Files
aimsight-labeling/auto/analyses/SNCF01_2026-10-22T1030_2026-11-01T0000/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

17 KiB
Raw Blame History

Labels: SNCF01, 2026-10-22T10:30Z → 2026-11-01T00:00Z (UTC)

Task: verify the 28 existing labels of the range (work/existing.jsonl) under the rules of 2026-11-17 (5 own noises on corr_x/corr_y; thermal only when the corrected curves keep it; a missing point not for itself; No data only when nothing is delivered). Result: 26 labels in labels.jsonl — 13 kept unchanged, 12 corrected, 3 removed, 1 added.

Context: no weather/ folder for SNCF01 (no MeteoSwiss weather: a station without a Swiss position); the system metrics (BusVoltage, BusCurrent, Power, ShuntVoltage) are all empty over the window and there is no sensor_*_deg_c: no temperature. No image quality metrics (older camera). Images on the hour (plus a few at the measurement minute), none during the outages. Cadence 15 min. 27 sensors, 3 references (BG.3.1, R1, R2), targets 16 to 21 m from the camera, about 6 to 8.5 mm per px (CV.1: no factor, px only). The sun is read on the noon images: 22/10 and 26/10 overcast, 25/10 sunny morning with clouds, 28, 29, 30, 31/10 full sun (noon_images.png).

Noise reference (median step between consecutive measurements over the window, overview.py): P4–P7 0.005–0.009 px (≈0.04–0.06 mm), P1–P3 0.009–0.014 px, BG.2/BG.3/F4.5/F4.6 ≈0.012 px, CV.2/CV.4 ≈0.03 px, CV.9 0.07–0.08 px.

