Files
aimsight-labeling/auto/analyses/NEBRASKA_2026-10-13T0000_2026-10-22T0000/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

10 KiB
Raw Blame History

Labels: NEBRASKA, 2026-10-13T00:00Z → 2026-10-22T00:00Z (UTC)

Requalification, under the new rules (labeling-rules.md section 1: only what stands out from the noise of the corrected curves), of the 4 labels of this analysis (work/requalify_input.jsonl, all reviewed by the user on 2026-11-17). The reading was not redone: the verdicts of the previous analysis (kept in work/analysis.before-requalify.md) stand. Each label was checked against the 5-noise threshold only. Result: 4 kept unchanged, 0 rewritten, 0 removed. labels.jsonl is the input without the id field, review fields included.

Context. One measurement every 4 h (52 instants in the range, 3 missing). REFCOLH22 is rejected at every measurement, so the compensation runs on REFH22 and REFP''21 only (the long label NEBRASKA-2). The image quality columns are empty on this camera. Weather: NEBRASKA has no weather/ folder, so there is no MeteoSwiss data. The station is indoors anyway. Distances and mm per px (inventory.py): POUJ20 7.9 m (3.10 mm/px), COLJ20 8.6 m (3.37), COLJ19 10.9 m (4.21), POUJ19 11.7 m (4.64), POUH19 16.8 m (6.72), POUH20 30.8 m (12.14), POUH21 37.8 m (14.87). COLH19, COLH20 and COLH21 have no sure factor (?), so they are given in px only.

1. My labels

2026-10-13T04:00Z → 2026-10-17T04:00Z (NEBRASKA-3, kept)

  • Phenomenon of the catalogue: "The camera tilts little by little" (here a pause and a slight backward movement), with "A compensation on two references".
  • Precedents: NEBRASKA-1 and NEBRASKA-6 (the same camera before and after).
  • Family: Displacement. It does not come back at equal temperature, and its cause is not established, so it is not Thermal.
  • Measured: ranges.py --k 5 over 10-10 → 17-10 20:00 with the level computed outside the label (--outside 2026-10-13T04:00Z/2026-10-17T04:00Z). In corr_y: COLJ20 +7.7 noises (14-10 00:00, +0.100 px, about 0.34 mm), COLH19 +7.4 (14-10 00:00), POUJ19 +6.7 (14-10 00:00, +0.071 px, about 0.33 mm), POUH19 +6.4 (14-10 00:00, about 0.42 mm). COLJ20 and COLH19 stay beyond 5 noises at several instants up to 16-10 16:00. COLJ23's ranges come from its quantised corr_y (noise 0.0014 px, values on two levels): they are not part of the event. Previous verdict (bounds, pose, link with BREITHORN and MICHIGAN): unchanged.
  • Images: 13-10 04:00 and 08:00 (previous analysis). Same scene, nothing in front.
  • Confidence: high on the threshold and the bounds. The cause stays not established.

2026-10-15T20:00Z → 2026-10-16T12:00Z (NEBRASKA-4, kept)

  • Phenomenon of the catalogue: "A gap common to several stations".
  • Precedents: none needed.
  • Family: No data. The station delivers no measurement at all.
  • Measured: inventory.py --start 2026-10-15T12:00Z --end 2026-10-16T16:00Z: 3 missing instants, between 15-10 20:00 and 16-10 12:00, for all the sensors (no partial curve). This is still a real No data under the new rule.
  • Images: none between 20:00 and 12:00.
  • Confidence: high.

