Files
aimsight-labeling/auto/analyses/NORDEND_2026-10-13T0745_2026-10-24T1630/analysis.md
T

16 KiB
Raw Blame History

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

Request: verify the 33 existing labels of the range (work/existing.jsonl), keep, correct, remove, add. DETECTORS: no.

Weather: data/NORDEND/weather/ does not exist (no MeteoSwiss data for this station): no weather; the weather is read on the images only.

Image quality metrics (peak to snr): empty on this camera (inventory.py), so quality.py does not apply; spots read on IRON thumbnails.

Verdict per existing label

Old id Verdict Evidence
NORDEND-181 kept thermal cycle of the 13th; pose and corr as described (days.py --bounds)
NORDEND-182 kept 10:00-10:30: CTR04 error_value up (0.27 at 10:00), PR04/PR07 rejected, PR14-PR16 jump (ranges.py --instants)
NORDEND-183 kept PR09 rejected 17:15 → 07:00, PR04/PR07 from 16:15; targets touched from 16:30
NORDEND-184 kept 05:15-06:00 CTR02, CTR03 rejected, no pose, targets jump by tens of noises; back at 06:15
NORDEND-185 kept cycle of the 14th, tilt first exceedance 07:15
NORDEND-186 kept CTR01 rejected 15:30-15:45; CTR03 error_value also up 15:00-16:00 (see section 2)
NORDEND-187 kept cycle of the 15th, pan/tilt from 07:45
NORDEND-188 kept CTR01 rejected 15:45
NORDEND-189 kept 10 steps missing 22:45 → 01:00, images missing too
NORDEND-190 kept cycle of the 16th
NORDEND-191 kept CTR01 rejected 08:15
NORDEND-192 kept 13:30 missing; 13:45 and 14:00 present, so two labels with 193
NORDEND-193 kept 14:15 missing
NORDEND-194 kept PR07 from 19:15, PR09 20:30 → 09:45
NORDEND-196 corrected (bounds, description) starts 19:00 (targets already shift at 19:15 as CTR04's error_value rises), ends 23:00 (common jumps continue 21:30-22:45 while CTR04's error_value stays raised after its return)
NORDEND-197 kept weak cycle of the 17th
NORDEND-198 kept 13:00-15:30 rolling deviations PR12-PR16
NORDEND-199 kept PR09 rejected 13:00 → 02:15
NORDEND-200 kept weak cycle of the 18th
NORDEND-201 kept PR09 18:00 → 10:45, PR04/PR07 evening; PR01 rejected four times too (not named, minor)
NORDEND-202 kept cycle of the 19th
NORDEND-203 kept 09:00 missing
NORDEND-204 kept 09:45 missing (09:15 and 09:30 present: two labels)
NORDEND-205 kept 20:45 missing
NORDEND-206 kept cycle of the 20th
NORDEND-207 kept PR07 touched 20:30, PR09 22:15 → 05:15
NORDEND-208 kept day pose below 5 noises but night cooling beyond (pan +13 noises at 01:15 on the 22nd): the label stands
NORDEND-209 corrected (split in two, description) two episodes separated by a calm from 12:15 to 14:15; "the other targets do not move" is false: PR15 corr_x −0.07 at 10:45-11:15, PR16 +0.076, PR14 +0.06; PR15 +0.25 corr_x at 16:45
NORDEND-210 kept pan on the morning of the 23rd still lower than the day before (0.016 vs 0.027)
NORDEND-211 corrected (split in two, description) two episodes (09:15-11:00, 12:00-14:00); "the other targets do not move" is false: PR05 corr_x −0.20 / corr_y +0.33 at 09:30, PR14 +0.12 at 10:00, PR07 −0.086 at 12:30, PR16 +0.055 at 12:15
NORDEND-212 corrected (description) bounds right (PR09 back at 08:00), but 06:00-07:30 the whole scene is hit by the dawn rain (new label): PR01, PR02, PR03, PR11, PR12, PR13 rejected; "the other targets remain measured" is false at that hour
NORDEND-213 kept rain day and night; the base step inside it is labelled apart (new label)
NORDEND-214 kept weak warming of the 24th; the two one-measurement jumps of the afternoon are labelled apart

Spanning labels (not changed): NORDEND-152 (PR04), NORDEND-169 (PR17), NORDEND-195 (CTR01) hold within the range: PR04 rejected scattered by day and night; PR17's curve changes offset 21-24; CTR01 rejected only on sunny afternoons and left-edge nights before 16T19:00, much of the time after.

1. My labels

Only the corrected and added labels are detailed; the kept ones are in the table above.

