# Labels: CLARIDEN, 2026-10-01T00:00Z → 2026-10-11T00:00Z (UTC) Task: check the 12 existing labels of the range (`work/existing.jsonl`, CLARIDEN-2 to CLARIDEN-13, all reviewed on 2026-11-17) and the 2 spanning labels (`work/spanning.jsonl`: CLARIDEN-1 and CLARIDEN-14), then look for what they miss. Result: 16 labels in `labels.jsonl` (11 kept as they are, 1 corrected, 4 added, 0 removed). Weather: there is no `data/CLARIDEN/weather/` folder (and no place on the CLARIDEN line of `stations.txt`), so no weather is available. The station is in a tunnel, so the weather counts little here. Station: 10 min cadence, at minutes :00, :10 … :50. There are 5 references (D-302-C, PR-291, PR-293, PR-295, PR-296), all retained at every instant except 2026-10-05T12:20Z. There are no image-quality metrics (export empty), so `quality.py` was not usable. Images are hourly, plus 10:30 on 5 Oct. ## Verdict per old id | id | verdict | evidence | |---|---|---| | CLARIDEN-2 | kept | `dump.py`: the corrected curves are first touched at 06:30 (D-302-B/PR_BG y, D-301-C x, P-299-B y) and last touched at 08:40 (D-301-C x, D-302-B y); from 08:50 only the error_value stays raised. The label's 06:15/08:45 bounds are 5 min off the measurement grid but cover exactly the same touched instants. The camera scale dips during the episode while the references' error_value is raised (D-302-C, PR-295), which matches "references moving away from the targets". | | CLARIDEN-3 | kept | D-303-B, D-304-B and D-306-B go down vertically from 17:50 (smoothed: D-303-B and D-306-B from about level to their lowest by 03:00–03:40 on the 2nd). The end at 02:30 is within one noise of the real bottom. No tilt step. The return of D-303-B that the description mentions is now labeled separately (added A). | | CLARIDEN-4 | **corrected** (start 21:45 → 21:30) | At 21:40 the episode is already there: P-301-B corr_x moves clearly, D-301-C corr_x and D-302_A corr_y jump, and the error of most targets rises (D-304-A, D-304-B). 21:40 was outside the label, so the bound is now the measurement before, 21:30. The end (23:00) and the description hold. | | CLARIDEN-5 | kept | Housing temperature: hourly 40.5 °C → 38.3 °C (21:30–22:20) → back by 00:00. The scale goes from about −5e-5 to −14.7e-5 at 22:50 and is back at the post-episode level (about −7e-5) at 23:20. The corrected curves are back at 23:00. | | CLARIDEN-6 | kept | 07:50 on 3 Oct is missing for all sensors. FLORIDA, JUNGFRAU, NORDEND and WILDSTRUBEL measure at 07:45 and 08:00, so the gap is specific to the station. There is no image instant at 07:50 (images are hourly). The voltage in that hour is 14.20 V, a little below the usual 14.33 V but far above the knee. | | CLARIDEN-7 | kept | Last measurement 17:40: D-301-C x, D-302-B/PR_BG x and y, and D-304-B y are touched, and the error of 12 sensors rises. The end bound (17:45) is off-grid; there is no measurement after it. | | CLARIDEN-8 | kept | There are no measurements from 17:50 on the 3rd to 10:20 on the 5th (244 steps), and no metrics from 18:00 on the 3rd to 09:00 on the 5th. The last image is 17:00 on the 3rd, the next 10:30 on the 5th. Voltage is 14.35 V before the gap and 14.20 V after it, with the battery at 100. The other stations have data on 4 Oct. | | CLARIDEN-9 | kept | The levels change across the gap (D-303-C, D-304-C, D-300-C, D-302_A, D-301-A, P-299-B, D-302-B…; D-306-B returns to its level from before the 1st). The scale moves from about −7e-5 to −16e-5. | | CLARIDEN-10 | kept | At 10:30 the tilt is −0.109 (against +0.005 afterwards), the whole scene is shifted, D-304-C is rejected and many corrected curves jump. 10:40 is mostly back; D-303-B y and D-302-B/PR_BG x are still touched until 11:10. The 10:30 image shows the blurred object on the right and the worker at the top left. | | CLARIDEN-11 | kept | 12:20: row C and D-302-C are rejected; D-306-B y and x, D-303-B, D-304-B, P-299-B, P-301-B and D-302-B jump; 12:30 is back. PR-296 and PR-295 also have their error raised at 12:20 (not in the text, not wrong). No image at 12:20, because the camera takes hourly images; the 12:00 image is