# Labels: SNCF01, 2026-10-04T00:00Z → 2026-10-14T06:45Z (UTC) Task: check the 30 labels already on the range (`work/existing.jsonl`) against the rules now in force (`labeling-rules.md`, section 1: a movement is labelled only beyond 5 of the target's own noises; a missing point is never labelled for itself). The 7 labels that cross the range bounds (`work/spanning.jsonl`) are checked inside the range only and are not changed. Result: 28 labels. **14 kept** as they are, **14 corrected**, **2 removed**, **0 added**. Context: - Station: a masonry wall along railway tracks. Three references (BG.3.1, R1, R2). There are no system metrics with values, so no temperature is available. - Targets are 16 to 21 m from the camera. 1 px is about 6 to 8.5 mm (`targets.csv`). CV.1 has no mm factor. - Weather: SNCF01 has no `weather/` folder (a station without a Swiss position), so there is no MeteoSwiss data. Sun or overcast is read on the noon images. There are cast shadows on 04, 05, 06, 08, 10 and 12/10, a bright overcast sky on 07, 09 and 11/10, and sun with clouds on 13/10 (`work/thumbs/…zone560-0-2500-900…png`). - The image quality columns are empty (older camera), so `quality.py` does not apply. Method: - Ran `inventory.py`, `overview.py` (15 min and 1 h, residuals) and `ranges.py --rolling 4 --instants 2`, plus `days.py --bounds` on P7 and on the pose (`--night 0-4 --smoothing 5`). - Wrote three scripts in `scripts/`: - `daily.py`: the deviation of each target per day. - `fast_instants.py`: the instants beyond 5 noises from the rolling ±4 median, with the six passages taken out. - `thermal_bounds.py`: first and last instant beyond 5 noises from the night level. - Ran `summary.py` per instant around each short event, `stations.py`/`inventory.py` on FLORIDA and nine other stations for the gaps, and looked at the images. ## Verdict per old id | Old id | Verdict | Evidence | |---|---|---| | SNCF01-8 | corrected (bounds, description) | Agitation beyond 5 noises from 09:00 (CV.4 y) to 16:15 (CV.2 y): new bounds 08:45→16:30, against 08:15→17:15. Old text copied from one day to the next; rewritten with this day's movers (CV.4, F4.2, BG.2, BG.3, and P jumping together at 11:30). Sunny day seen at 12:00. | | SNCF01-9 | kept | P7 corr_x up to 31 noises, corr_y 21. The pose returns all night and rests at dawn (pan flat 05:00-06:30), so it chains to -10. Text true for the day. | | SNCF01-10 | kept | P7 corr_x up to 36 noises. Pose returns until 06/10 05:00, chained to the next start. | | SNCF01-11 | corrected (end, description) | Last exceedance 15:00 (P2 y), so the end moves from 16:45 to 15:15. The 16:00 train is its own label. Movers rewritten (CV.2, CV.4 up to 10 noises; P1-P3 corr_y together at 10:45-11:00). | | SNCF01-12 | kept | 16:00 train seen on the image. BG.2, BG.3, F4.5, P4 stay under the threshold and jump up to 106 noises (BG.3 y). R1 and BG.3.1 rejected; no pose at that instant. | | SNCF01-13 | **removed** | 06/10 06:45: the CV targets and F4.2 are rejected (missing points). The targets left under the threshold stay below 5 noises: P2 y +3.3, F4.6 y +3.0, F4.5 x +2.8, P1 y +2.6, P3/P4 y +2.5, BG.2 x +2.1. Nothing stands out on the curves. | | SNCF01-14 | corrected (bounds, description) | Exceedances 07:15 (CV.4 x) to 13:00 (F4.5 x): 07:00→13:15, against 07:15→14:30. | | SNCF01-15 | corrected (end, description) | Ended at midnight (forbidden). The curves are back by 22:30. The pan reaches its night level at 02:30-03:00 on 07/10 and stays there (0.0144-0.0150° until 04:15); the tilt is back at 03:00, so the end is 07/10 03:00. The old text had a copy error ("the evening and the evening"). | | SNCF01-16 | corrected (end, description) | Ended at 21:00, but P5 corr_x is still beyond 5 noises from the next night level until 01:00 on 08/10, and P7 corr_y until 23:00. The pose returns until 04:00-04:15 on 08/10 (pan 0.0153°, tilt -0.1765°, then flat until 07:30), so the end is 08/10 04:15. This day is weaker and overcast: pan drops 0.006° against 0.013° on sunny days. Old text was a copy with an error. | | SNCF01-17 | corrected (description) | The old text said the