Files
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

11 KiB
Raw Permalink Blame History

Labels: FLORIDA, 2026-10-22T03:15Z → 2026-11-01T00:00Z (UTC)

This analysis first checked the 19 existing labels (all reviewed 2026-11-17) and added FLORIDA-79. It was then requalified under the new rule (labeling-rules.md, section 1): only what stands out from the noise of the corrected curves is labelled, at more than 5 of the target's own noises. Result: 18 labels kept unchanged, 0 rewritten, 2 removed (FLORIDA-47 and FLORIDA-52). See "## Requalification" below.

Weather: FLORIDA has no weather/ folder, so no weather is available.

Targets (inventory.py): Sellette 3 (33 m, 6.5 mm/px, twin of p1_r), Sellette 4 (48 m, 9.6 mm/px), Sellette 5 (64 m, 12.8 mm/px), Sellette 6 (80 m, 15.9 mm/px, Target issue FLORIDA-1), Massif 2 (99 m, 19.6 mm/px, Target issue FLORIDA-14), Pont Ernen (373 m, 74 mm/px, twin of r1_r). Only two references, p1_r and r1_r. The twins are not labelled for themselves. Their corr sits at zero with a near-zero noise, so a deviation "in noises" means nothing for them. The Target issue targets are read but are not used alone to justify a label. The requalification therefore rests on Sellette 4 and Sellette 5.

Verdict per old id (first pass, then requalification)

  • FLORIDA-46 (22 03:15→03:45, Displacement): kept. Sellette 4 corr_x steps up at 03:30 and stays. Against the night before, this is +11.7 noises at the extreme, 22 measurements above 5 noises on the morning, and a jump of 6.8 noises.
  • FLORIDA-47 (22 08:45→17:00, Thermal): removed at requalification: the thermal movement is compensated (details in section 2).
  • FLORIDA-48 (22 11:15→13:30, Visibility, r1_r lost): kept. Sellette 5 reaches +6.3 noises at 13:00, Sellette 4 +5.2.
  • FLORIDA-49 (23 03:45→09:15, Visibility, r1_r lost): kept. Sellette 5 reaches +6.4 noises (07:00–09:00), Sellette 4 +5.5.
  • FLORIDA-50 (23 12:15→20:15, Visibility, mist): kept. Sellette 4 corr_x +36 noises, corr_y ±22 noises. Sellette 5 +23 noises.
  • FLORIDA-51 (23 22:00→24 01:45, Visibility): kept. Sellette 4 corr_x +7.3 noises (rolling) or +10.4 noises (window level), corr_y +7.4 noises.
  • FLORIDA-52 (24 01:45→02:15, Visibility, r1_r lost for one measurement): removed at requalification: the curves do not move (details in section 2).
  • FLORIDA-53 (24 07:00→14:45, Visibility, r1_r lost): kept. Against the level outside the episode, Sellette 5 reaches +7.5 noises (12:00–14:30) and Sellette 4 +6.9.
  • FLORIDA-54 (24 21:45→25 21:15, Visibility, rain): kept. Sellette 4 corr_x +51 and −18 noises.
  • FLORIDA-55 (26 08:30→12:00, Visibility, r1_r lost): kept. Sellette 5 reaches +6.7 noises (10:30–11:45).
  • FLORIDA-56 (26 11:45→19:00, Thermal): kept. Sellette 5 corr_y is above its night level by up to +7.9 noises at 14:00, with 15 measurements around the maximum.
  • FLORIDA-57 (26 14:30→15:15, No data): kept. The two steps at 14:45 and 15:00 are absent for all sensors.
  • FLORIDA-58 (27 01:45→11:45, Visibility, r1_r): kept. Sellette 5 reaches −7.2 noises at 02:15 and +5.3 noises at 06:15–06:45. Sellette 4 reaches −6.3 noises.
  • FLORIDA-59 (27 09:00→09:30, No data): kept. The 09:15 step is absent for all sensors.
  • FLORIDA-60 (27 11:30→20:00, Thermal): kept. Against the night level, Sellette 5 corr_x reaches −8.8 noises at 12:45 and Sellette 4 −6.5.
  • FLORIDA-61 (28 07:00→10:30, Visibility, r1_r lost): kept. Sellette 5 reaches +6.4 noises and Sellette 4 +6.1.
  • FLORIDA-79 (28 10:15→19:45, Thermal, not reviewed): kept. Sellette 5 corr_y reaches +6.4 noises over its night level at 11:15, with 17 measurements above 5 noises from 09:15 to 15:00.
  • FLORIDA-62, 63, 64 (29, 30, 31, Thermal): kept. Sellette 5 corr_y reaches +8.1, +8.8 and +9.1 noises over its night level.