clean. | | CLARIDEN-12 | kept | First touched 18:10 (D-301-C x, D-302-B/PR_BG x, P-301-B x), last touched 22:00 (D-301-C x), 22:10 back. | | CLARIDEN-13 | kept | Hourly temperature drops from 40.5 °C to 38.0 °C (18:00–21:00) and is back at 40 °C by 23:00. The scale goes from about −16e-5 to −23e-5 between 19:20 and 22:20 and comes back after 22:40 (−18e-5, then −15e-5 at 23:10). | Spanning labels (not in `labels.jsonl`, not changed): - **CLARIDEN-1** (P-300-B Target issue): holds within the range. P-300-B is rejected at all 1196 instants of the range. - **CLARIDEN-14** (Displacement, 6 Oct 22:00 → 11 Oct 06:00): the bounds and the family hold, but **its description is wrong within the range**. It says "No camera movement". In fact the camera tilt decreases steadily from about +0.007° (7 Oct) to about −0.0025° (10 Oct), with a step at 09:00 on 8 Oct. D-303-B's vertical rise follows this tilt (see added labels C and D below). ## 1. My labels ### 2026-10-02T03:40Z → 2026-10-03T03:00Z (added A) - **Phenomenon of the catalogue**: "Targets move slowly, alone, then come back" (here the return phase of CLARIDEN-3). - **Precedents**: CLARIDEN-29 and CLARIDEN-30 (row B goes down or back up, Displacement, labeled separately from its outward movement). labeling-rules §3: "the outward and return of a drift are two labels". CLARIDEN-3 itself already notes the return in words. - **Family**: Displacement. The cause is not established. The rise does not follow the temperature (flat at 40.8–41.0 °C) and only partly the tilt. - **Measured**: D-303-B corr_y, smoothed (9 measurements): −0.078 at its bottom (03:00–03:40), −0.046 by 08:00, a plateau near −0.04 until 18:00, −0.02 by 22:00, and −0.005 from 02:40 on the 3rd, which is its level of 1 Oct (about 0). D-306-B and D-304-B stay low (−0.10 and −0.06) until the gap. The tilt goes from 0.0025° to 0.0003°; with D-303-B's gain of −11.7 px/deg (measured 5–11 Oct), that explains about a third of the rise at most. The error_value of D-303-B is ordinary, except during CLARIDEN-4. - **Images**: thumbnails of D-303-B and P-301-B at 2 Oct 12:00 and at 3 Oct 08:00, 12:00 and 16:00 (contrast): same sharp spot, nothing in front. - **Confidence**: medium. The curve is clear, but the cause is open, and where the start falls on a slow turning point is imprecise to within about an hour. ### 2026-10-02T21:30Z → 2026-10-02T23:00Z (CLARIDEN-4, corrected start) - **Phenomenon of the catalogue**: "Visibility drops over the whole scene, without an obstacle" (the station's recurring "view obstruction" episodes). - **Precedents**: CLARIDEN-2, -12 and -18 (same signature). - **Family**: Visibility, unchanged. - **Measured**: see the verdict table. The touched instants run from 21:40 to 22:50. - **Images**: 22:00 (per the original label). - **Confidence**: high. ### 2026-10-03T09:40Z → 2026-10-03T17:00Z (added B) - **Phenomenon of the catalogue**: none exactly. Closest: "Targets move slowly, alone, then come back". - **Precedents**: CLARIDEN-31 (row B and row A targets fidget, error up, Displacement, cause not established). - **Family**: Displacement. It is a single target, not the whole scene, so not Visibility. There is a single episode and P-301-B is otherwise clean (2 rejections over the whole period), so not Target issue. - **Measured**: P-301-B corr_y, raw: −0.03 to −0.05 from 08:30 to 09:40, then −0.013 at 09:50, +0.024 at 10:00, and +0.04 to +0.05 until about 14:30. After that it declines (+0.02 at 15:20, −0.01/+0.01 from 16:40). The noise is 0.010. P-301-B's error_value goes from about 0.035 to 0.055–0.065 from 09:50 (0.056, 0.065 at 10:00) and stays raised (0.05–0.07) until 13:30, then about 0.045 until 16:00. Raw y of P-301-B moves from 1013.43 to 1013.51–1013.54 px between 09:40 and 11:30, while D-303-B, D-304-B and P-299-B stay within 2 noises (dump.py against the base 03:00–07:30) and the camera pose does not change (tilt within ±0.001°, scale −5e-5 to −9e-5 as before). corr_x of P-301-B hardly moves (within 3 noises). Housing temperature flat (40.4–40.6 °C). - **Images**: thumbnails of P-301-B at 08:00, 12:00 and 16:00 on