Files
aimsight-labeling/auto/analyses/FLORIDA_2026-10-01T0000_2026-10-11T0945/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

14 KiB
Raw Blame History

Labels: FLORIDA, 2026-10-01T00:00Z → 2026-10-11T09:45Z (UTC)

Requalification on 2026-11-17 under the new rules (only what stands out from the noise of the corrected curves, at 5 of each target's own noises): 24 labels in, 18 kept unchanged, 2 rewritten (FLORIDA-15, FLORIDA-21), 4 removed (FLORIDA-2, FLORIDA-69, FLORIDA-19, FLORIDA-73). The verification work below (sections 1 and 3) stands; section 2 and the section "Requalification" are new.

Task: verify the 20 existing labels (work/existing.jsonl). No weather for FLORIDA (no weather/ folder). Older camera: the image-quality columns are empty, so quality.py does not apply. Two references only (p1_r, r1_r), each with its twin target (Sellette 3, Pont Ernen). Per-instant table: scripts/dump.py → work/dump.txt. Check for uncovered moments: scripts/uncovered.py (rejections, and corr beyond 3.5–5 noises of a ±6 h rolling median).

Verdict per old id

id verdict evidence
FLORIDA-2 kept Massif 2 rejected 07:30–07:45, back at 08:00
FLORIDA-3 kept hourly pose leaves its night level at about 09 h (pan, tilt), comes back in the evening; the housing temperature peaks at 13 h
FLORIDA-4 corrected 20:45 is already touched: error_value up on every sensor, pose tilt step, Sellette 4/5 corr shifted → start 20:30. Description: no registration from 21:00 (p1_r rejected), not only from 23 h; Massif 2 and Sellette 6 are kept 21:00–22:30
FLORIDA-5 kept gap of one measurement at 02:00; no change
FLORIDA-6 corrected bounds right (r1_r rejected 10:00–10:15). Cause: the 10:00 thumbnail shows the background behind r1_r in full sun, the spot drowned (as JUNGFRAU-50/55); the same lit background is seen on 03, 04, 05, 06, 07, 09 and 10/10 at 10:00. The old text said "cause not established, nothing in front"
FLORIDA-7 corrected "the other targets do not move" is false at 10:45: Sellette 4, 5 and 6 corr_x jump down together with an error spike of p1_r (new label)
FLORIDA-8 kept
FLORIDA-9 corrected the event is wider: rigid residual of r1_r drifts from 18:45 (to −0.2 px at 21:45), back at 22:15; corr_x of Massif 2, Sellette 5 and 6 up 18:45–21:45; then p1_r flickers (error up to 0.55, rigid dx +0.04 to +0.05) with joint drops of Sellette 4/5/Massif 2 at 22:30, 22:45 and 23:15, back at 23:30. r1_r thumbnails 19–22 h: smaller spot (p99 about 150 against 239 at 18:00 and 232 at 23:00). New bounds 18:30 → 23:30
FLORIDA-10, 11, 12, 13, 15 kept daily pose cycle seen on the hourly median with the housing temperature; FLORIDA-11 start (11:45) checked: hourly tilt leaves at 11–12 h
FLORIDA-16 kept start 03:15 (first touched point 03:30); end 08:00 (pose step back at 08:00 on the 9th)
FLORIDA-17, 18, 19, 21, 22 kept 18: Sellette 4 reaches the new level at 13:15–13:30
FLORIDA-20 corrected same as FLORIDA-6: the 10:00 thumbnail of the 9th shows r1_r's background lit by the sun

Spanning FLORIDA-1 (Sellette 6) and FLORIDA-14 (Massif 2): consistent within the range; nothing to change.

