Files
aimsight-labeling/auto/analyses/MICHIGAN_2026-10-12T1600_2026-10-23T2000/analysis.md
T

12 KiB
Raw Blame History

Labels: MICHIGAN, 2026-10-12T16:00Z → 2026-10-23T20:00Z (UTC)

Request: verify the 4 existing labels of the range (work/existing.jsonl): keep, correct, remove, add. DETECTORS: no.

Verdict per old id

Old id Verdict Evidence
MICHIGAN-5 corrected (end 16T16 → 16T12, description) The corr_x descent is real (median of the targets −0.17 at 15T16, all 14 targets below their level), but (1) the label says "the camera stops nodding forward and straightens a little": the tilt goes down on 13–14 Oct (0.0126 → 0.0097) then rises on 15–16 Oct to 0.0197 at 16T12, its highest of the range so far; the axis that follows the corr_x is the pan (r = −0.96 on 12T16–15T20), not the tilt; (2) it ends at 16T16 and says "then goes up again", while the return is MICHIGAN-7's (from 16T12): the trough is flat from 15T16 to 16T12 (−0.167, −0.157, gap, −0.132), the cut goes in the flat and the two phases touch (labeling-rules §2, §3). The expert sentences are kept.
MICHIGAN-6 kept No measurement 16T00, 16T04, 16T08 (inventory), no image between 15T20 and 16T12, no metric 15T23 → 16T10; BusVoltage 14.08–14.39 V and battery at 100 % just before and after. stations.py: the gap starts at 23:00 on the nine stations measuring then (BREITHORN, CLARIDEN, FLORIDA, JUNGFRAU, MICHIGAN, NEBRASKA, NIEDERHORN, NORDEND, SNCF01); LENZSPITZE starts at 02:10. Everything the label says that I could check holds.
MICHIGAN-8 kept 17T16 corr_x −0.089 (all targets below), 17T20 corr_y −0.160 (−3.9 noises, all 14 targets beyond 3 noises on the rolling level), 18T04 corr_x +0.118 and corr_y +0.131 (all 14 targets beyond 2 noises), back at 18T08 (corr_y +0.007; corr_x +0.085, in the band of the following days). Gap REF42−REFG43N: y −0.04 px at 17T20, x +0.05 px at 18T04; tilt 0.0228 then 0.0084, rotation swings. Temperature 31.25–32.5 °C without a brief drop at these hours. Thumbnails 17T12, 17T20, 18T04, 18T16: spots of the references and targets intact. NEBRASKA and BREITHORN do not move.
MICHIGAN-9 kept corr_y −0.201 at 19T04, −0.225 at 19T08 (−5.5 noises, the only day above 5 noises), corr_x −0.039 at 19T08, back at 19T12. Tilt 0.0278 then 0.0337 (max of the range), rotation −0.026, scale 0.99983 (min); gap REF42−REFG43N −0.07 then −0.05 px in y. Housing temperature: 31.0 at 19T04, 32.0 at 05:00, 29.5 from 07:00 to 09:00 (lowest of the range), 30.25 at 11–12:00: the first affected measurement precedes the drop, as the label says. The temperatures of NEBRASKA and BREITHORN drop already at 19T04. Thumbnails 19T04, 19T08: nothing in front. Confidence on the thermal cause medium: the same signature (nod forward, corr_y down) occurs at 17T20 without a drop (MICHIGAN-8).

Spanning labels (not changed):

  • MICHIGAN-3 (REFPOU22 lost): holds within the range: REFPOU22 rejected at all 65 instants (error_value at its level near 0.99), the compensation on REF42 and REFG43N only (2 retained at every instant).
  • MICHIGAN-7: wrong within the range on one point: it says "from 16 October … the camera nods forward again"; from 16T12 to 17T12 the tilt goes down (0.0197 → 0.0095, the camera straightens), it rises only from 17T16 (0.0172, then 0.020–0.034). The corr_x rise of 16T12 → 17T08 (−0.132 → +0.012) is the return of the descent of MICHIGAN-5 and follows the pan turning back (−0.0122 → −0.0169), not the tilt. The rest (step of corr_x on 17–18 Oct, slow descent of corr_y, faster tilt from the 17th with the housing temperature dropping) holds.