the 3rd (contrast): same sharp spot, nothing in front. No hourly image shows a cause. - **Confidence**: medium. The jump and the plateau are clear (+4 to +9 noises for four hours). The end is soft: corr_y is still +3 to +4 noises at 17:10–17:20 and the error of P-301-B rises again from 17:10 (0.058, 0.068), which runs into CLARIDEN-7 (17:30). An end at 17:30 would have been as defensible. ### 2026-10-06T13:00Z → 2026-10-06T13:30Z (added C) - **Phenomenon of the catalogue**: "A step of the camera that the compensation does not erase" (the camera moves at once and a remainder passes into the corrected curve of one target). - **Precedents**: CLARIDEN-22 (20 Oct 00:15 → 00:45, Displacement: a camera step that the compensation absorbs except on a few targets) and SNCF01-42 (a tilt step kept by the pose, corr_y of some targets keeps a step). Both say how to word it: Displacement, the remainder of the camera movement, not a movement of the structure. - **Family**: Displacement, as in the two precedents. Not Thermal displacement: the housing temperature is flat (40.35–40.54 °C). Not Obstruction: no rejection, no error rise. - **Measured**: tilt +0.004° until 12:50, +0.0053° at 13:00, +0.0065° at 13:10, +0.0069° at 13:30, and it stays near +0.006 to +0.007 on the 7th. Raw y of every sensor rises by about 0.07–0.09 px between the windows 11:00–13:00 and 13:20–15:20, references included (PR-291 954.53 → 954.62, D-302-C 667.09 → 667.19). Raw y of D-303-B rises by only 0.04 px (1020.02 → 1020.06). Its corr_y falls from +0.012 (12:50) to −0.014 (13:00), −0.027 (13:10), −0.059 (13:20): −7 noises between the two windows, and it stays low (smoothed −0.03 to −0.04 until the evening). D-303-C and D-304-C move by −1.6 and −2.5 noises only; all the other targets within 2 noises. Error of D-303-B ordinary (0.041–0.055). - **Images**: none (no hourly image shows anything; the step is seen on the pose). - **Confidence**: high for the step and its cause (the camera tilt); the label covers the step itself, the level kept afterwards is not a new event. ### 2026-10-08T08:50Z → 2026-10-08T09:10Z (added D) - **Phenomenon of the catalogue**: "A step of the camera that the compensation does not erase", as for C, in the opposite direction. - **Precedents**: CLARIDEN-22 and SNCF01-42, as for C. - **Family**: Displacement, as for C. - **Measured**: tilt +0.0039° at 08:50, +0.0020° at 09:00, +0.0008° at 09:10, then +0.0014 to +0.0027 (hourly near +0.0015 to +0.002 for the rest of the day). Raw y of the references falls by about 0.08–0.10 px (PR-291 954.49 → 954.40, D-302-C 667.02 → 666.92, D-302-B 1021.00 → 1020.90). Raw y of D-303-B falls by 0.07 px (1019.98 → 1019.91). Its corr_y goes from +0.017/+0.034 (08:40–08:50) to +0.060 (09:00) and +0.061 (09:10), and stays near +0.05 to +0.06 afterwards. D-303-C and D-304-C within 2 noises. The error of D-303-B rises a notch at the step (0.046 → 0.056–0.059) and stays there; D-304-B's too (+3 noises). The label text says "Ordinary recognition error": this is a small overstatement, not a wrong diagnosis. Housing temperature flat (39.95 °C). - **Images**: none. - **Confidence**: high for the step; medium for the exact share of D-303-B's rise that is due to the step (it sits inside its slow rise of these days). ### The link between D-303-B and the camera tilt (for C, D and CLARIDEN-14) Over 5 Oct 13:00 → 11 Oct 00:00 (hourly medians, 124 hours), D-303-B corr_y follows the camera tilt with r = −0.95. D-303-C corr_y (r = −0.91), D-304-C corr_y (r = −0.86) and D-301-C corr_x (r = −0.92) also follow it closely. These are exactly the targets and the directions that CLARIDEN-14 describes as drifting. Figure: `d303b_tilt.png` (D-303-B against the tilt, reversed, with the two steps in grey). The tilt itself drifts slowly from about +0.007° (7 Oct) to about −0.0025° (10 Oct), besides the two steps. The housing temperature stays at 39.7–40.7 °C, so this drift is not thermal. ## 2. What I left without a label - **5 Oct 23:10**: D-306-B corr_x +6 noises for one measurement, alone (scale −19.6e-5 at the same instant, back at 23:20). A single target, a