2026-10-16T20:00Z → 2026-10-17T20:00Z (NEBRASKA-5, kept)

  • Phenomenon of the catalogue: "A target or a small group jumps and stays shifted" (here with a partial return).
  • Precedents: none useful.
  • Family: Displacement, because no cause is seen.
  • Measured: the COLJ19 corr_x jump from 16-10 20:00 to 17-10 00:00 is +0.074 px, 6.6 noises with the noise of the whole range and 11.1 noises with the noise outside NEBRASKA-3 (ranges.py). Against the level outside NEBRASKA-3, COLJ19 corr_x stays out: +10.5 noises at 17-10 00:00, 8.0 to 10.3 from 04:00 to 16:00, +6.5 at 20:00 (half way back). That is about +0.29 mm at its peak (4.21 mm/px). ranges.py --rolling 3 does not flag it, because a step that lasts five measurements is followed by a level of ±3 measurements. A lasting jump is measured against its level before, so the outside run is the right one here.
  • Images: fixed thumbnails (previous analysis). Nothing in front, and the shift is too small to see.
  • Confidence: high.

2026-10-18T16:00Z → 2026-10-19T16:00Z (NEBRASKA-7, kept)

  • Phenomenon of the catalogue: "The camera drifts away when the temperature of its housing drops" (sheltered, in a building).
  • Precedents: NEBRASKA-8 to NEBRASKA-11 (the cold mornings at the end of the month).
  • Family: Thermal displacement. It follows the temperature trough and comes back with it.
  • Measured: the corrected curves keep the movement beyond 5 noises.
    • ranges.py --k 5 over the whole range, with no rolling, at 19-10 08:00: POUH20 corr_y +7.0 noises, COLH21 corr_x +6.5, POUJ19 corr_y +5.7, COLJ20 corr_x +5.2. corr_y has no drift over the range that would bias this level. corr_x does: POUH20, COLJ20 and POUJ20 stay above it from 18-10 to the end, which is NEBRASKA-6.
    • Against the night level before the cold morning (17-10 20:00 and 18-10 00:00, read in work/corr_y.txt, noise of the whole range), at 19-10 08:00: POUH20 corr_y +6.7 noises (+0.083 px, about 1.0 mm), POUJ19 corr_y +6.5 (+0.111 px, about 0.5 mm), COLJ19 corr_y +5.6 (about 0.23 mm).
    • Against the daytime level of 18-10 (04:00 to 16:00; scripts/thermal_dev.py, work/requal/n7_dev_preday.txt): POUH21 corr_x +6.6, POUJ19 corr_y +6.0, POUH20 corr_y +5.6, COLJ19 corr_y +5.5.
    • Against a level that brackets the event (18-10 16:00, 19-10 20:00 and 20-10 00:00), which also removes the drift underneath (n7_dev_bracket.txt): COLJ19 corr_y +5.0, COLH21 and POUH21 corr_x +4.9. This is at the threshold, because the curves come back a little higher than they left.
    • ranges.py --rolling 3, --rolling 6 and --rolling 1d flag nothing. The bump lasts about six measurements, as long as the rolling windows, so these windows absorb it. They do not measure this event.
    • So the label holds, but only just. I keep it unchanged.
  • Images: 19-10 00:00 and 08:00 (previous analysis). Indoor scene intact.
  • Confidence: medium. The label is above the threshold against the levels before the event, and at the threshold against a level that brackets it.

2. What I left without a label

No label was removed. None of the four labels rests on missing points, and none is a camera movement that the compensation absorbs:

  • NEBRASKA-3: up to 7.7 noises (COLJ20 corr_y). Kept.
  • NEBRASKA-4: no measurement at all on the station. Kept.
  • NEBRASKA-5: up to 10.5 noises (COLJ19 corr_x). Kept.
  • NEBRASKA-7: up to 7.0 noises against the range level, and up to 6.7 against the night before (POUH20 corr_y); 5.0 against a bracketing level. Kept.

What the previous analysis left without a label is still below the threshold, and was not looked at again (no search for new events):

  • 20-10 04:00–08:00: a small corr_y dip of the top targets, under 5 noises (median corr_y −2.9).
  • 17-10 08:00: a small corr_x peak, below 5 noises.
  • 17-10 00:00: a corr_y rise on COLH21 and COLH20, about 4 to 5 noises.
  • 18-10: a small step of POUH20 corr_x. In ranges.py --rolling 3 it reaches 5.3 and −5.4 noises at 18-10 04:00 and 16:00, single points inside the drift of the long label NEBRASKA-6, which describes POUH20 rising with the tilt.

