- 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>
16 KiB
Labels: SNCF01, 2026-10-14T06:45Z → 2026-10-22T10:30Z (UTC)
Task: verify the 32 existing labels of the range (work/existing.jsonl) under the rules of 2026-11-17, keep / correct / remove, and add what they miss. Result: 21 kept, 7 corrected, 4 removed, 1 added → 29 labels in labels.jsonl (built by scripts/build_labels.py, validated 0 errors).
Context: a masonry railway viaduct seen across the tracks, 24 targets at 16 to 21 m (0 m / no factor for CV.1, no mm for it), three references BG.3.1, R1, R2 (no distance). 15 min cadence, hourly IRON images on the hour. No weather (no weather/ folder: station without a Swiss position). No temperature: the metrics/ files exist (BusCurrent, BusVoltage, Power, ShuntVoltage) but are empty over the window (metrics.py: "No value in the requested window"). No image quality columns (older camera): quality.py not applicable.
Own noises used (whole window, median of successive differences, overview.py/ranges.py): P targets 0.006–0.013 px in both axes, BG.2/BG.3/F4.5/F4.6 ≈ 0.011–0.013, CV.2 0.03, CV.4 0.023, CV.6 0.04, CV.9 0.08, F4.2 0.016–0.02. Threshold 5 noises: ≈ 0.03–0.065 px for the P targets (≈ 0.2–0.4 mm).
Verdict per existing label
| id | Verdict | Evidence |
|---|---|---|
| SNCF01-38 | corrected (bounds, description) | First target beyond 5 noises on 14/10 at 09:45 (CV.2 cx −6.3, F4.5 cx −5.5; ranges.py --rolling 4), last 15:15 (CV.2 cx −5.7), CV.2 rejected 16:15 → 09:30 → 16:30 (was 06:45 → 16:45: the 06:45 start was only the range edge; 06:45–09:30 nothing beyond 5 noises). Description re-read on the day: sunny image at 13:00, the targets that jump that day. |
| SNCF01-39 | kept | P7 cx +24.8 noises at 15:15, beyond 5 noises 09:00→20:45; P6 cy +16.8; P1 cx −11; pose back overnight (days.py --bounds). |
| SNCF01-40 | corrected (description) | 14/10 10:45: CV.2, CV.4, CV.8 (and F4.3, usual) rejected; CV.6 cx −7.5, F4.2 cy −7.2, F4.6 cy −8.4 noises; BG.3.1 error +80 noises (0.656, under 0.7) not said in the old description: a reference touched explains why untouched targets jump. Bounds 10:30→11:00 right. |
| SNCF01-41 | kept | 01:45→02:15 and 02:15→02:45 two one-step gaps separated by one measurement = one No data (rule section 3). |
| SNCF01-42 | kept | 15/10 05:15→05:30 raw y of all sensors +0.25–0.35 px together (tilt step); corr_y P4 −0.06 px (−7 noises, −0.4 mm), P5 −0.045 (−6.4), P7 −0.047 (−7), CV.1 −0.05; R1 −0.02, R2 +0.027; levels kept the following hours. |
| SNCF01-43 | kept | 15/10 first exceedance 09:00 (6 sensors 5–7 noises), last 15:30; start 08:15 is two steps early (onset of a daytime agitation, F4.3 rejected 08:15–08:45 as usual): within uncertainty, not worth losing the review. |
| SNCF01-44 | kept | P7 cx +27.6 at 15:45, P5 cx +12.7, P6 cx +12.2, beyond 09:00 → gap 22:45. |
| SNCF01-45 | kept | 20 steps absent 22:45→04:00, no image 22:00→04:15. |
| SNCF01-46 | kept | P7 cx +16.6 at 13:00 (07:30→17:30), P1 cx −10.3, P5 cx +7.7; pose returns all night. |
| SNCF01-47 | kept | 16/10 exceedances 09:15, 10:45, 11:15, 12:30; CV.4 rejected 08:45 → start 08:15 one step early: kept. |
| SNCF01-48 | kept | CV.4 cx −129 noises (−3 px, −26 mm), cy +302 (+7 px) at 11:45, CV.2 rejected, back at 12:00. |
| SNCF01-49, -50 | kept | one step absent each, 13:45 and 14:30. |
| SNCF01-51 | corrected (description) | 16/10 16:00: all P are rejected (inventory), so "and some P" was wrong; the untouched targets that jump are BG.2 (cy −61 noises), BG.3 (−83), F4.5 (−55), plus the reference R2 (cx −32). Train seen on the 16:00 image (re-read: a blurred train across the lower scene). |
