Files
aimsight-labeling/auto/analyses/BREITHORN_2026-10-13T0000_2026-10-22T1600/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

9.8 KiB
Raw Blame History

Labels: BREITHORN, 2026-10-13T00:00Z → 2026-10-22T16:00Z (UTC)

This is a requalification of the 7 labels of the earlier analysis, under the new rule (labeling/labeling-rules.md §1): a label stays only where at least one target goes more than 5 of its own noises beyond its level on the corrected curves. I did not redo the reading. The earlier text is kept in work/analysis.before-requalify.md. The ids are those of work/requalify_input.jsonl. The earlier analysis called the drift BREITHORN-11; it is now BREITHORN-39.

Context: 4 h cadence. There is no weather (no data/BREITHORN/weather/ folder), and the camera is indoors. The image-quality columns are empty, so quality.py was not used. Distances (inventory): COLJ19 8.9 m and POUI19 9.4 m (3.3 and 3.5 mm/px); COLI20, POUI20 and COLH20 about 20–22 m (7.6–8.1 mm/px); COLI21 30 m (11.4); COLI22 and POUI22 40 m (15.0–15.2); COLI24 and POUI24 60 m (22.5–22.8). The references have no mm.

Result: 4 kept as they are, 1 rewritten, 2 removed.

1. My labels

2026-10-13T00:00Z → 2026-10-13T08:00Z (BREITHORN-10, kept as is)

  • Phenomenon of the catalogue: a pass-by hides the targets for the length of one measurement
  • Precedents: BREITHORN-9 (same signature, a vehicle at the bottom of the scene)
  • Family: Obstruction, because the object is seen
  • Measured: at 13/10 04:00, POUI22 corr_x +0.279 px (+33.8 noises, rolling ±3), which is about 4.2 mm, and corr_y −0.312 px (−29.4, about 4.7 mm). POUI24 corr_x +0.226 px (+25.6, about 5.2 mm). COLI22 corr_y −0.155 px (−24.5, about 2.3 mm). COLH20''' and COLI24 are rejected. Everything is back at 08:00.
  • Images: 13/10 04:00, a car with a light on at the bottom (earlier reading)
  • Confidence: high

2026-10-13T04:00Z → 2026-10-16T20:00Z (BREITHORN-39, kept as is)

  • Phenomenon of the catalogue: the camera tilts little by little (the return phase); cause per the expert: parallax between the near targets and the far references
  • Precedents: BREITHORN-2 and BREITHORN-13 (spanning)
  • Family: Displacement (expert)
  • Measured: ranges.py --k 5 over the label's own range, with no rolling (a rolling level would follow the drift and erase it). POUI19 corr_x reaches −6.4 noises (16/10 12:00) and −5.9 (16/10 16:00) against the median of the range. COLJ19 corr_x reaches −6.3 (15/10 00:00) and −6.7 (16/10 20:00). From the start, POUI19 comes back by about 0.04 px, which is about 0.14 mm, so these are small mm on near targets. Over 10–19/10, with --min-measurements 2, no range lasts. The excursion is made of single measurements at its extremes. It is a slow drift, over the threshold at its ends.
  • Images: fig_spot_sizes.png (earlier reading)
  • Confidence: medium on the threshold (a slow drift just past 5 noises at its ends); high on the cause (expert, borne out by the spot sizes)

2026-10-15T20:00Z → 2026-10-16T12:00Z (BREITHORN-12, kept as is)

  • Phenomenon of the catalogue: a gap common to several stations
  • Family: No data. There is no measurement from any sensor at 16/10 00:00, 04:00 or 08:00, and no image (inventory: one gap of 960 min). So the station delivers nothing, which still meets the new rule.
  • Confidence: high

2026-10-19T00:00Z → 2026-10-19T12:00Z (BREITHORN-14, kept as is)

  • Phenomenon of the catalogue: the camera moves through a thermal effect (cold trough of the housing)
  • Precedents: BREITHORN-17, 21, 24, 25, 27
  • Family: Thermal displacement (expert)
  • Measured: rolling ±3 over 17–21/10, at 19/10 08:00. COLH20 corr_y is +0.059 px (+11.4 noises, about 0.48 mm). COLI21 corr_x is −0.074 px (−6.1, about 0.84 mm). POUI22 corr_x is +0.038 px (+7.4, about 0.57 mm). REFPOU22 corr_y is +6.2. The corrected curves keep the thermal movement beyond 5 noises, so the label stays.
  • Confidence: high

2026-10-21T16:00Z → 2026-10-22T08:00Z (BREITHORN-15, description rewritten)

  • Phenomenon of the catalogue: a vehicle or machine that stays in front of the targets
  • Precedents: BREITHORN-16 (the next night)
  • Family: Obstruction
  • Measured: rolling ±3 over 19–22/10 16:00. POUI22 corr_y is +0.157 to +0.158 px (+12.0 and +12.1 noises, about 2.4 mm) and corr_x +0.121 px (+13.8, about 1.8 mm). COLI22 corr_x is −0.100 px (−9.9, about 1.5 mm) and corr_y −0.063 px (−7.3, about 0.9 mm). All of this falls over 21/10 20:00, 22/10 00:00 and 04:00, and COLI24 is rejected over the same three measurements. The ±8–11 at the edges (21/10 16:00, 22/10 08:00) are an artefact of the rolling window, which contains the three-measurement plateau.
  • Why rewritten: the old description began with the missing COLI24 curve, which is no longer labelled for itself. It also said COLI22 and POUI22 "deviate a little", while they go 7 to 14 noises beyond their level. The new text rests on that effect and keeps the bounds, the family and the image evidence. It loses its review field.
  • Confidence: high

