Files
aimsight-labeling/auto/analyses/NORDEND_2026-10-13T0745_2026-10-24T1630/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

17 KiB
Raw Blame History

Labels: NORDEND, 2026-10-13T07:45Z → 2026-10-24T16:30Z (UTC)

Request: requalify the 44 labels of this analysis (work/requalify_input.jsonl, verified 2026-11-17) under the new rules (labeling-rules.md section 1, families.md): keep only what stands out from the noise of corr_x, corr_y. Not a new analysis: no new events searched, no bounds moved. DETECTORS: no. The previous analysis is kept in work/analysis.before-requalify.md, the previous report in work/report.before-requalify.md.

Weather: there is no data/NORDEND/weather/ folder (no MeteoSwiss data for this station), so I have no weather. The weather comes from the images only. No image-quality metrics on this camera (inventory.py: export peak to snr empty).

Result: 35 labels remain (31 kept unchanged with their id and review, 4 rewritten), 9 removed.

1. My labels

Kept unchanged (31)

Each holds by the threshold: at least one target more than 5 of its own noises from its level over the label (scripts/requalify.py, output work/requalify.txt; thermal labels: scripts/thermal.py, output work/thermal.txt). Description, bounds, family and review field unchanged.

Id Bounds Family Largest deviation (target, column, noises, instant)
NORDEND-181 13/10 07:45 → 14/10 05:00 Thermal PR18 corr_x −7.4 (smoothed, against the night before), 15 measurements 11:00-15:45; PR16 corr_x −24 (ranges, night noise)
NORDEND-182 13/10 09:15 → 11:00 Visibility PR16 corr_y +13.1 at 10:30; PR16 corr_x +11.5; PR15 corr_y −10.4
NORDEND-265 13/10 12:15 → 12:45 Visibility PR08 corr_x −12.9, PR13 −9.8, PR16 −9.1 at 12:30
NORDEND-184 14/10 05:00 → 06:15 Visibility PR12 corr_y +105, PR12 corr_x +101 at 05:15
NORDEND-185 14/10 06:15 → 15/10 07:15 Thermal PR18 corr_x −13.8 at 11:30 (33 measurements), PR13 corr_y −12.9, 19 targets beyond
NORDEND-187 15/10 07:15 → 22:30 Thermal PR18 corr_x −13.0 at 11:00, PR06 corr_y −11.8, PR16 corr_x −11.7, 17 targets beyond
NORDEND-189 15/10 22:30 → 16/10 01:15 No data 10 instants missing for every sensor (inventory.py)
NORDEND-190 16/10 05:45 → 17/10 08:45 Thermal PR18 corr_x −15.5 at 10:45 (42 measurements), 18 targets beyond
NORDEND-192 16/10 13:15 → 13:45 No data 13:30 missing for every sensor
NORDEND-193 16/10 14:00 → 14:30 No data 14:15 missing for every sensor
NORDEND-266 16/10 19:00 → 23:00 Visibility PR08 corr_y −142 at 20:00, PR06 corr_x −94
NORDEND-197 17/10 08:45 → 18/10 07:15 Thermal PR10 corr_y −9.4, PR11 corr_y −9.1 at 10:30; 9 targets beyond
NORDEND-198 17/10 12:45 → 15:45 Visibility PR15 corr_x +28.5 at 14:30, PR16 +25.5, PR13 +20.9
NORDEND-200 18/10 07:15 → 19/10 06:15 Thermal PR15 corr_x −7.3, PR18 corr_x −6.7 at 09:00
NORDEND-202 19/10 06:15 → 20/10 07:00 Thermal PR18 corr_x −11.2 at 12:00, PR16 −9.6; 12 targets beyond
NORDEND-203 19/10 08:45 → 09:15 No data 09:00 missing for every sensor
NORDEND-204 19/10 09:30 → 10:00 No data 09:45 missing for every sensor
NORDEND-205 19/10 20:30 → 21:00 No data 20:45 missing for every sensor
NORDEND-206 20/10 07:00 → 21/10 07:15 Thermal PR18 corr_x −9.0 at 11:45 (28 measurements), 11 targets beyond
NORDEND-267 20/10 15:15 → 17:15 Visibility PR07 corr_x +14.6 at 15:30, PR10 corr_y +8.3 at 15:45
NORDEND-268 20/10 23:45 → 21/10 01:45 Visibility PR08 corr_x +20.5 at 00:30, PR15 corr_x −15.8, PR16 corr_y −12.2
NORDEND-269 21/10 10:15 → 12:15 Visibility PR16 corr_x +14.5 at 11:15, PR14 +13.1, PR15 −12.7
NORDEND-270 21/10 14:15 → 17:15 Visibility PR15 corr_x +41.3 at 16:45, corr_y +13.3
NORDEND-271 22/10 08:30 → 09:00 Displacement PR03 corr_y +21.0, PR02 corr_x +19.1 at 08:45
NORDEND-272 22/10 09:15 → 11:00 Visibility PR05 corr_y +32.5, corr_x −27.6 at 09:30; PR14 corr_x +20.2
NORDEND-273 22/10 12:00 → 14:00 Visibility PR08 corr_x +29.8 at 12:15, PR07 −12.5 at 12:30
NORDEND-275 23/10 05:45 → 07:45 Visibility PR10 corr_y −16.0 at 06:30, PR08 corr_x −13.1, PR15 +12.9
NORDEND-213 23/10 08:45 → 24/10 03:45 Visibility PR11 corr_y +29.8 at 10:00, PR10 corr_x +27.0, PR06 corr_y −22.0 (rolling 8)
NORDEND-276 23/10 14:15 → 14:45 Displacement the lasting step: night of the 24th against the night before the rain, PR06 corr_y −10.5 (18 measurements), PR02 corr_x +8.9, PR03 +7.2, PR05 corr_y −8.7
NORDEND-278 24/10 14:00 → 14:30 Displacement PR13 corr_y −8.6 at 14:15, PR17 −6.0
NORDEND-279 24/10 15:00 → 15:30 Displacement PR08 corr_x −10.4, PR15 corr_y +10.0 at 15:15