single point: noise. - **6 Oct 13:30 → 22:00**: D-303-B stays low after the step of C (smoothed −0.03 to −0.04). It is the level kept after the camera step, not a new event. From 18:00 it is inside CLARIDEN-12, from 22:00 inside CLARIDEN-14. - **8 Oct 05:40**: six targets (mostly row A: D-302_A, D-303-A, D-304-A, P-299-B, PR-293, D-306-B) at 4–5 noises for one measurement. Row A is the noisiest row; no error rise, no pose change. Noise. - **10 Oct 02:40**: three row A targets (D-300-A, D-301-A, D-303-A) at 4–5 noises for one measurement. Same reason. - **9–10 Oct**: D-301-C corr_x and D-303-B corr_y above their level for long stretches. They are inside CLARIDEN-14 (the slow drift, which follows the camera tilt) and need no other label. - **Single-step jumps of the references** (PR-291, PR-293 on 9 Oct 13:00; PR-296 on 6 Oct 20:10 inside CLARIDEN-12): at most 5–7 noises for one measurement, the noise of the references; nothing on the targets. ## 3. The documentation - `labeling-rules.md` §3 ("the outward and return of a drift are two labels") decided added A: CLARIDEN-3 named the return of D-303-B in words but its bounds stopped at 02:30 on the 2nd. - `families.md` on thermal movement (always labelled, the camera's included) confirmed keeping CLARIDEN-5 and CLARIDEN-13 next to the Visibility labels of the same evenings. - The precedents CLARIDEN-22 and SNCF01-42 gave the wording of C and D (a camera step, absorbed except on one target: Displacement, the remainder of the camera movement). CLARIDEN-29 and CLARIDEN-31 helped for A and B. - `stations.md` has no weather place for CLARIDEN and there is no `weather/` folder: no weather, as the instructions foresee. The station is in a tunnel. - There are no image-quality metrics for this station (export empty), so the quality reading of `tools/GUIDE.md` could not be used; the error of the references and the pose did the job. - **Contradicted by the data**: CLARIDEN-14 says "No camera movement" and "the camera housing temperature stays almost constant: no link with it". The second part holds; the first does not: the camera tilt drifts over the same days and the drifting targets follow it closely (r −0.86 to −0.95). The label stays (it is outside this range's set and its bounds and family hold), but its description should be corrected the next time its range is verified: the drift is mostly a remainder of the camera tilt that the compensation does not absorb on these targets, not an established movement. In the catalogue it is "The camera tilts little by little" (the compensation corrects some targets better than others), not "Targets move slowly, alone, then come back". ## Difficulties - The first analyst was cut off while writing this file (a filter stopped the reply); the labels were already complete and submitted. A second pass finished sections 1 (B, C, D), 2, 3 and the report from the work notes, the log of the first pass and fresh checks on the data (dump.py on 3 Oct 08:30–17:40, 6 Oct 12:20–14:30 and 8 Oct 08:00–10:00; raw positions of P-301-B and D-303-B). - The request said to read `work/prompt.txt`: there is no such file in the folder. The task was taken from `request.json` (its hints) and from the caller's message. ## Sources read - `instructions/analyst.md`, the analyst trade (`agents/agents/analyst.md`). - `tools/GUIDE.md`, `tools/DATA.md`, `labeling/families.md`, `labeling/labeling-rules.md`, `labeling/reading-rules.md`, `labeling/stations.md` (first pass). - `work/existing.jsonl`, `work/spanning.jsonl`, `labeling/auto/CLARIDEN.jsonl` (CLARIDEN-15, -18, -22, -29, -31). - Precedents (`labeling/scripts/precedents.py`): CLARIDEN-18, CLARIDEN-22, SNCF01-42, NORDEND-198, LENZSPITZE-15/-35, WILDSTRUBEL-83/-102. - `data/CLARIDEN/` through `tools/scripts/` (inventory, metrics, overview, ranges, plot, thumbnails, data.py) and `scripts/dump.py`, `scripts/smooth.py`, `scripts/tilt_fig.py`; displacement files of 2, 3, 5 and 6 Oct (raw positions); images 2026-10-05T10:30Z and 12:00Z. - The other stations' measurements on 3 and 4 Oct (FLORIDA, JUNGFRAU, NORDEND, WILDSTRUBEL), for CLARIDEN-6 and CLARIDEN-8.