Files
aimsight-labeling/auto/analyses/CLARIDEN_2026-10-20T0400_2026-11-01T0010/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

20 KiB
Raw Blame History

Labels: CLARIDEN, 2026-10-20T04:00Z → 2026-11-01T00:10Z (UTC)

Requalified 2026-11-17 under the new rule (labeling-rules.md, section 1: only what stands out from the noise of the corrected curves, 5 noises). 21 labels kept unchanged, 0 rewritten, 3 removed (CLARIDEN-27, CLARIDEN-43, CLARIDEN-67); details in section 2 and in "Requalification". The verification text below is the earlier work. Two of its entries are out of date: the 30/10 18:10 → 18:30 label (later CLARIDEN-67) and the thermal entries kept under D-117 (CLARIDEN-27, -43).

The task was to verify the 21 labels already on the range (work/existing.jsonl, all reviewed 2026-11-17) and the one spanning label (CLARIDEN-1).

Result: 12 kept unchanged, 9 corrected, 0 removed, 3 added; 24 labels in labels.jsonl. Labels are built from existing.jsonl by scripts/build_labels.py, so the kept ones are byte-identical apart from the dropped id. validate.py reports 0 errors. Its 8 "overlaps to look at" are distinct events on different target rows (row C drift, row B moves, row A jolts and returns). A script checked that every bound is a measurement instant of the station (10 min rate, minutes :00 to :50).

Verdict per existing id

id verdict evidence
CLARIDEN-23 kept Residuals: D-301-C corr_x rises over 20→24/10. D-304-C and D-303-C corr_y rise from 22/10 to 25/10 (6 h medians: D-304-C from about −0.42 to −0.36). Rigid dx of D-301-C goes up to +0.13, so these targets move in the image relative to the others. Temperature flat (39.6 to 40.6 °C over the range).
CLARIDEN-24 kept Inventory: one step missing between 06:10 and 06:30.
CLARIDEN-25 corrected Bounds were 20:45 → 21:45, which are not measurement instants. The present instants are 20:40 and 21:40 (inventory: 5 missing). Now 20:40 → 21:40; description unchanged.
CLARIDEN-26 corrected End was 08:45, not an instant. The last touched points are at 08:40 (D-303-B corr_y −0.042 against a level near +0.03; D-301-C corr_x −0.025). At 08:50 D-301-C is back and D-303-B is merging into the row B rise of CLARIDEN-28. End now 08:50.
CLARIDEN-27 kept The scale reaches its minimum at 08:30–08:40 with the temperature dip of the 22nd (hourly 39.58 °C at 06:00). With days.py smoothing 5 the deviation is −4.5 noises, just under the rule of 5 noises. It is kept because it matches the reviewed precedents CLARIDEN-5 and CLARIDEN-13 (sharp housing drops in the tunnel) and individual points exceed the threshold.
CLARIDEN-28 corrected Start was 08:45, not an instant. The row B rise is first seen after the visibility episode (D-306-B and D-303-B corr_y from 09:00). Start now 08:50.
CLARIDEN-29 kept D-306-B corr_y drops on the evening of the 22nd and the morning of the 23rd: rolling ranges, residual chart, rigid dy −0.166 at 23/10 16:10.
CLARIDEN-30 kept D-306-B and D-304-B corr_y partly rise again during the night of the 24th (6 h medians, residual chart).
CLARIDEN-31 kept Pose jolts at 24/10 19:20 and 25/10 06:30–07:20 (days.py: pan −5.1 and −7.5 noises). Row A corr_y +0.06 to +0.1, corr_x −0.04 to −0.05 (6 h medians). The corr_x of row A stays shifted until the end of the range (D-301-A about −0.14 → −0.18). Row A corr_y also drifts up a little on the morning of the 24th; this is not enough to move the start.
CLARIDEN-32 kept D-304-C corr_y goes from −0.36 (25/10) to −0.43 (29/10), D-303-C from −0.04 to −0.09, then flat.
CLARIDEN-33 corrected End was 15:15, not an instant. Instants present at 14:30 and 15:10, 3 missing. Now 14:30 → 15:10.
CLARIDEN-34 kept 09:00 and 09:30 present, 2 missing.
CLARIDEN-35 kept PR-284 corr_x jumps +15.3 (570 noises) and corr_y +11.9 at 12:10 only. Bounds 12:00 → 12:20 are right.
CLARIDEN-36 kept Hourly corr_y: D-306-B goes from +0.045 to −0.015 and D-304-B from −0.047 to −0.07 between 18:00 and 20:00 on 28/10. D-303-B corr_x steps down (residual chart).
CLARIDEN-37 kept 14:50 and 15:10 present, 1 missing.
CLARIDEN-38 corrected Bounds were 23:45 → 00:15, not instants. At 10 min the jump is between 23:50 (old level: D-302-B corr_y 0.050, PR_BG 0.050) and 00:00 (D-302-B 0.247, PR_BG −0.033). Raw y does not change; relative_y jumps (D-302-B −0.15 → +0.06). Now 23:50 → 00:10; description unchanged.
CLARIDEN-39 corrected End was 17:45, not an instant. Present at 17:00 and 17:50, 4 missing → 17:00 → 17:50. The description only said "cause not established", against rule 5 ("A No data states the metrics and the image…"). Now it states that the gap is CLARIDEN's own: ALASKA, FLORIDA, JUNGFRAU, LENZSPITZE, NADELHORN, NORDEND, SNCF01 and WYOMING all measured. last_sync has no entries from 17:09 to 17:40, so the station itself stops recording. The 17:00 image exists. BusVoltage stays at 14.35–14.36 V.
CLARIDEN-40 corrected Start was 13:45, not an instant. First touched point at 14:00 (D-301-C corr_x −0.093 against +0.01 at 13:50; D-302-B corr_y 0.315 against 0.262). Start now 13:50. End 15:10 holds: at 15:00 D-302-B corr_y is still 0.30, halfway; at 15:10 it is back at 0.265.
CLARIDEN-41 corrected Start was 18:45. First touched point at 18:50 (D-301-C corr_x −0.070, D-302-B corr_y 0.33, PR-295 corr_x 0.054). Start now 18:40. End 20:00 is right: PR-295 is rejected at 19:50.
CLARIDEN-42 corrected Start was 22:15. PR-295 is rejected from 22:30 (a touched reference) and is not rejected at 22:20. Start now 22:20; end 00:00 holds (P-301-B −5 noises at 23:40, curves crossing until 23:50).
CLARIDEN-43 kept Scale min at 23:30 (0.999787) with the housing drop (hourly 38.4 °C at 23:00). Pattern of the reviewed CLARIDEN-5 and -13.
CLARIDEN-1 (spanning) holds P-300-B is rejected at all 1690 instants of the range (inventory). Nothing to say.