remaining targets "barely move". In fact BG.3 corr_x moves +7.1 noises, and F4.5 and F4.6 corr_y move -6 noises: beyond the threshold, which is why the label holds. The 13:00 train is seen on the image. | | SNCF01-18 | kept | 16:00 train seen. BG.2, BG.3, F4.5, P1 jump (P1 -99 noises in y). | | SNCF01-19 | kept | 19:45: BG.3.1 rejected, R1 left out, and a pose is computed from them that spikes (pan, tilt, rotation, scale). BG.2, BG.3, F4.5, P2-P7 jump 26-75 noises. There is no image at 19:45 (the camera takes one on the hour). | | SNCF01-20 | kept | Exceedances 07:15 to 16:00, so 07:00→16:15 by the rule. The old 06:45→16:30 is one step off at each edge, which is within the tolerance. Sunny. | | SNCF01-21 | kept | P7 corr_x 16.5 noises by day; P1 and P5 overshoot at night as the text says (P5 corr_x -0.13 px against -0.09). | | SNCF01-22 | kept | No measurement of any sensor at 02:00. FLORIDA misses the same instant; none of the other nine stations does. The 02:00 image is missing (gap 01:00→02:15). Metrics empty. | | SNCF01-23 | kept | No evening return (P7 stays at -0.40 px until midnight). Return on the night of 10/10. | | SNCF01-24 | **removed** | 09/10, overcast: the only deviations are CV.4 x +5.1 at 09:15, one isolated measurement at the noise edge, and P5 x +5.8 at 07:15, which is the start of the thermal cycle of -23. CV.2 rejected once is a missing point. No agitation regime. | | SNCF01-25 | kept | SNCF01 alone that night; the 02:00 image exists. | | SNCF01-26 | kept | P7 corr_x 17 noises. Pose returns all night until 11/10 06:00-07:00, chained to the next start. | | SNCF01-27 | corrected (bounds, description) | Exceedances 07:15 to 12:45, so 07:00→13:00, against 07:15→14:15. | | SNCF01-28 | corrected (bounds, description) | Overcast. Seven single-measurement exceedances of 5.0 to 5.8 noises (CV.4 y, F4.2, CV.2, P1/P2 y) from 10:15 to 15:15, so 10:00→15:30, against 08:15→16:00. The P7 corr_x +6.9 at 16:15 is the nightfall step of the thermal cycle. Described as weak. | | SNCF01-29 | corrected (end, description) | Ended at midnight. Curves back by 21:30. Pan and tilt reach the night level they keep at about 02:45 on 12/10 (pan about 0.016°, tilt -0.1738°), so the end is 12/10 02:45. Copy error in the old text. | | SNCF01-30 | corrected (description) | All P targets are rejected at 14:15; only BG.2, BG.3 and F4.5 stay and jump (BG.3 y -91 noises). The old text "and some P" was copied from another passage. | | SNCF01-31 | corrected (description) | CV.8, F4.2 and CV.6 are rejected. CV.4 stays under its threshold, its error drops (0.46 against 0.55) and its corr_x jumps +0.15 px (+6.9 noises, about 1.3 mm), then comes back. The old text said "nothing else moves". | | SNCF01-32 | corrected (bounds, description) | Exceedances 07:45 to 14:30 (06:45 is -31), so 07:30→14:45, against 07:00→16:15. | | SNCF01-33 | kept | Night 12→13 without return: P7 corr_x stays at -0.345 px, 15 noises above the night before. Back by 14/10 04:00 (-0.446). | | SNCF01-34 | corrected (description) | Same as -30: all P rejected at 16:00; BG.2, BG.3, F4.5 jump. 16:00 train seen. | | SNCF01-35 | kept | SNCF01 alone; the 02:00 image is missing. | | SNCF01-36 | corrected (bounds, description) | Exceedances 08:00 to 13:45, so 07:45→14:00, against 08:30→14:30. | | SNCF01-37 | kept | Gap at 01:30, shared with FLORIDA. | ## 1. My labels ### Visibility, daytime shimmer (-8, -11, -14, -20, -27, -28, -32, -36) - **Phenomenon of the catalogue**: Atmospheric turbulence. - **Precedents**: `precedents.py` could not run (Ollama unreachable). I used a grep of `labeling/auto/` instead: NORDEND-262 (early-afternoon fluctuation, cause not established) and WILDSTRUBEL-82/90 (agitation by day, error rising). They show the same shape. Bounds here are read on the exceedances, not from sunrise to mid-afternoon. - **Family**: Visibility. The error of the agitated targets rises by day: CV.4 max 0.54 at night against 0.75-0.87 by day on 05/10, and CV.2 likewise. The agitation is a fast noise that comes back at once, only on sunny days (cast shadows at 12:00), and strongest on the low targets CV.2 and