2. What I left without a label

Removed by the requalification:

  • BREITHORN-40, 2026-10-13T20:00Z → 2026-10-14T12:00Z, Displacement (the upper targets rise together over the night). The label is mainly about corr_y. There, rolling ±1, ±2, ±3, 12 h and 1 d (12–16/10, and 13/10 04:00 → 16/10 20:00) find no target beyond 5 noises. COLI20 corr_y reaches +5.3 only with a fixed level computed outside the episode, or a 2 d rolling. COLJ19, POUI20 and POUI19 stay under 5 in corr_y. Only COLI20 corr_x passes, and only at 14/10 00:00: +5.2 in the 12–16/10 window, +5.9 to +6.2 in the 13–16/10 window, where its noise comes out smaller. The same sensor goes as far (+5.5 to +6.0) at 15/10 20:00, unlabelled. So this is COLI20's ordinary scatter under the drift, not a distinct event. The jump from one measurement to the next (+7.7 noises for COLI20 corr_y) is not a deviation from the level. The common tilt step of the three cameras of the building stays a fact of the pose. The compensation absorbs it to within the noise.
  • BREITHORN-37, 2026-10-14T04:00Z → 2026-10-14T12:00Z, Obstruction (lit objects behind the references, 14/10 08:00). No target goes beyond 5 noises. COLH20, POUI20 and COLI21 corr_x stay at about 3–4 noises (ranges.py lists no target range at that instant, rolling ±1 to ±3). Only the references cross, barely: REFI'21 corr_y −5.0 to −5.7, REFG44N −5.2, REFPOU22 +5.2 to +5.9 depending on the window. The rule asks for a target, and the earlier analysis already put the deviations at 3–5 noises.

Still below the threshold (from the earlier reading, unchanged): 15/10 12:00–20:00, the warm hours (3–4 noises); 18/10 20:00 (3–4); 19/10 20:00, the upper group dips (−4 to −5, COLI20 corr_y −5.1 with rolling ±3 in the 19–22/10 window, one measurement, no new label sought); 20/10 12:00 (2–3); 22/10 04:00 POUI24 (~3).

3. The documentation

  • labeling-rules.md §1 settled every case. The point that a missing point is not labelled for itself is what led to rewriting BREITHORN-15 around its effect.
  • How large a short episode looks in noises depends on the window ranges.py uses: the noise of COLI20 corr_x is 0.0073 over 12–16/10 and 0.0033 over 13–16/10. A case at the threshold (BREITHORN-40) passes or fails with the window. I decided it by comparing with the same sensor's unlabelled deviations in the same window. The rule could say which context window to use for a short label inside a drift.
  • For a drift label, the rule's "--rolling under a drift" cannot be applied to the drift itself, because the rolling level follows it. I measured it with no rolling over its own range.
  • Earlier precedents that guided the removed labels (NORDEND-112, JUNGFRAU-50/89) are not contradicted. Those labels were simply below the new threshold here.

Requalification

Commands (outputs in work/requal/):

  • python3 tools/scripts/ranges.py data/BREITHORN --columns corr_x,corr_y --k 5 --rolling 3 --start S --end E, with the windows 11–15/10 (label 10), 12–16/10 (labels 40 and 37), 17–21/10 (label 14) and 19/10 → 22/10 16:00 (label 15).
  • Label 40: the same with --rolling 1, 2, 12h, 1d, 2d, with --outside 2026-10-13T20:00Z/2026-10-14T12:00Z, and in the 13/10 04:00 → 16/10 20:00 window with --rolling 2, 3, 1d.
  • Label 39: --k 5 without rolling over 13/10 04:00 → 16/10 20:00, and --min-measurements 2 and 3 over 10–19/10.
id bounds family max deviation (noises) verdict
BREITHORN-10 13/10 00:00 → 08:00 Obstruction POUI22 cx +33.8, cy −29.4; POUI24 cx +25.6 kept
BREITHORN-39 13/10 04:00 → 16/10 20:00 Displacement COLJ19 cx −6.7, POUI19 cx −6.4 (no rolling, the drift itself) kept
BREITHORN-40 13/10 20:00 → 14/10 12:00 Displacement corr_y < 5 under rolling (COLI20 +5.3 at most without); COLI20 cx +5.2 to +6.2, equalled unlabelled on 15/10 removed
BREITHORN-37 14/10 04:00 → 12:00 Obstruction targets 3–4; references only −5.0 to +5.9 removed
BREITHORN-12 15/10 20:00 → 16/10 12:00 No data station delivers nothing kept
BREITHORN-14 19/10 00:00 → 12:00 Thermal displacement COLH20 cy +11.4, POUI22 cx +7.4, COLI21 cx −6.1 kept
BREITHORN-15 21/10 16:00 → 22/10 08:00 Obstruction POUI22 cx +13.8, cy +12.1; COLI22 cx −9.9 rewritten (description on the effect, not on the missing COLI24)

Difficulties

  • How large a short episode looks in noises depends on the context window (see section 3). BREITHORN-40 is a judgement at the threshold.

Sources read

instructions/analyst.md; labeling/labeling-rules.md §1; tools/GUIDE.md (ranges, days, metrics); tools/DATA.md (references in the cause table); the earlier analysis.md and report.md of this folder; work/requalify_input.jsonl. Tools: ranges.py (see above), inventory.py (distances, mm/px, image gap).