1. My labels

2026-10-12T16:00Z → 2026-10-16T12:00Z (MICHIGAN-5, corrected)

  • Phenomenon of the catalogue: "A compensation on two references" (family of what disturbs them), and "The camera tilts little by little" for the context of the month.
  • Precedents: MICHIGAN-4, -7 (the neighbouring pieces of the same month, validated), MICHIGAN-12 (validated: "it is the two references that move slightly away from the targets, and the compensation on them alone carries this gap over"), which I followed to state the mechanism.
  • Family: Displacement: a drift of several days that the housing temperature does not explain alone (r(corr_x, T) = −0.36 on 12T16–15T20), not Thermal displacement; consistent with MICHIGAN-4 and -7.
  • Measured: median of the targets of corr_x minus their level (scripts/table.py, work/table.txt): 12T16 −0.001, 13T16 −0.139, 14T00 −0.030, 14T16 −0.129, 15T00 −0.033, 15T16 −0.167, 15T20 −0.157, 16T12 −0.132, 16T16 −0.071; lower at 12:00–20:00 every day, higher at 00:00. corr_y: +0.03 to +0.09 until 15T08, −0.073 at 15T12, −0.023, −0.035, −0.098 at 16T12, −0.011 at 16T16. Gap REF42−REFG43N in raw x: +0.03 px on 12–14 Oct, 0.00 from 15T08 (r(corr_x, gap x) = +0.85). Pose: pan −0.0174 → −0.0089 at 15T16 (r(corr_x, pan) = −0.96), tilt 0.0126 → 0.0097 (13T08) → 0.0176 (15T12) → 0.0197 (16T12). overview.py --rigid base 12T16: targets' residual median ≤ 0.015 px, max ≤ 0.037 px; residuals of the references ≤ 0.06 px. Housing temperature 31.75–34.5 °C, peaks in the afternoons (13T12 34.5, 14T16 34.0, 15T16 34.5). Days: corr_x never above 5 noises of the night level (−2.0 at most, days.py --bounds). Neighbours: the tilts of NEBRASKA (0.0275 → 0.0239) and BREITHORN (0.0200 → 0.0172) go down over 13–16 Oct; MICHIGAN's rises on 15–16.
  • Images: thumbnails of REF42, REFG43N, COLH18S''', POUH19N', COLH21 at 12T16, 13T16, 14T00, 15T16, 15T20, 16T12, 16T16 (--fixed): spots intact, nothing in front.
  • Confidence: high on the curves and the mechanism; the cause of the change of the gap between the two references is not established.

2026-10-15T20:00Z → 2026-10-16T12:00Z (MICHIGAN-6, kept)

  • Phenomenon of the catalogue: "A gap common to several stations".
  • Precedents: none needed beyond the label itself.
  • Family: No data: no measurement for any sensor.
  • Measured: see the verdict table.
  • Images: none in the gap (no file between 15T20 and 16T12).
  • Confidence: high.

2026-10-17T12:00Z → 2026-10-18T08:00Z (MICHIGAN-8, kept)

  • Phenomenon of the catalogue: "A compensation on two references" / "The base changes without a reference lost".
  • Precedents: MICHIGAN-12 (same signature on 27 Oct, validated); JUNGFRAU-110, FLORIDA-72 (one-measurement common jumps through the base, Displacement).
  • Family: Displacement: no temperature drop at these hours, no visible cause.
  • Measured: see the verdict table; rolling level ±6 measurements, ranges.py --instants: 17T20 all 14 targets corr_y −4 to −5 noises; 18T04 all targets corr_x +2 to +3, corr_y +2 to +4.
  • Images: thumbnails 17T12, 17T20, 18T04, 18T16; whole image 17T20: indoor scene, dark, all spots present.
  • Confidence: high on the bounds; cause not established.

2026-10-19T00:00Z → 2026-10-19T12:00Z (MICHIGAN-9, kept)

  • Phenomenon of the catalogue: "The camera drifts away when the temperature of its housing drops, without a daily cycle".
  • Precedents: MICHIGAN-2, -10, -11, -13 (validated, same signature with a brief drop of the housing temperature).
  • Family: Thermal displacement, as the validated precedents of this station; see the reservation above.
  • Measured: see the verdict table.
  • Images: thumbnails 19T04, 19T08: nothing in front.
  • Confidence: medium on the cause (the plunge starts before the drop recorded at the camera; the neighbours cool at 19T04), high on the bounds.

