00/07:15 PB4 y −6/+8, 13:45 PB2_2 y +7); thermal pointer removed |
| WILDSTRUBEL-42 | kept | 06:00: PA3_2 x +7, PB4 x +6 |
| WILDSTRUBEL-43 | kept | 09:30: PB4 x +6 (rolling 4), +8 (rolling 8) |
| WILDSTRUBEL-44 | kept | 12:15: PA_9 corr_y +6 n (rolling 8), +7 n against its neighbours: at the threshold, kept |
| WILDSTRUBEL-45 | kept | PA_9 14:30: −556 n x, +303 n y (false point) |
| WILDSTRUBEL-47 | removed | the label itself says "the corrected curves do not show this wave" |
| WILDSTRUBEL-48 | corrected | day agitation (07/09 06:30 PA_9 −6, 07:15 PB1_2 y −8, PA_9 +6; 10:15 PA3_2 −6/−7); thermal pointer removed |
| WILDSTRUBEL-49 | kept | PA_8 rejected intermittently; PA_9 +6/+7 n at 07:15 and 12:30 while PA_8 is out |
| WILDSTRUBEL-50 | kept | PB1_2 corr_x hourly −6 to −8 n |
| WILDSTRUBEL-51 | kept | 09:45: PB4 x +6 |
| WILDSTRUBEL-52 | kept | 14:30: PB_5 x −7/−6 |
| WILDSTRUBEL-53 | kept | PB1_2 02:45: x −12 n, y +6 n, back at 03:15 |
| WILDSTRUBEL-54 | corrected | day agitation and sunrise base losses (08/09 06:45 PB4 y +7, 07:45 PA2_2 −6); thermal pointer removed |
| WILDSTRUBEL-55 | removed | the label itself says "the corrected curves do not show this wave" |
| WILDSTRUBEL-56 | kept | PA_8 out almost all day: PA_9 hourly +10 n at 08h, PA2_2 +7, PA3_2 +7 |
| WILDSTRUBEL-57 | kept | PB1_2 corr_y hourly −6 to −9 n, corr_x −4/−6 n |
| WILDSTRUBEL-58 | kept | 08:30: PA2_2, PA_9, PB_5 +5 (rolling 4), PA_9 +7 (rolling 8) |
| WILDSTRUBEL-59 | kept | 09:00: PA2_2 x +6 |
| WILDSTRUBEL-60 | kept | 13:30: PA2_2 x −6 (rolling 4), −6.3 against its neighbours |
| WILDSTRUBEL-61 | corrected | sunrise base losses and agitation (09/09 04:45 PB4 −8, 06:30 PA2_2 +8, PA3_2 +7/+8, PB4 y +7); thermal pointer removed |
| WILDSTRUBEL-62 | removed | the label itself says "the corrected curves do not show this wave" |
| WILDSTRUBEL-63 | kept | PA_8 out: PA_9 hourly +9 n at 06h, PA3_2 +10, PA2_2 +7/+8 |
| WILDSTRUBEL-64 | kept | PB1_2 corr_x hourly −6 to −11 n, not back by the evening |
| WILDSTRUBEL-65 | kept | 09:45, no reference retained: PB_5 x +6, y −5 (rolling 8), +5 against its neighbours |
| WILDSTRUBEL-66 | kept | PA_9 10:15: +802 n x, −783 n y (false point); PB2_2 +5 |
| WILDSTRUBEL-67 | kept | 12:15: PB1_2 corr_x −5 n (rolling 8): at the threshold, kept by the same criterion as the others |
| WILDSTRUBEL-68 | kept | 12:45: PB_5 x +6, PA2_2 x +5 |
The thermal camera cycle, removed every day (WILDSTRUBEL-2, -8, -16, -22, -32, -39, -47, -55, -62)
- Phenomenon of the catalogue: "The camera moves through a thermal effect" (labelled only when the curves keep it beyond the noise).