1. My labels

2026-10-28T10:15Z → 2026-10-28T19:45Z (FLORIDA-79, added in the first pass, kept)

  • Phenomenon of the catalogue: The camera moves through a thermal effect
  • Precedents: FLORIDA-56 and FLORIDA-60 (same station; a thermal label that starts at the last rejection of the morning r1_r loss) and FLORIDA-62 to 64 (same shape). precedents.py was not run (usage limit in the first pass).
  • Family: Thermal displacement. The pose and Sellette 5 corr_y follow the housing temperature (19.7 → 24.3) and come back with it. error_value stays at its level.
  • Measured: rotation goes from about +0.01 at night to −0.014 at 12:00–14:00, then is back to +0.005 at 19:30. Tilt goes from +0.006 to +0.017. Pan stays flat. Sellette 5 corr_y is above its night level by up to +6.4 noises at 11:15 (days.py --bounds), with 17 measurements above 5 noises from 09:15 to 15:00. That is about +0.09 px, or about 1.2 mm.
  • Images: none (not looked at, usage limit in the first pass).
  • Confidence: medium. The cycle is clear on the pose, and it clears the threshold, but by a small margin.

2. What I left without a label

Removed at the requalification

  • FLORIDA-47, 2026-10-22T08:45Z → 2026-10-22T17:00Z, Thermal displacement: removed, compensated thermal. The pan drifts with the housing temperature, but the compensation absorbs it.
    • Sellette 5: against the night of the 21st, corr_y peaks at +2.9 noises (14:30). Its corr_x exceeds 5 noises only at 13:00–13:15 (+5.3), which is inside FLORIDA-48, the loss of r1_r.
    • Sellette 4: against its level after the step of FLORIDA-46 (ranges.py --rolling 1d), it passes 5 noises only at 06:00 (+5.1, a single point before the cycle) and at 13:00–13:15 (FLORIDA-48).
    • Sellette 6: its corr_y reaches +8.9 noises (15:00–17:00), but it is under its Target issue (FLORIDA-1). It was rejected from 09:00 to 14:45 that day, so its curve is unreadable and does not carry a label alone.
    • The label's own description already said that the corrected curves show almost nothing.
  • FLORIDA-52, 2026-10-24T01:45Z → 2026-10-24T02:15Z, Visibility: removed, missing points only. r1_r and Pont Ernen are rejected at 02:00 for one measurement. Against the local level (ranges.py --rolling 8, 23 20:00 → 24 05:00), nothing exceeds 5 noises around 02:00. At 02:00, Sellette 5 corr_x is 0.110 against a level of 0.110 (0 noises). Sellette 6 is at 0.180 against 0.156 (about +1 noise), and Sellette 4 at 0.190 against 0.190 (0 noises). Nothing moves in corr_y either. The earlier ranges in that window (Sellette 4 at 00:00–01:00, Sellette 5 at 22:45) belong to FLORIDA-51.

Unchanged from the first pass

  • 23rd, 09:15–12:15. Rotation goes to −0.02 as the temperature rises to 20.5, then the mist starts. This is −3.1 noises, and Sellette 5 corr_y does not move: no thermal label.
  • 24th. No thermal cycle above 5 noises against the night level: Sellette 5 corr_y peaks at +4.8 noises and corr_x at +2.3. The pose is mostly absent.
  • 25th. The cycle is masked by the rain (FLORIDA-54).
  • 29th–31st, evenings. Sellette 5 corr_y goes 6 to 8 noises below the window level from 18:00 to 23:45. This is the lower night level after the warm days (the next night minus the reference night is −1.2 to −2.8 noises), not an event of its own.
  • Sellette 6 and Massif 2. Both are under their Target issue labels (FLORIDA-1, FLORIDA-14). Their large excursions while accepted (for example Massif 2 corr_y +17 noises on the 26th to 28th) are read as part of these regimes.