| SNCF01-52 | kept | P7 cx +12.9 at 14:15 (10:30→18:15), P5 cy +15.1, P6 cx +7.4, P4 cy +10.9. |
| SNCF01-53 | kept | first 08:15 (CV.4 cy −16.9), last ~12:15, CV.2 rejected 16:30. |
| SNCF01-54 | kept | CV.2 cx +168, cy +431 noises at 10:00 (+5.2/+12.4 px ≈ 32/77 mm); seen on the 10:00 image per its text. |
| SNCF01-55 | removed | Only F4.2 rejections (missing points) and an error_value rise; F4.2 accepted night points within ≈2.5 own noises, P2 within 3–4, no target beyond 5 noises at night 17→18/10 (ranges.py --rolling 4, whole window). |
| SNCF01-56 | kept | P7 cx +16.9 at 15:15, P5 cy +17.7 at 14:45, P6 cx +9.6, back in the evening. (Minor wording "the evening and the evening".) |
| SNCF01-57 | corrected (start) | 18/10 08:00 is itself touched: BG.3 cy −5.8, CV.4 −5.1, F4.6 −5.7, P2 −5.3, P4 −5.9 noises, back at 08:15 (sun arriving on the wall, image 08:00) → start 07:45 (rule: start = measurement before the first touched point). |
| SNCF01-58 | removed | Same as 55, night 18→19/10: F4.2 rejected at times, accepted points ≈2.5 noises; P2 deviations 17:00–19:00 are the evening of the thermal cycle (covered by 56). |
| SNCF01-59 | kept | 18/10 20:15: P3 cy −82 noises (−0.77 px, −4.8 mm), BG.2 −56, BG.3 −79, F4.5 −50, R2 −25; "some P" = P3, true; deduced (no image at 20:15). |
| SNCF01-60 | corrected (family Thermal → Displacement, description) | See section 1. |
| SNCF01-61, -63 | kept | 4 steps absent 09:00→10:15; 1 step absent 20:45→21:15. |
| SNCF01-62 | removed | One F4.2 rejection at 13:45 (missing point); its neighbours and all targets within ±1.6 noises. |
| SNCF01-64 | corrected (description) | 20/10 05:45: all P rejected (P3 too), so "some P" was wrong; the untouched ones are BG.2, BG.3, F4.5, F4.6 (cy −46 noises), and R2 (−35). It falls inside the new R1 episode, said. |
| SNCF01-65 | removed | F4.2 rejected 07:00–09:00 and CV.6 all night (CV.6 is a Target issue, spanning SNCF01-6): missing points; P2, P3 within 3.4 noises. The real anomaly of that morning (R1, 04:45→08:00) is a different event, added. |
| SNCF01-66 | kept | 12:45→21/10 10:00, two one-step gaps then 80 steps; isolated measurements at 13:15 and 13:45 do not make a resumption. |
| SNCF01-67 | kept | 21/10 exceedances 12:30, 13:00 (5 targets), 14:00, 14:30, 15:15. |
| SNCF01-68 | corrected (description) | "the compensation absorbs it almost entirely" is contradicted per target: corr_x of P5 dips from −0.050 (12:00) to −0.105 px (18:00), ≈7.6 noises (≈0.35 mm), P7 from −0.395 to −0.441 (≈6.4 noises), with the evening tilt rise; back at the 20:15 measurement. Bounds kept. |
| SNCF01-69 | kept | 19:30→20:15 and 20:15→22/10 10:30; single isolated measurement at 20:15 (P5–P7 cx +5–9 noises against a rolling level spanning the gaps: not read as an anomaly). |
1. My labels
2026-10-14T09:30Z → 2026-10-14T16:30Z (SNCF01-38, corrected)
- Phenomenon of the catalogue: Atmospheric turbulence (heat shimmer).
- Precedents: the same station's 13/10 and 15–18/10 day labels (same signature every sunny day); the precedents tool did not answer (Difficulties).
- Family: Visibility: fast one-measurement noise of the low targets, only by sunny day, night calm; not thermal (no slow movement), not Obstruction (many targets, whole day).
- Measured:
ranges.py --rolling 4: 14/10 targets beyond 5 noises at 09:45, 10:15, 10:30, 10:45, 11:00, 11:30, 12:00, 12:15, 12:45–13:15, 13:30–14:15, 14:30, 15:00, 15:15 (CV.4 up to 7–10, F4.3 up to 10, CV.6 7.5, BG.2 8.3, F4.6 8.4); CV.2 rejected 10:45, 11:45, 13:30, 16:15. Nothing 06:45–09:30. - Images: 13:00, sun with hard shadows on the wall.