- Precedents: the rule of 2026-11-17 (
labeling-rules.mdsection 1, "Thermal movement is labeled only when the corrected curves keep it"); the labels' own texts (WILDSTRUBEL-47, -55, -62: "the corrected curves of the targets do not show this wave"; -2, -32: "only barely more than the noise"). The search for precedents timed out (see Difficulties). - Family: none kept.
- Measured: tilt +45 to +72 mdeg above its night level from 04h to 07h, −19 to −34 mdeg at 10–11h, back at 13h, every day; housing temperature from about 36 °C at night to 48–50 °C. Hourly median of the corr of the targets minus the night level (
work/thermal_check.txt): outside PB1_2 (its own labels), within ±4 n most hours; +5 to +7 n only on PA2_2/PA3_2 at the sunrise hour (05–06h) when the compensation loses PA_8, R2 and sometimes all references (sunrise, under the daily Visibility labels), and on PA_9/PA2_2/PA3_2 on 08 and 09/09 when PA_8 is out (WILDSTRUBEL-56, -63). The small daytime positive offset of PA2_2 and PA3_2 (+2 to +4 n) does not follow the tilt: it stays the same at 07h (tilt +60) and at 11h (tilt −30). In mm: a few tenths of a mm on the PA targets at most. - Images: none needed (no effect to explain).
- Confidence: high.
The daily Visibility labels (heat shimmer by day), corrected (WILDSTRUBEL-3, -9, -17, -23, -33, -41, -48, -54, -61)
- Phenomenon of the catalogue: "Atmospheric turbulence" plus "The sun in the field or behind a spot" at sunrise.
- Precedents: the labels themselves (reviewed);
families.mdboundary Thermal displacement or Visibility. - Family: Visibility, unchanged.
- Measured: by day, single-point deviations beyond 5 n on several targets every day (rolling 4: 79 instants in the batch, all by day except the PB1_2 night spikes of 03/09 02:45, 06/09 02:45 and 08/09 02:45); at night none. In mm on 02/09: ±0.5 to 1 mm on most targets, up to about 1.5 mm on PA_9 (farthest, 39 m), against about 0.2 mm at night (figure
day_0209_corr_mm.png). - Change: only the text: they pointed to "a separate label, thermal movement of the camera", which is removed; now "the camera moves in the morning through a thermal effect and the compensation absorbs it: the corrected curves do not keep its wave beyond the noise (not labelled)". WILDSTRUBEL-33 also pointed to the removed passage of 09:15 (WILDSTRUBEL-35): now stated without a separate label. Bounds and the rest of the text unchanged.
- Confidence: high on the change.
Obstruction labels removed (WILDSTRUBEL-5, -14, -21, -27, -28, -29, -35, -37)
- Phenomenon of the catalogue: "A pass-by hides the targets for the length of one measurement" / "A vehicle that stays in front of the targets".
- Measured: the hidden targets are rejected (missing points); the other targets stay within 5 noises at every instant of the label with both rolling windows; largest values against the two neighbours: -5 PA2_2 y +5.1 n, PB1_2 x +4.6 n; -14 PB_5 x −5.6 n, PA2_2 y +5.2 n (daytime neighbours are themselves agitated; local level under 5 n); -27 PA3_2 x +5.6 n (12:15, rolling ≤ 5 n); -28 PB1_2 y −5.5 n (inside the PB1_2 daylight label); -21, -29, -35, -37 nothing above 3 n.
- Confidence: high for -21, -29, -35, -37 (nothing at all); medium for -5, -14, -27, -28 (one value at 5–6 n against the two neighbours only, under 5 n against a local level of ±1 to 2 h).
Kept labels
The 41 others hold on the curves (table above), with the bounds of the expert's review. For the borderline passages (WILDSTRUBEL-13, -44, -60, -65, -67), the deviation reaches 5 to 6 noises with one of the two rolling windows: kept by the rule.
2. What I left without a label
- Removed old labels, with their deviation: see the table (thermal: hourly ≤ +4 n outside the sunrise base losses and the PA_8 days; passages: ≤ 5 n with a local level).