1. My labels (added)

2026-10-02T08:30Z → 2026-10-02T09:45Z

  • Phenomenon of the catalogue: A reference loses its brightness, without an obstacle
  • Precedents: none useful
  • Family: Visibility. The reference is less well seen; it is not an object (Obstruction)
  • Measured: rigid residual of r1_r dx −0.03 (08:30), −0.18 (08:45), −0.13 (09:00–09:15), −0.11 (09:30), −0.02 (09:45); Massif 2 corr_x +0.20 to +0.23 and Sellette 6 corr_y −0.28 to −0.33 (08:45–09:15); pan −0.021 against −0.046; scale 1.0006 against 1.0013
  • Images: r1_r thumbnails: peak 40 at 02/10 09:00, against 255 at 11:00 and on 03/10 09:00
  • Confidence: high

2026-10-02T10:30Z → 2026-10-02T11:00Z

  • Phenomenon: none exact (a reference near its threshold). Family: Visibility, consistent with FLORIDA-22. Measured: p1_r and Sellette 3 error 0.63 against 0.42 at 10:45; rigid p1_r dx +0.09; Sellette 4/5/6 corr_x −0.10 to −0.12, back at 11:00. Images: none at 10:45. Confidence: medium (cause)

2026-10-02T18:00Z → 2026-10-02T18:30Z

  • Massif 2 alone at 18:15: error 0.68, corr_y +0.17 (7 noises). Visibility like FLORIDA-2, 7 and 9. Confidence: medium

2026-10-04T08:30Z → 2026-10-04T09:00Z

  • Precedents: WYOMING-17 (one-measurement jump of a reference alone, Displacement, cause not established). Measured: at 08:45 Sellette 5 −0.10, Massif 2 −0.13, Sellette 6 −0.17 in corr_x against about −0.03; rigid r1_r dx +0.128; error of r1_r normal (0.20); tilt and scale jump. Images: none at 08:45; r1_r normal at 08:00 and 09:00. Confidence: medium

2. What I left without a label

Removed at the requalification

  • FLORIDA-2, 2026-10-01T07:15Z → 08:00Z, Visibility: missing points only. Massif 2 is rejected for two measurements; no target leaves its level beyond 5 noises inside the label (no range of ranges.py at k 5, plain or rolling 6 h). Sellette 6 starts a dip of its corr_y at 08:00, up to −14 noises at 09:00 (rolling 6 h), which belongs neither to Massif 2's rejection nor to this label (see below).
  • FLORIDA-69, 2026-10-02T10:30Z → 11:45Z, Visibility: missing points only. Massif 2 rejected for four measurements. The only deviation inside, Sellette 4 and Sellette 5 corr_x at 10:45 (−6.3 and −6.2 noises; rolling −5.6 and −5.5), is the p1_r jump that FLORIDA-68 already labels.
  • FLORIDA-19, 2026-10-09T08:00Z → 20:00Z, Thermal displacement: no thermal cycle that can be told apart beyond the noise. There is no clean night level that day: the night before is rain (FLORIDA-16), and the night after carries the post-rain shift (FLORIDA-18). Compared with the night after, Sellette 4 corr_x sits at −17.7 noises (08:00–12:45), Massif 2 corr_x at −11.4 (15:00–19:45) and Sellette 5 corr_x at −10.3 and −7.8. This is the shift that is not yet set, not a daily cycle: compared with the nights before the rain, Sellette 4 is at +0.2 noises over 08:00–13:30. Under a rolling level of 6 h, only one point of Sellette 4 at 12:00 reaches −5.9 noises. The label itself said its effect was mixed with the shift.
  • FLORIDA-73, 2026-10-09T09:45Z → 10:15Z, Visibility (r1_r rejected, sunlit background): under the threshold. Against a local level (rolling 6 h, or ±4 measurements) no target goes beyond 5 noises. The jump at 10:00 is +0.081 px for Sellette 5 corr_x (4.5 noises of the whole range) and +0.14 px for Sellette 6 (4.7 noises). The 7.5 noises of the plain run are measured from the median of the whole range, which sits below the post-rain shift. The same phenomenon on 02/10 (FLORIDA-67) holds, at 10.8 noises.