3. The documentation

  • The 5-noise rule removed the two labels whose own description said there was no effect on the curves (FLORIDA-47) or only one missing point (FLORIDA-52).
  • For a jump-and-return, the level must be measured outside the episode (ranges.py --outside). The window level over ten days hid FLORIDA-53: Sellette 4 at 5.1 and Sellette 5 at nothing against the window, but +6.9 and +7.5 noises locally. For a step (FLORIDA-46), the level must be read against the night before. Against several days before, the step is only about 2.4 noises, because the earlier days sat at another level.
  • What is missing: the rules do not say whether a Target issue target (here Sellette 6, rejected half the time) may alone carry a movement label. I did not let it, following the thermal description rule ("without those that a Target issue makes unreadable").
  • On a two-reference station, every loss of r1_r gives Sellette 4 and Sellette 5 the same, repeatable offset, close to the threshold (about 5 to 7 noises). The one-measurement loss at 02:00 on the 24th gave none.

Requalification

Commands (from the repository root, --out in work/rq):

  • python3 tools/scripts/ranges.py data/FLORIDA --columns corr_x,corr_y --k 5 --start 2026-10-22T03:15Z --end 2026-11-01T00:00Z → work/rq_ranges_k5.txt
  • .venv/bin/python tools/scripts/days.py data/FLORIDA --columns corr_x,corr_y --sensors <target> --bounds --start 2026-10-21T00:00Z --end 2026-11-01T00:00Z, run for Sellette 4, Sellette 5, Sellette 6 and Massif 2 → work/rq_days_*.txt. This gives the thermal deviation from the night level.
  • ranges.py --sensors "Sellette 4","Sellette 5" --start 2026-10-21T20:00Z --end 2026-10-22T09:00Z --outside 2026-10-22T03:30Z/2026-10-22T09:00Z (FLORIDA-46), and with --rolling 1d to 23 03:00 (FLORIDA-47).
  • ranges.py --rolling 8 --start 2026-10-23T20:00Z --end 2026-10-24T05:00Z (FLORIDA-51, FLORIDA-52), plus summary.py --column corr_x --step 15min --start 2026-10-24T01:00Z --end 2026-10-24T03:00Z.
  • ranges.py --outside <label> on a window of ±6 h around each r1_r loss (FLORIDA-48, 49, 53, 55, 58, 61) → work/rq_outside.txt.

Values (largest deviation of a readable target, in its own noises):

id family deviation verdict
FLORIDA-46 Displacement S4 corr_x +11.7 (vs night before), jump 6.8 kept
FLORIDA-47 Thermal S5 +2.9 corr_y; S4, S5 > 5 only inside FLORIDA-48 removed
FLORIDA-48 Visibility S5 +6.3, S4 +5.2 kept
FLORIDA-49 Visibility S5 +6.4, S4 +5.5 kept
FLORIDA-50 Visibility S4 +36, S5 +23 kept
FLORIDA-51 Visibility S4 +10.4 corr_x, +7.4 corr_y kept
FLORIDA-52 Visibility ≤ 1 at 02:00 removed
FLORIDA-53 Visibility S5 +7.5, S4 +6.9 (outside level) kept
FLORIDA-54 Visibility S4 +51 / −18 kept
FLORIDA-55 Visibility S5 +6.7 kept
FLORIDA-56 Thermal S5 corr_y +7.9 kept
FLORIDA-57 No data 2 steps absent, all sensors kept
FLORIDA-58 Visibility S5 −7.2 / +5.3, S4 −6.3 kept
FLORIDA-59 No data 1 step absent, all sensors kept
FLORIDA-60 Thermal S5 corr_x −8.8, S4 −6.5 kept
FLORIDA-61 Visibility S5 +6.4, S4 +6.1 kept
FLORIDA-79 Thermal S5 corr_y +6.4 kept
FLORIDA-62/63/64 Thermal S5 corr_y +8.1 / +8.8 / +9.1 kept

python3 labeling/scripts/validate.py labeling/auto/analyses/FLORIDA_2026-10-22T0315_2026-11-01T0000/labels.jsonl: 18 labels, 0 errors.

Difficulties

  • First pass: the usage limit was reached, so no images were looked at for FLORIDA-79, no precedents query was run, and no GIF was made.
  • Requalification: the rule does not say which level to use for a step or a jump-and-return: the window, outside the episode, or the night before. I used the local level outside the episode, and the night before for the 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, work/requalify_input.jsonl, the previous analysis.md and report.md of this folder. Tools: inventory.py, ranges.py, days.py --bounds, summary.py, plot.py.