1. My labels

Kept and corrected labels are justified in the table above. Below are the three added labels.

2026-10-25T19:10Z → 2026-10-29T04:00Z

  • Phenomenon of the catalogue: "Targets move slowly, alone, then come back" (the return leg).
  • Precedents: CLARIDEN-15 (return of the row C drift of 6–11/10, labelled separately and naming the outward label) and CLARIDEN-32 (same dates, row C). Both guided the wording and the "return labelled apart" structure.
  • Family: Displacement. The cause is not established. Temperature is flat (39.6 to 40.6 °C), so this is not Thermal displacement. The image is good and nothing hides the targets, so it is not Visibility.
  • Measured: 6 h medians of corr_y.
    • D-304-A goes from −0.03 (23/10) to +0.09 (25/10 12:00), +0.058 (26/10 06:00), +0.022 (28/10 00:00), −0.020 (29/10 00:00), then fluctuates between −0.045 and +0.007.
    • D-303-A goes 0.05 → 0.13 → 0.05.
    • D-301-A goes −0.09 → −0.02 → −0.09.
    • D-302_A goes 0.03 → 0.09 → 0.02.
    • P-299-A and P-298-A follow less.
    • corr_x of row A stays shifted, as CLARIDEN-31 says.
    • Rigid residual dy: D-304-A max +0.126 at 25/10 19:50 and +0.037 at the end; D-303-A +0.084 → +0.014; D-301-A +0.072 → +0.026. The targets come back in the image with their curves.
    • No pose step on these days.
    • The end is aligned on CLARIDEN-32. At 29/10 04:00 the hourly values are within the later fluctuation.
  • Images: none looked at specifically for these days; the 28/10 19:00 and 20:00 images were described in CLARIDEN-32.
  • Confidence: medium. The return is clear on the 6 h medians, but it is gradual and the end is fuzzy within a few hours.