- Confidence: medium (cause deduced; onset of an agitation is fuzzy by one or two steps).
2026-10-14T10:30Z → 2026-10-14T11:00Z (SNCF01-40, corrected)
- Phenomenon: A pass-by hides the targets for the length of one measurement.
- Precedents: SNCF01-54 (object seen at the bottom of the scene).
- Family: Obstruction: several low targets and a reference hit at one measurement, all back at the next.
- Measured: 10:45: CV.2, CV.4, CV.8 rejected; BG.3.1 error 0.656 (level 0.44, threshold 0.7); CV.6 cx −7.5, F4.2 cy −7.2, F4.6 cy −8.4 noises; 11:00 back.
- Images: none at 10:45 (hourly). Confidence: medium.
2026-10-16T15:45Z → 2026-10-16T16:15Z (SNCF01-51, corrected)
- Phenomenon: pass-by (train), with both references BG.3.1 and R1 rejected (no pose).
- Family: Obstruction, seen.
- Measured: 16:00, 22 sensors error beyond 5 MAD; BG.2 cx −26 / cy −61, BG.3 −34 / −83, F4.5 −20 / −55 noises (BG.3 −1.05 px ≈ −6.3 mm), R2 −32/−22; all P, CV, F4.1–F4.4, F4.6 rejected.
- Images: 16:00, a train blurred across the whole lower scene. Confidence: high.
2026-10-18T07:45Z → 2026-10-18T15:45Z (SNCF01-57, corrected start)
- Phenomenon: heat shimmer (sunny day).
- Family: Visibility.
- Measured: 08:00 five targets beyond 5 noises in corr_y (BG.3, CV.4, F4.6, P2, P4), back at 08:15; later 09:00–15:15 one-measurement jumps (CV.4 up to 15.6 noises at 11:45).
- Images: 08:00, the sun arrives on the arch (shadow edge across the wall). Confidence: medium.
2026-10-19T07:15Z → 2026-10-20T12:45Z (SNCF01-60, corrected to Displacement)
- Phenomenon of the catalogue: "The base changes / the camera tilts", closest: "A step of the camera that the compensation does not erase" mixed with "targets move apart in the image"; not a daily thermal cycle.
- Precedents: spanning SNCF01-7 (month-long drift, Displacement, reviewed): the scene slowly deforming in the image while the camera tilts; this day is its largest single step, labelled apart (rule 1, "what falls under a long label").
- Family: Displacement rather than Thermal displacement: the boundary of
families.mdasks that a thermal movement come back at equal temperature; here no temperature exists, 19/10 is overcast (no sun on the wall at 08:00 or 13:00, unlike every other cycle day), the pose (pan −0.007°, tilt −0.004°, scale +2·10⁻⁴) does not come back before the outage, and the targets do not return either. - Measured:
overview.py --rigid 2026-10-19T04:00Z: overall movement tx −0.35 px, ty −0.24 px by 15:00 then flat until 20/10 13:00; residuals of the targets themselves: P2 dx −0.10 px, P5 dx +0.07–0.10, P7 dx +0.04–0.07, P5 dy +0.04–0.05 px: targets move relative to each other. corr_x: P2 from ≈0.22 to ≈0.09 px (≈ −11 noises, −0.8 mm), P5 +0.08 px (+11.7 noises), P6 +11.5, P7 +9.1 (night 20/10 vs night 19/10,days.py). sigma0 beyond 5 noises from 19/10 10:30, R1 error elevated from 16:45 (ranges.pyon error_value). On 21/10 P2 ≈0.10 and P5 ≈ −0.055 px: still near the new levels. - Images: 19/10 08:00 and 13:00: overcast, soft image, no sun. Confidence: medium (the shift is certain; its cause and whether the structure moves are not established).
2026-10-20T04:45Z → 2026-10-20T08:00Z (added)
- Phenomenon of the catalogue: "A reference loses its brightness, without an obstacle" / "A reference dropped: the base changes" (family of the cause).
- Precedents: FLORIDA-66 and FLORIDA-71 (Visibility: a reference weakens at night, its measured position slides against the targets, the pose follows it; FLORIDA is SNCF01's neighbour), read by grep in
labeling/auto/because the precedents tool did not answer. - Family: Visibility: reference less well seen, nothing in front, in a night of poor visibility for faint targets; not Obstruction (no object, spot whole on the thumbnails), not Displacement (comes back).