3. The documentation

  • labeling-rules.md section 1 was clear on the threshold. One gap: for a thermal bump that lasts about one day at a 4 h cadence, --rolling (named for thermal cycles) absorbs the bump itself. The night level asked for by the rule had to be computed by hand from the per-instant tables. A ranges.py option for "level = the night before" (like days.py --bounds, but per target) would settle such cases directly. days.py --bounds works on the median of the targets, which hides a deviation carried by two or three targets (there: never above 3.7 noises).
  • The full-range level of ranges.py without rolling is biased in corr_x by the long drift (NEBRASKA-6). It is usable in corr_y here, which has no drift.
  • No precedent was looked up again: the families and the bounds are unchanged.

Requalification

Commands (output in work/requal/):

  • python3 tools/scripts/ranges.py data/NEBRASKA --columns corr_x,corr_y --k 5 --start 2026-10-13T00:00Z --end 2026-10-22T00:00Z → full.txt: at 19-10 08:00, POUH20 cy +7.0, COLH21 cx +6.5, POUJ19 cy +5.7, COLJ20 cx +5.2; the COLJ19 cx jump at 17-10 00:00 is 6.6 noises.
  • … --start 2026-10-10T00:00Z --end 2026-10-17T20:00Z --outside 2026-10-13T04:00Z/2026-10-17T04:00Z → n3_outside.txt: NEBRASKA-3, COLJ20 cy +7.7, COLH19 cy +7.4, POUJ19 cy +6.7, POUH19 cy +6.4; NEBRASKA-5, COLJ19 cx +10.5 (17-10 00:00) to +6.5 (17-10 20:00).
  • … --start 2026-10-17T00:00Z --end 2026-10-21T00:00Z --rolling 1d --instants 2 → n7_rolling1d.txt: no range.
  • … --start 2026-10-13T00:00Z --end 2026-10-22T00:00Z --rolling 3|6 --instants 2 → full_rolling3.txt, full_rolling6.txt: only POUH20 cx ±5.3–5.4 on 18-10 (under NEBRASKA-6).
  • python3 scripts/thermal_dev.py <folder> <reference instants> <span> → n7_dev_preday.txt, n7_dev_bracket.txt: NEBRASKA-7 per target, see section 1.
  • python3 tools/scripts/inventory.py data/NEBRASKA --start 2026-10-15T12:00Z --end 2026-10-16T16:00Z → n4_inventory.txt: 3 instants missing for all the sensors.
id bounds family max deviation (noises) decision
NEBRASKA-3 13-10 04:00 → 17-10 04:00 Displacement 7.7 (COLJ20 cy, level outside) kept
NEBRASKA-4 15-10 20:00 → 16-10 12:00 No data no measurement at all kept
NEBRASKA-5 16-10 20:00 → 17-10 20:00 Displacement 10.5 (COLJ19 cx, level outside) kept
NEBRASKA-7 18-10 16:00 → 19-10 16:00 Thermal displacement 7.0 (POUH20 cy, range level); 6.7 vs the night before; 5.0 bracketing kept

validate.py: 0 errors. It reports 1 overlap to look at, NEBRASKA-3 with NEBRASKA-5. These are two distinct events (a camera bump of the top group, and COLJ19's own jump), and the overlap already existed.

Difficulties

  • The prescribed --rolling cannot measure a one-day thermal bump at a 4 h cadence (see section 3). I used the night level by hand instead.

Sources read

instructions/analyst.md; tools/GUIDE.md (ranges, days, the reading steps); labeling/labeling-rules.md (sections 1 to 5); the analysis folder: work/requalify_input.jsonl, labels.jsonl, analysis.md and report.md (previous versions, copied to work/*.before-requalify.md), work/corr_x.txt, work/corr_y.txt, work/days.txt. Tools: ranges.py (full range, outside, rolling 3, 6, 1d), inventory.py (distances, mm/px, the gap). Script: scripts/thermal_dev.py. data/NEBRASKA/ listing (no weather/).