NORDEND-198 still says that PR09 and CTR01 are rejected "from 13:00 and into the night (separate label)". That separate label (199) is now removed. The label is right in substance and the user validated it, so I kept it as it is. The phrase is a minor inaccuracy.

2026-10-14T15:15Z → 2026-10-14T16:00Z (NORDEND-186, rewritten)

  • Phenomenon of the catalogue: The sun in the field or behind a spot; A reference loses its brightness (here: lit background).
  • Precedents: the label itself (05/10 signature of CTR01 in the sun), NORDEND-198 (sun on the quarry, scene disturbed).
  • Family: Visibility, unchanged: the sun lights the background of the references, nothing hides them.
  • Measured: error_value CTR03 0.09 (14:30) → 0.196 (15:15), 0.212 (15:30), 0.207 (15:45), 0.172 (16:00), 0.095 (16:15). CTR01 rejected 15:30-15:45 (0.63, 0.59). CTR02, CTR04 at their level. The references' corr_y move apart: CTR02 0.088 → 0.114, CTR03 −0.073 → −0.098. corr_y PR13 0.110 (15:15) → 0.075 (15:45) → 0.083 (16:00) → 0.104 (16:30); PR11 0.122 → 0.095 → 0.086; PR12 0.128 → 0.098. In noises: PR13 −7.7 (window ±3 h, --outside), −7.2 (--rolling 8); PR12 −5.3; PR11 −5.7 at 16:00. Magnitude about 1 mm (PR13: 0.03 px × 43 mm/px). The effect lasts until 16:00-16:15 (PR15 −7.4 at 16:15), one measurement past the end. The bounds were not moved, as the request says.
  • Images: zone thumbnails 14:30, 15:30, 16:15 (sun-references.png): at 15:30 the rock behind CTR01, CTR02 and CTR03 is in full sun (background 96 against 15 and 10). It is in shadow at 14:30 and 16:15. The whole image of 15:30: sunny day, the wall of the targets in shadow.
  • Why rewritten: the old text was about CTR01's missing curve, which is no longer labelled. The episode does move the other curves, through reference CTR03, beyond the noise: the description now says that.
  • Confidence: medium. The deviation is a few noises above the threshold. On the same afternoon, PR13 makes a −7.4 excursion at 12:15 without a cause.