- 02/09 13:15–13:30, PA2_2 corr_y +6 n then −11 n (about +1 mm then −1.8 mm), alone, no rejection; raw y moves by half a pixel and comes back; nothing on the 13:15/13:30/13:45 thumbnails (the spot next to the pole, unchanged). The largest point of the day's agitation; left under the daily Visibility label WILDSTRUBEL-9 (heat shimmer), not a separate label.
- 02/09 10:00 PB2_2 corr_x +7 n, 06/09 13:45 PB2_2 corr_y +7 n: alone, no rejection, nothing on the thumbnails; part of the day's agitation (under WILDSTRUBEL-9 and -41).
- Sunrise base losses (no or three references retained, e.g. 03/09 05:00, 04/09 05:45–06:00, 06/09 07:00–07:15, 07/09 07:15, 08/09 06:45, 09/09 04:45, 06:30): single-point jumps of 5 to 8 n, sometimes up to 10 n; already described inside the daily Visibility labels, as the reviewed labels do; not split.
- PB1_2 on 03/09 at 02:45: +5.7 n with rolling 4, but 0.041 against 0.048 just before: not a deviation, the window median is pulled by the drop of 03:30; the start of WILDSTRUBEL-15 at 02:45 is right.
- Nothing beyond 5 n at night outside PB1_2. From 09/09 21:45 to the end of the range: no measurement (spanning WILDSTRUBEL-69).
3. The documentation
labeling-rules.mdsection 1 (rules of 2026-11-17) settled the thermal and the passage removals; theranges.py --rollingreference in the rule was the right tool for daytime short events: against the window median, the day agitation and the thermal offsets make every daytime point "out of level".- What was missing: how to judge a one-point passage by day when the day agitation itself reaches 5 to 7 noises; I used "beyond 5 n with a rolling local level (±1 h or ±2 h)" and kept the borderline cases. A sentence on which rolling window to use for 15-min stations would make this reproducible.
- The thermal labels of this station were written under D-117 (camera cycle labelled every day); their own texts already said the corrected curves do not keep the wave, consistent with removal under the new rule.
- WILDSTRUBEL-53 (Obstruction, cause "not established, not seen") sits against
families.md("a jump then the return … without an established cause: Displacement"); kept because its description deduces the cause from a repeated signature (PB1_2 spikes at 02:45–03:00 on 03, 06 and 08/09, same hour, nothing visible on the night thumbnails); to be weighed by the user. stations.md(WILDSTRUBEL: road, six references, R3 invisible by day) matched the data.- Spanning labels: WILDSTRUBEL-46 (R3 Target issue from 07/09): R3 rejected 07, 08, 09/09 by day (24+23, 18+28, 4+43): holds within the range. WILDSTRUBEL-69 (No data 09/09 21:45 → 22/09 08:45): no measurement, image or pose after 09/09 21:45 in the range: holds. WILDSTRUBEL-70 (Displacement hidden in the outage): within the range it is only the outage; nothing to contradict.
Difficulties
precedents.pydid not answer within the call (the Ollama container was starting); the verification rests on the rules and on the reviewed labels themselves.- No weather folder for the station.
- The inline guide asks for one command per call without loops; I used two small scripts in
scripts/as the analyst instructions allow.
Sources read
instructions/analyst.md, tools/GUIDE.md, tools/DATA.md, labeling/families.md, labeling/labeling-rules.md, labeling/reading-rules.md, labeling/stations.md, work/existing.jsonl, work/spanning.jsonl, request.json; data of data/WILDSTRUBEL/ through inventory.py, ranges.py (default, --rolling 4, --rolling 8, --instants), days.py, summary.py, plot.py, thumbnails.py, and my scripts scripts/check_labels.py, scripts/thermal_check.py (reading through tools/scripts/data.py); thumbnails of 06/09 02:30–03:00, 08/09 02:30–03:00 (PB1_2, PB2_2), 02/09 10:00, 13:15–13:45 and 06/09 13:45 (PA2_2, PB2_2, PA3_2).