Verdict per old id

Old id Verdict Evidence
SNCF01-70 kept 22/10 10:30–13:30: 5+1+3 missing steps separated by single measurements (12:00, 12:30): one No data. Metrics empty, no image in the gaps.
SNCF01-71 corrected (description) Curves flat within the day but P5.x ≈+10, P7.x +8…+12, P6.y +8…+10 noises above the 21/10 evening level (P5.x −1, P7.x −2): the corrected curves keep the cycle; the old text said the compensation absorbs it (would mean removal). P7.x back from +9 to +2 at the isolated 15:30 measurement with the pan step. Overcast on the images.
SNCF01-72 removed Turbulence on 22/10 not seen: one exceedance over the 11 measurements (P7.x −5.6 noises at 15:30, the thermal return); error_value steps ≤0.013 (vs 0.03–0.07 on sunny days); overcast images.
SNCF01-73 kept 15:15 missing, 15:30 single measurement, then 259 steps missing to 25/10 08:30: one No data.
SNCF01-74 corrected (end 15:00→13:15, description) Exceedances 08:30 (CV.4 +5.1, F4.2 −5.9) to 13:00 (CV.1 −5.7, P6.y −5.3, P7.y −6.1); after 13:00 only F4.3 (Target issue SNCF01-5). Sun 09:00–13:00, overcast 12:00 and 14:00. Named targets fixed (CV.4, F4.2, F4.6, CV.9, P1; the old list named BG.3, CV.6, not moving that day).
SNCF01-75 kept P5.y, P6.y +10…+13 noises by day vs ≈0 the following night; back ≈21:00; pan plateau ≈22:00–22:30. Bounds hold.
SNCF01-76 corrected (description) 25/10 23:45: P7.x jumps +0.048 px (≈7.5 noises, ≈0.3 mm) with its error_value 0.253→0.319, P5's too; BG.3.1 at 0.685 (threshold 0.7), R1 0.396. The old text said "the other targets do not move": P7 does, its point is wrong.
SNCF01-77 corrected (description) 26/10 day: P7.x up to +11 noises at 10:00 (tilt drop), P5.x +5…+7 vs night −1…0: holds. Old text (copied) said the curves come back at nightfall: they do not (evening episode, outage at 22:00). Overcast day, smaller pose cycle.
SNCF01-78 kept 14:45, 15:00 missing; common gap (other stations) as stated; I did not recheck the other stations.
SNCF01-79 corrected (start 14:30→15:15, description) Event at the resumption after the 15:00 gap: P6.y +5→+8 at 15:15, P5.x +7→+11 at 15:30, then P5.x/P6.x to +19…+22, P6.y +14, P4.y −10, P1.x −8, P1.y −7, P3.y −7 noises by 19:00–22:00 (≈0.15 px ≈1 mm on P5). error_value of R2, BG.3.1, P1, P3, CV.1, F4.2 above 5 MAD from 17:15; sigma0 max of the range at 17:00. Images 16:00–21:00: nothing clear (18:00 vs 25/10 18:00). Back at level on 28/10. Gap rule: an event starting at the resumption starts at that measurement.
SNCF01-80 kept 63 + 68 missing steps around the lone 27/10 14:00 measurement: one No data.
SNCF01-81 corrected (07:15–16:15 → 08:00–14:45, description) Exceedances 08:15 (CV.4 −6.2, CV.9.y −5.1, P2.y −5.4) to 14:30 (CV.4.y +5.9); nothing beyond 5 after. 07:15–07:45 the pose is still at its night level.
SNCF01-82 kept P5.y/P6.y up to +20, P7.x +16 noises; back ≈21:30; pan plateau 22:30; start 07:30 vs pan leaving at 07:45: one step.
SNCF01-83 kept CV.2 at 10:30: −9.5 px in x, +12.6 px in y (≈−59 / +78 mm), back at 10:45; CV.4 rejected.
SNCF01-84 kept 21:15: BG.3.1, R1 rejected, no pose; BG.3.y −72, BG.2.y −49, F4.5.y −45, R2 −33 noises (BG.3 ≈−0.86 px ≈−5 mm).
SNCF01-85 kept 22:45: same signature, BG.3.y −70, F4.5.y −48, BG.2.y −47 noises.
SNCF01-86 corrected (08:00–16:00 → 10:45–15:30, description) Exceedances 11:00 (CV.9.y +5.0) to 15:15 (CV.4.x −5.6); before 11:00 nothing beyond 5 (CV.4 rejected at 08:30 only: a missing point).
SNCF01-87 corrected (start 08:30→07:45, description) Pan 14.21 m° at 07:45 → 12.11 at 08:00, tilt leaving at 07:30–08:00: the morning turning point is 07:45, three steps before the old start. Text had "the evening and the evening". P5.y/P6.y +16…+19 noises.
SNCF01-88 kept 15:00 missing, image present, common gap as stated.
SNCF01-89 kept P5.y +21, P6.y +20, P7.x +13 noises; return slow, curves back ≈00:30–01:00, pan rising all night to 06:00: without rest, ends at the next start (31/10 07:45). Start 08:00 vs turning point 07:45: one step.
SNCF01-90 corrected (08:30–19:00 → 09:15–15:30, description) Exceedances 09:30 (CV.2.y −6.8) to 15:15 (F4.2.y +5.2); hourly error_value steps back to night values from 16:00; the old end ran into the evening trains (labelled apart).
SNCF01-91 kept 18:00: P3/P4 −48/−56 noises, BG.3 −64; 18:15 clean (targets within 3 noises): two labels.
SNCF01-92 kept 18:30: P1 −55, P2 −66, P3 −61 noises in y.
SNCF01-93 corrected (description) 21:30: P3.x −10, P1.x −6, BG.3.x −5 noises, pan jump, BG.3.1 error_value 0.622; the old text named P1, P2, BG.2 in x and y. 21:45: F4.2 rejected alone (same event).
SNCF01-94 removed F4.2 rejected at 21:45 alone: a missing point; no curve beyond 5 noises at 21:45 (all targets within 2 noises). Already inside SNCF01-93.
SNCF01-95 removed F4.2 rejected at 31/10 07:15 alone: a missing point; at 07:15 no target beyond 5 noises.
SNCF01-96 corrected (07:45–15:45 → 08:15–16:00, description) Exceedances 08:30 (BG.3.y −6.7) to 15:45 (BG.2.x −5.1); P2 (up to 11 noises), CV.4, CV.2, P1 move most: the old list omitted P2 and P1.
SNCF01-97 corrected (description) Holds as a thermal day, but the curves do not come back: P6.y +15, P5.y +11…12 noises at 23:30 vs ≈0 on the previous nights, partly the tilt step of 12:30 (added label). The old text said they come back.