2026-10-30T18:10Z → 2026-10-30T18:30Z

  • Phenomenon of the catalogue: "A reference dropped: the base changes", in a variant where the reference is not dropped but shifts under its threshold.
  • Precedents: WYOMING-17 (reviewed): references jump alone for one measurement, the targets' raw positions do not move, the compensation passes the jump into the targets; Displacement, cause not established. NIEDERHORN-126: same family and wording for a reference rejected once.
  • Family: Displacement. No object is seen or deduced, so not Obstruction. The scene is not degraded (error_value of the targets flat), so not Visibility.
  • Measured:
    • At 18:20, corr_x jumps against 18:10: D-300-A −0.117 → −0.211, D-301-A −0.167 → −0.236, D-302_A −0.266 → −0.340, D-303-A −0.036 → −0.126, D-304-A −0.105 → −0.200, P-298-A −0.045 → −0.093, P-299-A −0.098 → −0.140, PR-283 and PR-284 about −0.06. D-303-C and D-304-C move about +0.03.
    • Everything is back at 18:30.
    • Raw x of the targets does not change (D-301-A 531.43; D-303-A 294.6).
    • Reference PR-291: raw x 991.53 → 991.63 → 991.57; rigid dx +0.0845, its maximum of the range. Its error_value is 0.089 → 0.142 → 0.121, under the 0.7 threshold. The other references are unchanged.
    • corr_y is not visibly affected.
  • Images: 18:00 (whole image) shows the targets and the frame at the top, nothing in front. 19:00 exists. No image at 18:20.
  • Confidence: medium-high on the mechanism (PR-291 alone), low on the cause.

2026-10-31T09:20Z → 2026-10-31T15:00Z

  • Phenomenon of the catalogue: "The camera moves through a thermal effect". It is labelled even when compensated (D-117).
  • Precedents: CLARIDEN-5, -13, -27, -43 (reviewed): scale against housing temperature, always on drops. This one is a warming, so the direction is opposite.
  • Family: Thermal displacement. The pose follows the housing temperature both ways, and the error_value does not rise during the warming. The return coincides with the Visibility episode CLARIDEN-40: two labels, as on the 22nd.
  • Measured:
    • Housing temperature, hourly: 07:00 40.05, 09:00 40.43, 11:00 40.63, 13:00 40.78, 14:00 40.30, 15:00 40.20 °C.
    • Camera scale at 10 min: 0.999857 at 09:40, 0.999881 at 09:50, about 0.99989–0.99991 until 13:50, 0.999878 at 14:10, 0.999847 at 15:00.
    • Pan, hourly: 0.0042 (09:00), 0.0027 (10:00–13:00), back to 0.0044 at 15:00.
    • days.py --bounds --smoothing 5: pose_scale 31/10 exceeds 5 noises from 10:50 to 12:50, max +6.0 at 10:50. It is the only day of the range where the pose exceeds 5 noises.
    • Rigid scale of the targets: 1.000012 → 1.000033 over the same hours.
    • Corrected curves, 30 min medians:
      • D-301-C corr_x from about 0.00 to +0.04 from 09:30;
      • D-304-C corr_y from −0.42 to −0.39 from 09:40;
      • D-303-C corr_y from −0.088 to −0.07;
      • D-304-B corr_y from −0.06 to −0.04 from 09:30;
      • all back by 15:00–15:10.
    • Start: 09:20, the measurement before the first lasting deviation (09:30). The 09:10 point of D-301-C is isolated and within the noise.
    • End: 15:00, pose back.
  • Images: 12:00 (whole image) shows a lit object at the far end of the tunnel between the target rows, nothing in front of the targets. 09:00 and 14:00 also exist.
  • Confidence: medium. The co-movement of scale, pan and temperature is clear, but small (the housing temperature varies by less than a degree).

2. What I left without a label

  • 20→22/10: row B in x drifts slowly (6 h medians). D-306-B goes from +0.042 to +0.003, about 3.5 noises; P-301-B from −0.02 to −0.04; D-303-B a little; P-299-B goes up and back. In the image (rigid dx), D-306-B barely moves against the other targets (−0.006 → −0.018), so most of it comes from the base. It is small, without dates of its own, and overlaps the reviewed row B labels. If the expert wants it, it would continue the D-306-B horizontal decline already named in CLARIDEN-14.

  • 20/10 18:40 and 21/10 03:20: row A corr_y minima (overview window minima of −5.6 to −7 noises). They do not exceed 5 noises against a rolling level of ±6 measurements; they are a slow wander.

  • Single-sensor points: P-298-B corr_y at 31/10 02:40 (+5.3 noises), PR-284 corr_y at 25/10 02:50 (+5.5). They are single points on the two noisiest sensors, with no other sensor and no reference moving.