Rewritten

  • FLORIDA-15 (07/10) and FLORIDA-21 (10/10), Thermal displacement: both said that the corrected curves show almost nothing (the compensation absorbs it), which under the new rules would mean removal. Against their night level the curves do keep part of the cycle. On 07/10, Sellette 6 corr_y is at +9.7 noises over 11 measurements, 14:45–18:00. On 10/10, outside FLORIDA-22: Sellette 6 corr_x +6.5 (16:00–17:15), Massif 2 corr_x +6.1 (16:15–17:15), Sellette 5 corr_x −7.1 (18:45–19:45) and Sellette 6 corr_y −9.0 (18:30–19:45). The descriptions now say this, and the labels lose their review field.

Kept with a thin margin

  • FLORIDA-18 (09/10 08:00–13:30, onset of the post-rain shift, return FLORIDA-30 on the 15th). The shift itself crosses the threshold. Compared with the nights before the rain (nights of 05, 06 and 07/10), the night levels after it are Sellette 4 corr_x +5.6 and +6.5 noises and Sellette 5 corr_x +5.8 and +6.8 noises (nights of 10 and 11/10). But inside the label's own bounds, Sellette 5 sits at only +3.3 noises (median), and Sellette 4 reaches its new level at 13:45–15:30 (+7.1 noises), after the label's end. The only points beyond 5 noises inside are 10:00 (Sellette 5 +7.7, Sellette 6 +9.7), the r1_r moment of the removed FLORIDA-73. Kept, because the bounds may not be moved and the event is real; a later pass could extend its end to about 15:30.

Still without a label (from the verification)

  • About 10:00 on sunny days (03 to 06/10), the sun lights r1_r's background: a spike of the camera pose and small jumps of Massif 2 (and of Sellette 4 and 5 on the 5th), about 3 to 4 noises, mostly inside the thermal labels.
  • Isolated single-sensor deviations of 3.5 to 5 noises: Massif 2 on 01/10 06:30–07:00 and 04/10 05:45; Sellette 4 and 5 on 03/10 and 05/10 at night, and 07/10 at 22:45.
  • Twins Pont Ernen and Sellette 3: their corr flicker in quantised steps near zero (they follow their references). Not read.
  • Seen at the requalification, not labelled (the instruction was not to look for new events): Sellette 6 corr_y dips on 01/10 08:00–09:15 (−14 noises at 09:00, rolling 6 h), between FLORIDA-2 and FLORIDA-3. Sellette 6 is almost always rejected (it has a spanning Target issue, FLORIDA-1). To look at in a later pass.

3. The documentation

  • Helpful: families.md (a hindered reference: family of the cause), the reading rules on two references, and overview.py --rigid, which showed the reference slides the old labels had missed. Precedents JUNGFRAU-50/55 (sun behind a reference) and WYOMING-17.
  • Not covered by the docs: on a two-reference station, rigid residuals of a reference of ±0.1 px are frequent. There is no rule for the size of a joint jump that is worth a label; I used about 4 to 5 noises on at least two targets together.

Requalification

Targets read: Massif 2, Sellette 4, Sellette 5 and Sellette 6. The references have corr of 0, and the twins Pont Ernen and Sellette 3 have a quantised noise near zero, so they are not read. Outputs are in work/requal/; the scripts are scripts/requal.py (it keeps the largest range of each label from a ranges.py output) and scripts/nights.py (night medians before and after the rain).