2. What I left without a label

The corrected curves of all the targets are agitated together all through the range, through the compensation on two references (MICHIGAN-3): the raw positions are given to the hundredth of a pixel and a change of one or a few hundredths of the gap between REF42 and REFG43N becomes about a tenth of a pixel on all the targets (r(corr_y, gap y) = +0.8, r(corr_x, gap x) = +0.8 on each part of the range). Left without a label, because they stay under the threshold that the labelled moments pass (median of the targets under 3.5 noises of its level; days.py: never above 5 noises outside 19 Oct; ranges.py --rolling 6 --k 3 --instants 6: only 17T20, 19T04, 19T08 in the range):

  • the corr_x of all the targets higher at the 00:00 measurement from 19 to 23 Oct (+0.12 to +0.15, the other hours +0.02 to +0.08), following the pan; the lowest points 20T04 (−0.043, all targets −2 to −4 noises against the rolling level, pan jumping to −0.0134), 21T08 (−0.033, −2 to −3 noises), 22T16, 23T20. 20T04 is the closest to a label: about as large as the 18T04 part of MICHIGAN-8, but alone in x, and as deep as 21T08 against the median level of these days (−2.2 and −2.0 noises): I treat both as the troughs of this agitation, not as events;
  • small dips of the corr_y of all the targets with a nod of the camera (tilt up, rotation down), about half of those of MICHIGAN-8 and -9: 16T12 (−0.098, at the resumption), 17T04 (−0.082), 19T16 (−0.101), 21T00 (−0.083), 21T08 (−0.073), 23T08 to 23T20 (−0.094, −0.082, −0.047, −0.078, tilt 0.019 → 0.028 at 23T08 and staying there): the latter precede MICHIGAN-10 (24 Oct, outside the range), whose camera nod starts at 23T08 already;
  • the daily wave of the corr_x on 13–15 Oct, described in MICHIGAN-5 (under 5 noises, no separate thermal label: labeling-rules §1).

3. The documentation

  • families.md ("A compensation on two references") and reading-rules.md (§5: neighbouring cameras read together; at least three references) framed everything; stations.md (MICHIGAN: two references since 09.10, a measurement every four hours, indoors) too.
  • What was missing: nothing tells how big a common deviation must be to be labelled on a station compensated on two references, where all the targets always move together; the "5 noises" rule exists only for the thermal cycle. I took the moments already validated as the threshold (about 4 noises of the median of the targets).
  • What misled: the validated labels MICHIGAN-5 and -7 tell the camera's tilt from the narrative of the three cameras of the building ("stop and slight backward movement from 13 to 16, resumption from the 17th"); on MICHIGAN itself the tilt rises on 15–16 Oct and goes down on 16–17 Oct, and the corr_x follows the pan. A description of a camera axis is to be read on the station's own pose (labeling-rules §5, "checked axis by axis").
  • The raw positions are rounded to the hundredth of a pixel: the gap between the two references moves by steps of that size, which overview.py --rigid shows only as residuals of a few hundredths; computing the gap REF42 − REFG43N directly (scripts/table.py) made the mechanism plain.
  • No weather: data/MICHIGAN/weather/ does not exist (station without a Swiss position, indoors).

Difficulties

  • plot.py --zone takes one range per flag (--zone A/B --zone C/D), not several after one flag.

Sources read

  • agents/agents/analyst.md, instructions/analyst.md, work/prompt.txt, work/existing.jsonl, work/spanning.jsonl
  • tools/GUIDE.md, tools/DATA.md, labeling/families.md, labeling/labeling-rules.md, labeling/reading-rules.md, labeling/stations.md
  • labeling/auto/MICHIGAN.jsonl (MICHIGAN-2, -4, -10, -11, -12, -13 for the edges and the precedents)
  • precedents: precedents.py (JUNGFRAU-110, -112, FLORIDA-72, NORDEND-80, WYOMING-17, NEBRASKA-10, MICHIGAN-2, -11, -13, -15)
  • tools: inventory.py, overview.py --rigid (work/overview.txt), metrics.py (work/metrics.txt), ranges.py --rolling (work/ranges_rolling.txt), days.py --bounds (work/days.txt), stations.py (all stations at the gap; MICHIGAN, NEBRASKA, BREITHORN over the range), thumbnails.py, plot.py; whole image 2026-10-17T20:00Z
  • my script scripts/table.py (per instant: medians of the targets, gap REF42 − REFG43N, pose, hourly temperature; work/table.txt)