  • ranges --k 4 --instants 4: every group instant falls inside a labelled moment. The 30/10 18:20 group stays under 4 noises per sensor there but was found by its coherence; it is labelled.

  • Weather: CLARIDEN has no weather/ folder (a tunnel with no Swiss position), so no weather is available. The instruction in effect (re-read after it changed during the session) reads the weather from data/<STATION>/weather/.

  • Removed by the requalification (2026-11-17):

    • CLARIDEN-27, 2026-10-22T08:00Z → 09:20Z, Thermal displacement. Its own description says the camera moves and the compensation absorbs it. Level and noise taken outside the label, window 21/10 20:00 → 22/10 09:20: no target is beyond 5 noises over 08:50 → 09:20, the part after the view obstruction. Earlier in the label, the corrected curves do pass 5 noises (D-301-C corr_x −12.1, D-303-B corr_y −8.8, D-302-B/PR_BG corr_y +8.5, up to 08:50), but that is the bump of the Visibility label CLARIDEN-61, which stays. This is a compensated camera thermal movement, so it is not labelled.
    • CLARIDEN-43, 2026-10-31T23:10Z → 2026-11-01T00:10Z, Thermal displacement. Same case. Level from 31/10 12:00, no target beyond 5 noises over 23:20 → 00:10. The deviations of 22:40 → 23:20 (P-301-B corr_x +11.2, D-301-C −9.4, D-302-B corr_y +7.7, D-304-B −7.0) belong to the Visibility label CLARIDEN-71, which stays. Compensated thermal: removed.
    • CLARIDEN-67, 2026-10-30T18:10Z → 18:30Z, Displacement (the one-measurement jump through reference PR-291). Window 30/10 06:00 → 31/10 06:00, level and noise outside the label. The 18:20 jumps of row A in corr_x are D-300-A 4.2 noises (−0.094 px against a noise of 0.019), D-303-A 3.7, D-304-A 3.6, D-301-A about 3.7 and P-299-A about 1.8. They are the size of the ordinary jumps of the same day: largest jumps of 4.0 to 5.2 noises at other hours. ranges.py --k 5 finds nothing in corr_x or corr_y, and --k 2 finds nothing on the named targets at 18:20. It is under the threshold, so it is removed. The earlier analysis already noted "under 4 noises per sensor, found by its coherence".

Requalification

Commands, from the repository root:

  • python3 scripts/requalify.py <folder> 5 outside. For each non-No data label, this runs tools/scripts/ranges.py data/CLARIDEN --columns corr_x,corr_y --k 5 on the label ± max(12 h, its length), with --outside <label> (level and noise outside the label). It keeps the ranges inside the label. Output: work/rq/outside_k5.txt.
  • For the steps and drifts, the level must be the one before the label, so the run uses --outside <label start>/<window end> on the named targets (work/rq/before_level.txt, work/rq/extra.txt, work/rq/c36.txt).
  • For the thermal labels, the level is the night: --outside the thermal hours, with the window starting the evening before (work/rq/c68.txt).

Values (largest deviation in noises of a named target inside the label):