2026-10-21T07:15Z → 2026-10-22T06:30Z (NORDEND-208, rewritten)

  • Phenomenon of the catalogue: The camera moves through a thermal effect.
  • Precedents: NORDEND-210 (the next day, the camera does not come back).
  • Family: Thermal displacement, unchanged.
  • Measured: the corr_y of PR06, PR05, PR01, PR11 and PR12 drop together in the evening (from about 18:00-20:00) and stay low until 05:00. Against the night before, in noises over the label (smoothed by 3): PR06 −6.6 at 00:15 (10 measurements), PR05 −6.0 (14), PR11 −5.5, PR01 −5.3. ranges.py with the night level and night noise gives PR05 corr_y −10.7 (39 measurements) and PR12 −10.5 (28). About 1-2 mm (PR06 0.032 px × 38 mm/px; PR01 0.036 × 59). By day the curves stay within a few noises. The old label said the pose goes beyond its level at night (pan +13 noises at 01:15).
  • Why rewritten: the old text said "the corrected curves stay flat". The night remainder is a coherent group movement beyond 5 noises: the label holds, its description did not.
  • Confidence: medium (close to the threshold, but on five targets together).

2026-10-22T06:30Z → 2026-10-23T05:15Z (NORDEND-210, rewritten)

  • Phenomenon of the catalogue: The camera moves through a thermal effect.
  • Family: Thermal displacement, unchanged.
  • Measured: between 07:30 and 09:00, the corr_y of PR01, PR12, PR05, PR10, PR11 and PR06 rise together by about 0.03-0.05 px (1.5-3 mm). They stay up all afternoon and night, the mirror of the drop on the 21st evening. PR01 corr_y +6.2 at 14:45 (9 measurements), PR12 +6.0, PR05 +5.9; PR18 corr_x −6.1 (11 measurements, 09:00-20:00). With the night noise: PR12 corr_y +21 for 74 measurements. The PR02/PR03 jump of 08:45 and the episodes of 09:30 and 12:15 are separate labels (271, 272, 273), left out of the reading.
  • Why rewritten: same as 208.
  • Confidence: medium.

2026-10-24T07:30Z → 2026-10-24T16:00Z (NORDEND-214, rewritten)

  • Phenomenon of the catalogue: The camera moves through a thermal effect.
  • Family: Thermal displacement, unchanged.
  • Measured: corr_y of PR13 0.083 (08:00) → 0.105-0.12 (12:30-16:15), +9.3 noises at 15:30 (9 measurements); PR06 0.035-0.045 → 0.07-0.08; PR18 corr_y +6.2. About 1.5 mm (PR13 0.035 px × 43 mm/px). Not back at the end of the label. The night level is the night of the 24th, after the base step of the 23rd, so the step does not enter it. The one-measurement jumps of 14:15 and 15:15 (278, 279) are excluded.
  • Why rewritten: the old text said "the curves stay flat".
  • Confidence: medium.