2026-10-13T12:15Z → 2026-10-13T12:45Z (added)

  • Phenomenon of the catalogue: Visibility drops over the whole scene (a brief drop of the light)
  • Precedents: NORDEND-255 (PR07 rejected one measurement, PR15/PR16 jump, weak PR07 spot, Visibility), NORDEND-69 (validated): same family and wording
  • Family: Visibility: the whole group's spots weaken, no object in front
  • Measured: 12:30: PR07 rejected (error_value 0.82); corr_x PR16 −0.073, PR15 −0.062, PR14 −0.051, PR08 −0.039, PR06/PR05 −0.038, PR13 −0.036; refs ±0.003; error_value of PR11, PR14, PR16 up; 12:45 back
  • Images: thumbnails 12:15/12:30/12:45 of PR07, PR14, PR16: at 12:30 spots smaller (PR07 peak 94 vs 246), background brighter; whole image 12:30: overcast, nothing in front
  • Confidence: high for the event, medium for the cause

2026-10-16T19:00Z → 2026-10-16T23:00Z (NORDEND-196 corrected)

  • Phenomenon of the catalogue: A reference loses its brightness, without an obstacle; A reference dropped: the base changes
  • Precedents: the old label itself; NORDEND-173 (CTR04 discarded, all targets jump)
  • Family: Visibility (unchanged)
  • Measured: CTR04 error_value 0.034 → 0.047 (19:15) → 0.084 (19:30), rejected 19:45-20:45; at 19:15 corr_x PR16 +0.071, PR15 +0.045, PR06/PR08 +0.035; after the return (21:00), CTR04 error_value 0.05-0.19 until 22:30 and the targets jump together: PR08 +0.115 (21:45), −0.107 (22:30), +0.048 (22:45); 23:00 back (PR08 −0.005)
  • Images: 20:00 (old label); not redone
  • Confidence: high

2026-10-20T15:15Z → 2026-10-20T17:15Z (added)

  • Phenomenon of the catalogue: A faint spot lost when the light is missing
  • Precedents: NORDEND-69 (PR07 late afternoon, weak spot, Visibility, validated)
  • Family: Visibility; not Target issue (PR07 is not rejected every evening at this hour)
  • Measured: PR07 corr_x +0.104 at 15:30 (error 0.27), rejected 15:45, −0.058 at 16:30, −0.048 at 17:00 (error 0.13-0.28); PR09 corr_y +0.054, PR15 +0.045 at 15:45; refs flat; 17:15 back
  • Images: thumbnails 14:30, 15:30, 15:45, 16:45: the image darkens to the noisy night image; PR07 spot faint, nothing in front
  • Confidence: medium

2026-10-20T23:45Z → 2026-10-21T01:45Z (added)

  • Phenomenon of the catalogue: Part of the scene goes dark at night
  • Precedents: the left-edge night labels of NORDEND (same station, other targets)
  • Family: Visibility: spots weaken without an object
  • Measured: error_value PR14 0.17 → 0.42 (rejected 00:15), PR16 0.20-0.37, PR15 0.08-0.18 (00:00-00:30); corr_y PR15 −0.063, PR16 −0.084 (00:15); corr_x of most targets +0.02 to +0.05 (00:00-00:45); PR15 corr_x −0.064 to −0.090 from 00:45 to 01:30, −0.014 at 01:45; refs flat
  • Images: thumbnails 23:30, 00:15, 00:30, 01:15 (fixed frame): PR14, PR15, PR16 spots much smaller (p99 138 vs 238), PR12 a little
  • Confidence: high for the event, low for the cause

2026-10-21T10:15Z → 12:15Z and 2026-10-21T14:15Z → 17:15Z (NORDEND-209 split)

  • Phenomenon of the catalogue: A faint spot lost when the light is missing
  • Precedents: NORDEND-209 itself
  • Family: Visibility (unchanged)
  • Measured: first: PR07 rejected 10:30-10:45, PR09 10:30-11:15; PR15 corr_x −0.064/−0.073/−0.069 (10:45-11:15), PR16 +0.076 (11:15) and +0.048 (12:00), PR14 +0.062 (10:30); 12:15 back. Second: PR07 +0.045 (14:30), then rejected 14:45-15:15 and 16:30-16:45; PR09 rejected 14:45-16:45; 16:45 PR15 corr_x +0.253 and corr_y +15 noises, PR16 +0.078, PR14 +0.054, all error_values about doubled; 17:00 PR16 +0.036; 17:15 back
  • Images: zone thumbnails 10:00, 10:45, 15:00, 16:45: spots smaller at 10:45, noisy night image at 16:45
  • Confidence: high