1. My labels

Only the corrected and added labels are detailed here; the kept ones carry their old text and the evidence of the table above.

2026-10-22T10:30Z → 2026-10-22T15:30Z (SNCF01-71, corrected)

  • Phenomenon of the catalogue: The camera moves through a thermal effect
  • Precedents: SNCF01-68 (21/10, same station, the day before, same setting); the precedents tool did not answer (see Difficulties)
  • Family: Thermal displacement: the pose has its day profile (pan low, scale high, return at 15:30) and the curves keep a day shift; no nuisance on the images (overcast)
  • Measured: pan 5.8–8.0 m° (night ≈13), scale +194…+248 ppm, at its extreme 13:30; P5.x +8…+10, P7.x +8…+12, P6.y +7…+10 noises against the window night level, against −1/−2 on 21/10 evening; 15:30: pan 10.1, scale +141, P7.x +2, references' corr shift (BG.3.1.x −9, R1.y −9, R2.y +9 noises)
  • Images: 13:30, 14:00: overcast, diffuse light
  • Confidence: medium: few measurements, no night of the same day

2026-10-25T08:30Z → 2026-10-25T13:15Z (SNCF01-74, corrected)

  • Phenomenon of the catalogue: Atmospheric turbulence
  • Precedents: SNCF01-67, -81, -86, -90, -96 (same station, same template); none useful beyond
  • Family: Visibility (turbulence): one-measurement jumps on many targets with the error_value agitated (CV.4 step 0.042 by day vs 0.002 at night), sunny hours only; not Thermal (that is the slow component, labelled apart)
  • Measured: 21 exceedances by day (rolling ±4 median), CV.4 ×4 up to 7 noises, F4.2, F4.6, CV.9, P1; last at 13:00 (outside the Target issue F4.3)
  • Images: 09:00, 10:00, 11:00, 13:00 sunny with sharp shadows; 12:00 and 14:00 cloudy
  • Confidence: high on the agitation, medium on the end (agitation of error_value lingers to 14:30)

2026-10-25T23:30Z → 2026-10-26T00:00Z (SNCF01-76, corrected)

  • Phenomenon of the catalogue: A pass-by hides the targets for the length of one measurement; A target left under the threshold latches onto…
  • Precedents: SNCF01-84/-85 (train signature, stronger)
  • Family: Obstruction (a passage, deduced)
  • Measured: CV.1, CV.2, CV.4, CV.5, CV.8, CV.9, F4.2 rejected at 23:45; P7.x −0.4465 → −0.3983 → −0.4439 px (≈7.5 noises); P7 and P5 error_value 0.25→0.32; BG.3.1 0.685
  • Images: none at 23:45 (hourly)
  • Confidence: medium (cause deduced)

2026-10-26T08:15Z → 2026-10-26T22:00Z (SNCF01-77, corrected)

  • Phenomenon: The camera moves through a thermal effect
  • Family: Thermal displacement
  • Measured: pan 13.4 → 9–10 m°, tilt −163 → −166, scale +70 → +160 ppm (half the sunny days); P7.x +11 at 10:00, +4…+6 by day; P5.x +5…+7; no return before the outage
  • Images: 12:00 overcast
  • Confidence: medium (amplitude modest; the evening is mixed with SNCF01-79)

2026-10-26T15:15Z → 2026-10-26T22:00Z (SNCF01-79, corrected)

  • Phenomenon: Visibility drops over the whole scene / a reference less well seen (cause not established)
  • Family: Visibility as before (reviewed choice): error_value up on references and many targets, sharp-edged episode, sigma0 at its maximum; not a thermal return (the curves go the other way)
  • Measured: see table; at 15:15 P6.y +8 (vs +4/+5), 15:30 P5.x +11 (vs +7); R2 error_value up to +36 noises at 18:15; 16 sensors above 5 MAD at 18:15
  • Images: 16:00–21:00 night images, 18:00 compared with 25/10 18:00: same scene, nothing clear
  • Confidence: medium on the start, low on the cause