2. What I left without a label

Removed labels

Id Bounds Family Why Measured
NORDEND-183 13/10 16:00 → 14/10 07:15 Visibility missing points only: PR09, CTR01, PR04, PR07 rejected at night. None is a reference of the compensation (CTR01 is local), so their loss cannot move the other curves --rolling 8 on 16:00 → 05:00: scattered one- or two-measurement deviations of other targets, 5-10 noises (PR14 corr_y −10.7, PR16 corr_x −10.4, PR18 corr_y +10.2 at 16:15-17:00); the mist from 05:15 is NORDEND-184
NORDEND-188 15/10 15:30 → 16:00 Visibility missing point only: CTR01 rejected at 15:45; CTR03 not disturbed (error 0.08-0.12) PR08 corr_y −5.4 for one measurement (±3 h window); none beyond 5 noises with --rolling 8
NORDEND-191 16/10 08:00 → 08:30 Visibility missing point only: CTR01 rejected once nothing beyond 5 noises
NORDEND-194 16/10 19:00 → 17/10 10:00 Visibility missing points only (left-edge night) the big jumps 19:00-23:00 are NORDEND-266; after 23:00 (--rolling 8): PR15 corr_x +10.2 (23:15-00:15), PR08 +9.0, PR16 +7.6, PR10 corr_y −7.9 at 09:30, short and scattered, not on the rejected targets
NORDEND-199 17/10 12:45 → 18/10 02:30 Visibility missing points only 12:45-15:45 is NORDEND-198; after 15:45: PR16 corr_x −8.8 (15:45-16:30), PR10 +5.9, CTR01 +6.2 at 01:00, PR06 +5.4 at 02:15
NORDEND-201 18/10 16:45 → 19/10 11:00 Visibility missing points only PR18 corr_y ±9.3 (16:45-19:15, the evening thermal return under the rolling level), PR07 corr_x +9.3 at 22:45 (2 measurements), PR05 +8.2, scattered 5-7 noise points
NORDEND-207 20/10 20:15 → 21/10 05:30 Visibility missing points only the 00:00-01:30 deviations are NORDEND-268; PR07 corr_y +11.7 at 20:30-21:00 (3 measurements, the faint target itself, under its threshold), PR17 +12.7 at 21:00-21:45 (PR17 has its spanning Obstruction NORDEND-169)
NORDEND-274 22/10 20:15 → 23/10 08:00 Visibility missing points only before the rain (to 05:45): PR01 corr_x +9.3 at 21:30, PR07 +8.2 at 02:45, PR13 +6.3, PR18 +6.1, all single points; 05:45-07:45 is NORDEND-275
NORDEND-277 23/10 14:45 → 24/10 07:30 Thermal compensated thermal: the label itself said "the compensation absorbs it" ranges.py, level of the 5 h before, --outside the label: nothing beyond 5 noises. The night of the 24th deviates from the rainy afternoon only by the base step (NORDEND-276)

Under the threshold, or not tied to a label

  • The left-edge night rejections (PR09, CTR01, PR04, PR07) are now unlabelled. Over the range, PR09 is rejected at 34 % of the instants, CTR01 at 44 %, PR04 at 22 % and PR07 at 14 % (data.load). CTR01 and PR04 already have spanning Target issue labels (NORDEND-195, NORDEND-152). PR09 is rejected every night of the range, but I did not create a label for it (no new events): see section 3.
  • On 13/10, 16:15-17:00 (dusk, inside the thermal label 181), several targets deviate together for one or two measurements, by 7-10 noises (PR14, PR16, PR18, PR06, PR15). This is beyond the threshold, but it lay inside the removed label 183, whose bounds and text did not describe it. It is a candidate for a new short label in a future analysis.
  • Similarly, PR07's own accepted points jump before or between its rejections (20/10 20:30, +11.7 noises; 18/10 22:45, +9.3). These are single points of the noisiest target, not described by the removed labels.
  • The items of the previous section 2 stay valid: the error_value spikes without effect (21/10 17:45, 22/10 07:00, 23/10 08:30, 24/10 09:30); the corr of the references move by a few thousandths on 24/10 11:15.
  • The day noise at NORDEND reaches 5-7 noises without any cause (PR13 corr_y −7.4 at 14/10 12:15). The short labels kept here all exceed 8 noises, except 186 (7.7).