2026-10-22T08:30Z → 2026-10-22T09:00Z (added)

  • Phenomenon of the catalogue: none sure (a pass-by hiding targets for one measurement would fit, not seen)
  • Precedents: WYOMING-17 (one-measurement jump, cause not established, Displacement); JUNGFRAU-55
  • Family: Displacement: jump then return without an established cause (families.md boundary)
  • Measured: 08:45: PR02 corr_x +0.079, corr_y +0.147, error 0.145; PR03 corr_x −0.050, corr_y +0.221, error 0.231; others and refs flat; 09:00 back
  • Images: thumbnails 08:30, 08:45, 09:00 (zone, contrast): nothing seen
  • Confidence: high for the event, cause unknown

2026-10-22T09:15Z → 11:00Z and 2026-10-22T12:00Z → 14:00Z (NORDEND-211 split)

  • Phenomenon of the catalogue: A faint spot lost when the light is missing
  • Family: Visibility (unchanged)
  • Measured: 09:30: error_value PR03 0.15, PR05 0.21, PR06 0.20, PR15 0.21; PR05 corr_x −0.203, corr_y +0.328; PR14 corr_x +0.079 (09:30), +0.124 (10:00), +0.094 (10:15), back 10:30; PR09 rejected 09:30-10:45. Second: PR07 corr_x +0.044 (12:15), −0.086 (12:30); PR16 +0.055 (12:15), PR15 −0.051 (12:30); PR09 rejected 12:30-13:15 and 13:45
  • Images: thumbnails 08:30, 08:45, 09:30, 10:00 and 11:45, 12:15, 12:30, 14:15: spots smaller at 09:30, much weaker at 12:15-12:30, nothing in front
  • Confidence: high

2026-10-22T20:15Z → 2026-10-23T08:00Z (NORDEND-212, description only)

  • Bounds kept; the description adds the dawn rain of the 23rd (new label) and moves PR01's rejections to it. Measured: PR09 rejected 20:30 → 04:30 and 06:00 → 07:45.

2026-10-23T05:45Z → 2026-10-23T07:45Z (added)

  • Phenomenon of the catalogue: Rain
  • Precedents: NORDEND-213 (the rain of the same day), NORDEND-184 (dawn mist of the 14th)
  • Family: Visibility: drops on the window seen, whole scene
  • Measured: error_value of all sensors 0.12-0.23 at 06:15; rejected PR01 06:00-07:00, PR02, PR03, PR11, PR12, PR13 06:15-06:45; corr_x PR15 +0.052/+0.133 (06:00/06:15), PR16 +0.065, PR08 −0.082, PR06 −0.061 (06:45); 07:15 most targets −0.02 to −0.04; 07:45 back
  • Images: 06:45 whole image: dark hazy scene, blurred drops in the lower left of the window
  • Confidence: high

2026-10-23T14:15Z → 2026-10-23T14:45Z (added)

  • Phenomenon of the catalogue: The base changes without a reference lost
  • Precedents: LENZSPITZE-105/106/108 (base change, Displacement); JUNGFRAU-117 (targets stay shifted after rain)
  • Family: Displacement: lasting step without an established cause, all references kept
  • Measured: corr_x CTR02 −0.002 (14:15) → −0.040 (14:30) → −0.028 (14:45), −0.028 (16:45), −0.030 (17:45 on); CTR03 +0.004 → +0.035 → +0.027, +0.030. Night medians before (22nd and 23rd, 02:00-05:00) and after (24th, 04:00-07:00): corr_x CTR02 0.000 → −0.019, CTR03 0.021 → 0.043, PR02 0.013 → 0.045, PR03 0.028 → 0.052, PR12 0.062 → 0.105; corr_y CTR02 0.091 → 0.109, CTR04 −0.014 → −0.022, PR06 0.091 → 0.050, PR08 −0.015 → −0.049; hourly medians stable through the 24th until 16:30 (no return). Raw x of the references move with the targets (no raw jump of the targets)
  • Images: none useful (rain)
  • Confidence: medium (the step is clear; its cause is unknown, and the levels change again on the 25th-26th, outside the range)

2026-10-23T14:45Z → 2026-10-24T07:30Z (added)