2026-10-28T08:00Z → 2026-10-28T14:45Z, 2026-10-29T10:45Z → 15:30Z, 2026-10-30T09:15Z → 15:30Z, 2026-10-31T08:15Z → 16:00Z (SNCF01-81, -86, -90, -96, corrected)

  • Phenomenon: Atmospheric turbulence
  • Family: Visibility (heat shimmer), as before
  • Measured (rolling ±4 median, own noise; accepted measurements; trains excluded): 28/10 35 exceedances (CV.4 ×8 up to 10, CV.2 ×6, P2, P3, CV.1); 29/10 17 (CV.4 ×6 up to 8.5, F4.2, CV.6, BG.3, BG.2); 30/10 42 (CV.2 ×8 up to 10, F4.2, CV.6, F4.6, F4.5, CV.4); 31/10 49 (CV.4 ×8, P2 ×7 up to 11, CV.2 up to 11, P1, BG.3, F4.2). Nights: 0 to 3. Magnitudes: CV.4 ±0.2–0.26 px (≈1.7–2.2 mm), CV.2 up to 0.37 px (≈2.3 mm), P2 0.12 px (≈0.7 mm). Hourly max error_value step (median of CV.4, CV.2, BG.3.1, R2, P2): 0.01–0.07 between 08 and 15 h on sunny days, ≤0.005 from 16 h
  • Images: 12:00 of each day (31/10 also 13:00): full sun, sharp shadows
  • Confidence: high on the phenomenon; bounds ±1 step (the agitation fades in and out)

2026-10-29T07:45Z → 2026-10-29T23:15Z (SNCF01-87, corrected)

  • Phenomenon: The camera moves through a thermal effect
  • Family: Thermal displacement
  • Measured: pan 14.21 (07:45) → 12.11 (08:00) → ≈3 (15:15); tilt −161.4 → −169.5; scale +40 → +286 ppm; P5.y +19, P6.y +17, P7.x +11 noises at 15:45; back by 21:15–22:15; pan plateau 22:45
  • Images: 12:00 sunny
  • Confidence: high

2026-10-30T21:15Z → 2026-10-30T22:00Z (SNCF01-93, corrected)

  • Phenomenon: A pass-by …; a reference less well seen
  • Family: Obstruction (deduced passage), as before
  • Measured: 21:30: P3.x −10, P1.x −6, BG.3.x −5 noises (P3 ≈−0.1 px ≈−0.6 mm), pan 9.82 → 12.33 → 9.73, BG.3.1 error_value 0.622; CV.2 rejected; 21:45 F4.2 rejected
  • Images: none at 21:30 (21:00 and 22:00 exist, before and after)
  • Confidence: medium

2026-10-31T07:45Z → 2026-11-01T00:00Z (SNCF01-97, corrected)

  • Phenomenon: The camera moves through a thermal effect
  • Family: Thermal displacement (plus the added Displacement for the step)
  • Measured: pan 15.8 → 2.8 m° (15:00) → 10.3 (23:30); scale +4 → +290 → +120 ppm; P6.y up to +27, P5.y +22, P4.y +15, P7.x +12 noises; at 23:30 P6.y +16, P5.y +12 (previous nights ≈0)
  • Images: 12:00, 13:00 sunny
  • Confidence: medium (the part due to the step and to an incomplete thermal return cannot be separated without temperature)

2026-10-31T12:15Z → 2026-10-31T12:45Z (added)