  • python3 tools/scripts/ranges.py data/FLORIDA --columns corr_x,corr_y --start 2026-10-01T00:00Z --end 2026-10-11T09:45Z → ranges_all.txt; the same with --rolling 6h → ranges_roll6h.txt; scripts/requal.py on both → per_label.txt.
  • Thermal labels: ranges.py … --sensors "Massif 2,Sellette 4,Sellette 5,Sellette 6" --start <the night before> --end <the night after> --outside <label> --min-measurements 3 (level = the night level) → thermal_outside.txt.
  • FLORIDA-73: --rolling 4 on 09/10 08:00–13:30; FLORIDA-18: --outside 2026-10-08T03:15Z/… from 05/10, and scripts/nights.py → f73_f18.txt, f18_detail.txt, nights.txt.
id family verdict largest deviation (noises)
FLORIDA-2 Visibility removed none ≥5 (missing points of Massif 2)
FLORIDA-3 Thermal kept vs night: Massif 2 cx +7.9, Sellette 4 cx +7.4, Sellette 6 cx +6.6, Sellette 6 cy +6.1 (8 meas.)
FLORIDA-65 Visibility kept Massif 2 cy −34
FLORIDA-5 No data kept no measurement from any sensor
FLORIDA-66 Visibility kept Sellette 6 cy −12.5 (rolling −10.8)
FLORIDA-67 Visibility kept Massif 2 cx +12.3 (rolling +10.8), Sellette 5 cx +7.0
FLORIDA-68 Visibility kept Sellette 4 cx −6.3, Sellette 5 cx −6.2 (rolling −5.6, −5.5)
FLORIDA-69 Visibility removed missing points; the 10:45 jump belongs to FLORIDA-68
FLORIDA-8 Thermal kept vs night: Sellette 4 cy −15.9 (46 meas.)
FLORIDA-70 Visibility kept Massif 2 cy +6.2 (rolling +6.9)
FLORIDA-71 Visibility kept Sellette 4 cx −7.2; rolling: Massif 2 cx +6.0
FLORIDA-10 Thermal kept vs night: Massif 2 cx +9.0, Sellette 6 cx +8.2, Sellette 4 cy −8.9
FLORIDA-72 Displacement kept Massif 2, Sellette 5, Sellette 6 cx −5.6 (rolling Sellette 6 −5.7)
FLORIDA-11 Thermal kept vs night: Sellette 6 cx +14.4, Massif 2 cx +11.0, Sellette 6 cy +10.5
FLORIDA-12 Thermal kept vs night: Sellette 6 cx +12.3, Massif 2 cx +10.6, Sellette 6 cy +10.1
FLORIDA-13 Thermal kept vs night: Sellette 6 cx +8.6, Massif 2 cx +7.8
FLORIDA-15 Thermal rewritten vs night: Sellette 6 cy +9.7 (11 meas.)
FLORIDA-16 Visibility kept Sellette 4 cx +35, Massif 2 cy −30
FLORIDA-17 No data kept no measurement from any sensor
FLORIDA-18 Displacement kept shift of the night level +5.6 to +6.8 (Sellette 4, Sellette 5); within its bounds only 10:00
FLORIDA-19 Thermal removed no separable cycle; rolling 6 h: one point −5.9
FLORIDA-73 Visibility removed rolling 6 h / ±4: none ≥5; jump 4.5–4.7
FLORIDA-21 Thermal rewritten outside FLORIDA-22, vs night: Sellette 6 cy −9.0, Sellette 5 cx −7.1, Sellette 6 cx +6.5
FLORIDA-22 Visibility kept Sellette 4 cx +37, Massif 2 cy −26

validate.py: 0 errors. One overlap to look at, FLORIDA-65 (ends 08:45) with FLORIDA-66 (starts 08:30); it predates this pass and its bounds stay.

Difficulties

  • The noise of ranges.py depends on the window: a short window gives a larger noise, a night window a smaller one. Deviations here use the noise of the whole range or of the thermal window; FLORIDA-73 and FLORIDA-18 are close to the threshold whichever window is used.
  • Usage limit reached during the work. The thermal labels were checked on their hourly pose and temperature, not edge by edge at the native step.

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/FLORIDA.jsonl (labels near the range), work/existing.jsonl, work/spanning.jsonl, precedents (one query). Tools: inventory, overview (common, residuals, rigid), days --bounds, metrics, thumbnails; images 02/10 09:00 and 12:00. Requalification: instructions/analyst.md, labeling/labeling-rules.md section 1, tools/GUIDE.md (ranges), work/requalify_input.jsonl, labeling/auto/FLORIDA.jsonl (09–16/10); tools: ranges.py, inventory.py, tools/scripts/data.py (through scripts/nights.py).