id family deviation verdict
CLARIDEN-23 Displacement D-304-C corr_y +17.0, D-303-C +16.4, D-301-C corr_x +10.6 (level before 20/10) kept
CLARIDEN-24, -60, -64, -34, -37, -66 No data every sensor missing (inventory, earlier analysis) kept
CLARIDEN-61 Visibility D-301-C corr_x −13.7, D-302-B/PR_BG corr_y +9.0 kept
CLARIDEN-27 Thermal displacement nothing beyond 5 after the view obstruction; compensated removed
CLARIDEN-62 Displacement D-304-B corr_y +6.7 to +7.2, D-306-B +6.9 (window level), D-303-B +6.2 kept
CLARIDEN-29 Displacement D-304-B corr_y −7.6, D-302-B/PR_BG −7.5, D-303-B −6.0 (level before 17:00) kept
CLARIDEN-30 Displacement D-306-B corr_y +5.2 at 04:50 (level before 01:00); D-304-B +5.7 just after kept, near the threshold
CLARIDEN-31 Displacement D-306-B corr_y +13.4, D-302_A +8.4, D-304-B +8.4 kept
CLARIDEN-32 Displacement D-303-C corr_y −5.9, D-304-C −5.6, in repeated ranges 27/10 → 29/10 (level before 25/10 18:00) kept, near the threshold
CLARIDEN-63 Displacement P-298-A corr_y −12.8, D-301-A −10.6, D-304-A −9.6 (level before 25/10 19:10) kept
CLARIDEN-35 Target issue PR-284 corr_x +544, corr_y +449 kept
CLARIDEN-36 Displacement D-306-B corr_y −10.1 from 19:30, D-304-B −5.8 (level before) kept
CLARIDEN-65 Target issue D-302-B corr_y +20.2, corr_x −8.8, PR_BG corr_y −8.1 (level before) kept
CLARIDEN-67 Displacement row A corr_x 3.6 to 4.2 at 18:20 removed
CLARIDEN-68 Thermal displacement D-301-C corr_x +5.5 (09:40), +5.2 (10:20), +5.1 (11:20) against the night level, 4.1 to 4.7 in between until 13:00; D-304-C corr_y +5.2 (09:50); D-303-C corr_x −5.5 (10:50) kept: the corrected curves keep part of the cycle, near the threshold
CLARIDEN-69 Visibility D-301-C corr_x −9.7, D-302-B/PR_BG corr_y +6.2 kept
CLARIDEN-70 Visibility D-301-C corr_x −7.6, P-299-A +6.2, D-304-A +5.5 (reference PR-295 −10.0) kept
CLARIDEN-71 Visibility D-302-B/PR_BG corr_x +12.0, D-303-B +11.9, P-301-B +11.7 kept
CLARIDEN-43 Thermal displacement nothing beyond 5 after the view obstruction; compensated removed

No kept label rests in its description on what is no longer labelled, so none is rewritten. The kept ones are byte-identical to the input, without id, and keep their review. validate.py: 21 labels, 0 errors, 8 overlaps between distinct row events.

3. The documentation

  • What helped:
    • Rule 2 ("bounds are measurement instants, at the minute") drove 8 of the 9 corrections. The existing labels carried 15-minute-grid minutes (:15, :45) on a 10-minute station.
    • Rule 5 ("A No data states the metrics and the image") drove the rewrite of CLARIDEN-39.
    • The last_sync metric (one entry per minute) was decisive in showing that the station itself went silent from 17:09 to 17:40.
    • overview.py --rigid separated "the reference PR-291 moves" from "the targets move" on 30/10 18:20.
    • days.py --bounds gave the only thermal day above 5 noises.
  • Contradiction to flag: CLARIDEN-27 and CLARIDEN-43 (reviewed) stay at −4.5 and −3.5 noises on the smoothed scale. Strictly, the rule "no thermal label under 5 noises" would remove them. I kept them because the user validated them and individual points do exceed it. Should the rule be read on smoothed or raw points?
  • Precedents: WYOMING-17 was the right model for the reference-shift jump. The CLARIDEN thermal precedents only cover housing drops; the 31st is the first warming case. Nothing contradicted by the data.
  • validate.py does not check that bounds fall on the station's measurement instants; this is how the :15/:45 bounds passed. A check against the station's instants would catch them.

Difficulties

  • The instructions file and tools/GUIDE.md changed on disk during the session. The weather moved from a fetch command to data/<STATION>/weather/; CLARIDEN has neither a stations.txt place nor a weather folder, so the outcome is the same.

Sources read

Requalification: instructions/analyst.md, tools/GUIDE.md (ranges), labeling/labeling-rules.md sections 1–3, work/requalify_input.jsonl, this analysis, report.md, inventory.py (distances, mm/px).

Earlier work: instructions/analyst.md (twice), tools/GUIDE.md, tools/DATA.md, labeling/families.md, labeling/labeling-rules.md, labeling/reading-rules.md, labeling/stations.md, labeling/auto/CLARIDEN.jsonl (labels at and around the range: CLARIDEN-1, 5, 11–15, 20–22), work/existing.jsonl, work/spanning.jsonl, request.json, ~/aimsight-data/extraction/stations.txt (CLARIDEN line, no place). Precedents: precedents.py (WYOMING-17, NIEDERHORN-126, CLARIDEN-11, NORDEND-65/89, SNCF01-42). Tools: inventory, overview (common, residuals, rigid at 10 min and 6 h, with --follow), ranges (--rolling 6, --instants 3 and 4), days --bounds --smoothing 5, metrics (3 h and 1 h), summary (10 min, 30 min, 1 h, 6 h), plot. Raw: displacement CSVs (instants of all stations around the gaps), metrics/ BusVoltage, last_sync, link, GPU on 30/10. Images: 20261030T1800Z, 20261031T1200Z (whole). No quality metrics on this camera (empty columns).