Requalification

Commands (from the root of aimsight-app):

  • python3 labeling/auto/analyses/NORDEND_2026-10-13T0745_2026-10-24T1630/scripts/requalify.py <input> work/requalify.txt. For each label, it runs tools/scripts/ranges.py data/NORDEND --columns corr_x,corr_y --k 5 with:
    • movement labels up to 4 h: --start S-3h --end E+3h --outside S/E;
    • longer ones: --start S --end E --rolling 8;
    • thermal labels: --start S-5h --end E --outside S/E. The script keeps the target ranges (CTR02-04 excluded) inside the label.
  • scripts/thermal.py <input>: for each thermal target, night level = median of the 5 h before the start; noise = own noise over the label; deviation smoothed by 3; instants of the non-thermal labels skipped; PR04, PR17 and CTR01 left out (labels of their own) (work/thermal.txt).
  • ranges.py --k 5 --rolling 8 on the left-edge nights without the overlapping movement labels (work/req_missing_*.txt).
  • Base step: ranges.py --start 2026-10-22T23:00Z --end 2026-10-24T07:30Z --outside 2026-10-23T05:30Z/2026-10-24T07:30Z --min-measurements 6.
  • overview.py --residuals corr_x,corr_y on 21-22, 22-23 and 23-24/10 (work/req_10-*/).
  • summary.py of error_value and corr_y around 186 and 188.
  • inventory.py --start --end: the six No data gaps are instants missing for every sensor.
  • thumbnails.py --zone 1600,750,500,650 at 14/10 14:30, 15:30, 16:15.

Values: the largest deviation of each kept label is in the table of section 1; those of the removed labels are in the table of section 2. Kept 31, rewritten 4 (186, 208, 210, 214), removed 9 (183, 188, 191, 194, 199, 201, 207, 274, 277). validate.py: 35 labels, 0 errors.

3. The documentation

  • labeling-rules.md section 1 and families.md ("A gap", "What is not labeled") decided the removals. A left-edge target that is not a reference of the compensation cannot move the others, so its night rejections are missing points only.
  • The rule for thermal labels ("5 noises from the night level") does not say which noise. The night noise (ranges.py --outside) is two to four times smaller than the noise over the whole day. With it, every day of the range exceeds the threshold, even the ones the user validated as "absorbed". I used the noise over the label (day and night) and a deviation smoothed over 3 measurements. Even so, 208, 210 and 214 exceed the threshold (6-9 noises), only 277 does not. The rule could name its noise.
  • PR09 is rejected every night of the range, whatever the weather. That fits the Target issue boundary of families.md ("rejected every night ... under the station's ordinary regime"). The old labels called it Visibility ("poorer view at night"). A Target issue for PR09 over the range, like CTR01's NORDEND-195, is worth deciding with the user. I did not add it: the request says not to look for new events.
  • The removal of 183 leaves a dusk jump of many targets (13/10 16:15-17:00) without a label (section 2).
  • The precedents were not queried again: this is a requalification, and the precedents of the kept labels are in work/analysis.before-requalify.md.

Difficulties

  • The request forbids moving bounds. For 186, the effect lasts one measurement past the end (16:15).
  • While I was working, labeling/labeling-rules.md and instructions/analyst.md changed on disk (uncommitted edits: the new section 1, the weather and the mm). The content I applied is the current one.

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, work/spanning.jsonl, the previous analysis.md and report.md, the Target issue labels and the labels around 24/10 in labeling/auto/NORDEND.jsonl. Tools: ranges.py, overview.py, summary.py, inventory.py, thumbnails.py; scripts scripts/requalify.py, scripts/thermal.py, scripts/build_requalified.py. Images: whole image 14/10 15:30, zone thumbnails 14/10 14:30, 15:30, 16:15.