  • Phenomenon of the catalogue: The camera moves through a thermal effect (cooling)
  • Precedents: WYOMING-7 (evening cooling with rain, thermal), NORDEND-210 (camera not coming back at night)
  • Family: Thermal displacement: the pose follows the housing temperature and comes back with it on the 24th
  • Measured: housing temperature 28.2 °C (14:00) → 25.0 (21:00) → 23.5-24.5 (00:00-07:00), back to 28.6 at 12:00 on the 24th; pan (hourly) 0.010-0.018 by day on the 23rd → 0.022-0.027 at night → 0.011-0.013 at 11:00-13:00 on the 24th; tilt −0.124 → −0.140 → −0.126; corr flat apart from the rain and the base step
  • Images: none
  • Confidence: medium (pose partly hidden by the two-reference instants of 15:00-17:30)

2026-10-24T14:00Z → 2026-10-24T14:30Z (added)

  • Phenomenon of the catalogue: A step of the camera that the compensation does not erase (one measurement)
  • Family: Displacement: jump and return, cause not established
  • Measured: 14:15: raw y of CTR04 −0.5 px, CTR02 and CTR03 −0.4, PR16 −0.5 (all together); tilt −0.132 vs −0.125/−0.127; error_value of all sensors 0.11-0.14 (vs 0.03-0.06); corr_y PR06 −0.036, PR14 −0.029, PR18 −0.028, PR15 −0.027, PR13 −0.022; corr_x PR16 +0.026; 14:30 back
  • Images: zone and reference thumbnails 14:00, 14:15: nothing in front, overcast
  • Confidence: medium

2026-10-24T15:00Z → 2026-10-24T15:30Z (added)

  • Phenomenon of the catalogue: The base changes without a reference lost (one measurement)
  • Family: Displacement
  • Measured: 15:15: corr_y PR16 +0.068, PR15 +0.057, PR06 +0.054, PR13 +0.053, PR12 +0.048, PR18 +0.046; corr_x PR06 −0.038, PR11 −0.033, PR16 −0.033; corr_x CTR02 −0.012, CTR03 +0.010; CTR04 error_value 0.077 (vs 0.05), others at level; raw y of the targets unchanged; 15:30 back
  • Images: thumbnails 15:00, 15:15: darker background at 15:15, nothing in front of the references
  • Confidence: medium

2. What I left without a label

  • 2026-10-14 15:00-16:00: CTR03 error_value up to 0.21 (sun on the rock), the references' corr_y move apart a little (±0.027), targets' corr_y +0.02 to +0.05 during the afternoon thermal return: within the time of NORDEND-186 and the thermal label; not split.
  • 2026-10-20 18:15: PR07 corr_x −0.054 for one measurement, error_value at its level: PR07 is the noisiest target (faint spot); left as noise.
  • 2026-10-24 11:15: the corr of the three references move by a few thousandths (7-8 noises of their own tiny noise), targets flat.
  • Error_value spikes on all sensors without effect on corr: 2026-10-21 17:45, 2026-10-22 07:00, 2026-10-23 08:30, 2026-10-24 09:30.
  • Isolated single-sensor deviations of about 6 noises (PR07, PR09, PR13, CTR04), listed by scripts/uncovered.py 6.

3. The documentation

  • Helpful: labeling-rules.md "one label, a single event" and "a range ends at the measurement after the last touched", which decided the splits of 209 and 211 and the end of 196; families.md "The base changes without a reference lost"; the precedents NORDEND-69/255 for PR07 in dim light.
  • The old labels 209 and 211 said "the other targets and the references do not move": contradicted by the residuals (PR05, PR14, PR15, PR16 jump by tens of noises). The earlier reading likely followed the rejections, not the corr of the measured targets.
  • The base step of the 23rd is hidden under the rain label and only shows when the night levels before and after are compared (scripts/levels.py): the guide could suggest comparing night levels on each side of a long Visibility episode.
  • No weather folder for NORDEND, and no image quality metrics: the causes rest on the IRON thumbnails.

Difficulties

  • The thermal labels cover the whole range, so a coverage check of the jumps has to ignore them (scripts/uncovered.py).
  • The Write tool refused .md files in this session; analysis.md and report.md were written through Bash.

Sources read

agents/agents/analyst.md, instructions/analyst.md, tools/GUIDE.md, tools/DATA.md, labeling/families.md, labeling/labeling-rules.md, labeling/reading-rules.md, labeling/stations.md, work/existing.jsonl, work/spanning.jsonl, precedents.py (queries cited above). Tools: inventory.py, overview.py, ranges.py --rolling 8 --instants 3, days.py --bounds, metrics.py, summary.py, plot.py, thumbnails.py; scripts scripts/tab.py, scripts/levels.py, scripts/uncovered.py, scripts/build.py. Images: whole images 2026-10-13 12:30 and 2026-10-23 06:45; thumbnail sheets in work/.