- Measured: R1 error 0.274 → 0.318 at 05:00 (+25 own noises), elevated until ≈09:30;
overview.py --rigid 2026-10-20T04:00Z: R1 dy −0.05 → −0.097 px at 05:30, back to −0.008 at 08:00, targets' own residuals ≈0. corr_y P7 −0.057 → 0.000 px at 05:30 (+8.6 noises, +0.39 mm), P6 0.142 → 0.19 (+7.4, +0.32 mm), P5 0.106 → 0.142 (+5.1), R2 +0.035; back at 08:00 (P7 −0.049, P6 0.156). The train of 05:45 inside (SNCF01-64). - Images: thumbnails R1/R2/P7 at 19/10 05:00 (witness), 20/10 05:00, 06:00, 07:00, 09:00 (
--fixed): spot of R1 in place, whole, nothing in front. - Confidence: medium (the effect is clear; "visibility" as cause is deduced from the night's other faint targets).
2026-10-20T05:30Z → 2026-10-20T06:00Z (SNCF01-64, corrected)
- Family: Obstruction, deduced (train signature). Measured: 05:45: BG.2 cy −56, BG.3 −74, F4.5 −47, F4.6 −46 noises, R2 −35; every P, CV, F4.1–F4.4 rejected, BG.3.1 and R1 rejected (no pose). Images: none at 05:45. Confidence: high.
2026-10-21T10:00Z → 2026-10-21T19:30Z (SNCF01-68, corrected description)
- Phenomenon: The camera moves through a thermal effect (sunny 21/10, image 13:00).
- Family: Thermal displacement, kept: pose cycle of the day with partial remainder on P5/P7.
- Measured: see table. Confidence: medium (no temperature, cut by outages on both sides).
The 21 kept labels: evidence in the table above.
2. What I left without a label
- F4.2 night rejections 17→18/10, 18→19/10, 19/10 13:45, 20/10 07:00–09:00 (old labels 55, 58, 62, 65): missing points; accepted F4.2 within ≈2.5 own noises at night; P2, P3 error rising without corr effect (P2, P3 within 3–4.5 noises).
- References' corr alone on 19/10 17:15–17:30 (BG.3.1 cx −7, R1 cy −7, R2 cy +9 noises): a reference residual, not a target; P targets within 5 noises; inside SNCF01-60.
- Isolated measurements 21/10 20:15 and 22/10 10:30: P5–P7 cx +5–9 noises against a rolling level that spans 15 h gaps: the return of the evening dip of 21/10 and the start of 22/10 (SNCF01-71 of the next batch), not separate events.
- Single scattered daytime exceedances inside the shimmer labels (CV.9, F4.3 Target issue, CV.6 Target issue): covered.
- R2 error elevated 19/10 18:30 → 20/10 12:00 (+13 noises) without own corr effect beyond what SNCF01-60 and the added label carry.
3. The documentation
families.mdboundary "Thermal displacement or Displacement" (comes back at equal temperature) decided SNCF01-60; the catalogue line "A reference loses its brightness" and FLORIDA-66/71 decided the added label.labeling-rules.mdsection 1 (missing point not labelled) removed 55, 58, 62, 65; section 2 (start = measurement before the first touched point) corrected 57 and 38.days.py --boundson the median of the targets says "never > 5 noises" most days, because P7/P5 rise while P1/P2 drop: the cycle is only visible per sensor (--sensors P7,P5). Worth a sentence in the guide.ranges.py --rolling 4misses slow shifts of a few hours (the R1 episode of 20/10): it was found on the per-slot table ofdays.pyand onoverview.py --rigid.- The old labels' "and some P" sentence was copied from one train label to another although all P were rejected at 16/10 16:00 and 20/10 05:45 (rule "a sentence taken from a model is reread").
stations.mdsays SNCF01 has no system metrics: the metric files exist but are empty; consistent.
Difficulties
precedents.pynever answered: the Ollama container runs but lists no model (/api/tagsreturns an empty list); several sessions were stuck on it. Precedents were read bygreponlabeling/auto/*.jsonlinstead (FLORIDA-66, FLORIDA-71).- No weather and no temperature for this station: the thermal and visibility causes are deduced from the pose and the images only.
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, labeling/auto/SNCF01.jsonl (edges of the range), work/existing.jsonl, work/spanning.jsonl, request.json; data/SNCF01 through inventory.py, overview.py (window, --rigid 19/10 and 20/10), ranges.py (--rolling 4, error_value), days.py --bounds (median, P7, P5, P1, P6, P4), summary.py, metrics.py, plot.py, thumbnails.py; images 14/10 13:00, 16/10 16:00, 18/10 08:00, 19/10 08:00, 19/10 13:00, 21/10 13:00; precedents by grep: CLARIDEN-12, FLORIDA-66, FLORIDA-71, NIEDERHORN-150/151/32/158, MICHIGAN-17.