  • Phenomenon of the catalogue: A step of the camera that the compensation does not erase
  • Precedents: none useful (tool unavailable; a grep on "the camera moves" gave thermal cases of other stations, MICHIGAN-9, LENZSPITZE-117, not a lasting step)
  • Family: Displacement: a lasting change of the camera that the compensation does not erase, not proved thermal (no temperature, no return at equal hour: night tilt −168/−165 m° on 01/11 and 02/11 vs −161/−163 before); rather than Thermal displacement, whose boundary requires a return
  • Measured: tilt −167.69 (12:15) → −172.75 (12:30) → −174.45 m° (13:00); overview.py --rigid 2026-10-31T11:45Z: ty −0.10 → −0.28 px at 12:30 (all sensors move up together), median target residual 0.02–0.03 → 0.04–0.05 px; P6.y +18 → +22…+25, P7.y +4 → +9, P3.y +6 → +10…+12, P5.y +16 → +21 noises (≈0.25–0.35 mm on P6, P7, P3); stays all afternoon. From 17:00 the references' corrections shift and stay (BG.3.1.y +3 → +10…+17, R1.x −3 → −9…−16, R2.x 0 → +5…+11 noises), through 02/11; targets not beyond 5 noises at 17:00
  • Images: 12:00 and 13:00: same framing, full sun; nothing in front
  • Confidence: medium: the step is clear; its cause (sun on the camera, mount) is not established

2. What I left without a label

  • 26/10 daytime agitation: 3 exceedances (CV.4 +5.4 at 09:45, P5.y +5.1, P7.x +6.5 at 10:00 with the tilt drop): overcast day, part of the thermal label; no turbulence label.
  • 22/10: P7.x −5.6 noises at 15:30: the thermal return, within SNCF01-71.
  • 27/10 14:00 isolated measurement: large deviations against the rolling neighbours (P5.x −11, P6.x −10, R1.x +16) are an artefact of neighbours across the outage; inside No data SNCF01-80.
  • 28/10 07:15 and 08:00: R1.x ±5, BG.3.1.y ±5–6 noises: references only, at the resumption.
  • CV.7 (Target issue SNCF01-3): accepted at 28/10 14:00 and 30/10 13:15 with +4.3 / +8.8 px jumps (63 / 126 noises): covered by its Target issue.
  • 31/10 17:00: the references' corrections shift (above) without any target beyond 5 noises at that instant: mentioned in the added label.
  • Removed labels: SNCF01-72 (22/10, max 5.6 noises on one point, the thermal return), SNCF01-94 (F4.2 missing point; targets ≤2 noises at 21:45), SNCF01-95 (F4.2 missing point; no target beyond 5 noises at 07:15).

3. The documentation

  • Helped: labeling-rules.md sections 1 and 2 (5 noises, gap edges, morning turning point, lasting jump bounded on the jump), families.md boundaries (Thermal vs Visibility; "A gap"), reading-rules.md section 5 (rigid read before saying "camera"). tools/DATA.md for the mm factors.
  • The spanning labels: SNCF01-7 (the long drift, to 01/11) says the drift flattens at the end of the month; within the range it is right, but the step of 31/10 12:30 adds a jump on the last day, now labelled apart. SNCF01-1 to -6 hold within the range (BG.1, F4.1 rejected at all 517 measurements; CV.7 499, F4.4 482, F4.3 445, CV.6 279 rejected; CV.7's last accepted points jump as said).
  • The old turbulence labels had been bounded on the error_value agitation and shared one copied list of targets; the thermal labels shared a copied sentence ("partly all at once at nightfall"), wrong on 26/10 and 31/10.
  • Missing: no weather and no temperature for SNCF01; the sun had to be read on the images (useful: it separates the sunny shimmer days from the overcast ones).
  • The precedents tool could not reach its Ollama server: precedents were read from the station's own labels and a grep.

Difficulties

  • labeling/scripts/precedents.py failed: Ollama unreachable (docker start ollama-bge-m3 needed); no RAG precedents.
  • overview.py --table and my tables: two measurements at 22/10 12:30 (15 s and 19 s), duplicate instant in the export.

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, labeling/auto/SNCF01.jsonl (labels from 19/10 to the range end)
  • Tools: inventory.py, overview.py (window; --rigid 2026-10-31T11:45Z on 31/10–02/11), ranges.py --rolling 4 --instants 2, metrics.py, plot.py, thumbnails.py --zone; whole images 25/10 18:00, 26/10 18:00, 31/10 12:00, 13:00
  • My scripts: scripts/table.py (pose + corr in noises per instant), scripts/agitation.py (daily exceedances day/night), scripts/build_labels.py; inline checks of error_value steps by day and hour