CV.4. This is not a pose agitation without an error rise (that would be thermal). - **Measured**: rolling ±4 deviations in own noises (window noise CV.4 0.022 px ≈ 0.19 mm, CV.2 0.030 px ≈ 0.18 mm): - CV.4 up to 12.4 noises (06/10 11:30, ≈ 2.4 mm). - CV.2 up to 9 noises (≈ 1.7 mm). - F4.2 up to 9.8 noises. - Group jumps of BG.3 and the P targets in corr_x at single instants (04/10 11:30, 10/10 09:45, 12/10 10:30) or in corr_y (05/10 10:45). - Number of exceedance instants per day: 04/10 15, 05/10 31, 06/10 13, 08/10 30, 10/10 24, 11/10 7, 12/10 19, 13/10 8. - Night: none except the thermal nightfall steps of P6/P7. - **Images**: noon zone sheet of the 10 days, plus the whole images of 05/10 16:00, 07/10 13:00, 07/10 16:00 and 12/10 16:00. - **Confidence**: high on 04, 05, 06, 08, 10 and 12/10. Medium on 13/10. Medium-low on 11/10, which only just crosses the threshold. ### Thermal displacement (-9, -10, -15, -16, -21, -23, -26, -29, -33) - **Phenomenon of the catalogue**: The camera moves through a thermal effect. - **Precedents**: the reviewed SNCF01 labels themselves (chained cycles ending at the pose rest). - **Family**: Thermal displacement. Every day, pan and tilt drop and scale rises, then come back; the curves follow and come back. There is no temperature, so the link is deduced. - **Measured**: P7 corr_x daily peak in noises (and mm): - 04/10: 31 (1.2 mm) - 05/10: 36 (1.4 mm) - 06/10: 26 - 07/10: 12 - 08/10: 16.5 - 09/10: 13 - 10/10: 17 - 11/10: 16 - 12/10: 25 - 13/10: 13 P1 corr_x moves the other way. The pose comes back slowly every night, with the slow monthly drift on top (tilt up, rotation down, as in SNCF01-7). The ends are therefore read where pan and tilt stop rising and stay, or at the next start. - **Images**: noon sheet: sunny on the strong days, overcast on 07, 09 and 11/10. - **Confidence**: high on the cause and the family; medium on the night ends (the pose drifts). ### Obstruction, passages (-12, -17, -18, -19, -30, -31, -34) - **Phenomenon of the catalogue**: A pass-by hides the targets for the length of one measurement (a reference behind it makes the base jump); for -31, a target left under the threshold latches onto a light. - **Family**: Obstruction. Trains are seen at 05/10 16:00, 07/10 13:00, 07/10 16:00 and 12/10 16:00. The others are deduced from the same signature (the camera takes its images on the hour). - **Measured**: the remaining targets jump at the instant. In mm: BG.3 -1.19 px ≈ 7 mm, P1 -0.96 px ≈ 6 mm, P4 -0.81 px ≈ 5 mm. For -17, only BG.3 +0.08 px ≈ 0.5 mm and F4.5 ≈ 0.4 mm. For -31, CV.4 +0.15 px ≈ 1.3 mm. - **Confidence**: high, except -31 (medium: no image). ### No data (-22, -25, -35, -37) - **Phenomenon of the catalogue**: A gap common to several stations (09 and 14/10, with FLORIDA). For 10 and 13/10, SNCF01 alone; on 10/10 the image exists ("The station runs, the measurement is missing"). - **Confidence**: high that nothing was delivered; the cause is not established. ## 2. What I left without a label - **06/10 06:45 (old -13, removed)**: CV and F4.2 rejected. The remaining targets move 2-3.3 noises in a common direction (P2 y +3.3, F4.6 y +3.0, F4.5 x +2.8, P1 y +2.6, P3 y +2.6, P4 y +2.5, BG.2 x +2.1). Under the threshold. - **09/10 by day (old -24, removed)**: CV.4 x +5.1 at 09:15, a single isolated value. P5 x +5.8 at 07:15 is the start of the thermal cycle. CV.2 rejected at 09:30 is a missing point. Overcast. Max 5.1 noises with no other exceedance: not an agitation regime. - **07/10 by day (never labelled)**: CV.4 x +5.4 at 11:15 and CV.2 y -6.2 at 13:45, two isolated single values on a day of about 1,500 daytime target values. Overcast; the error stays at its night level outside the passages. I count it as tail noise, consistent with 09/10. P6 y -5.1 and P7 y -6.1 at 18:30-18:45 fall inside the thermal label -16 (nightfall). - **References**: R1, R2 and BG.3.1 corr at 08/10 12:30-13:30 (±6-8 noises) are inside -20. References are not labelled targets, and no target moves with them beyond the shimmer. - **Target issue targets** (CV.6, CV.7, F4.3, F4.4, BG.1, F4.1): their noise is covered by SNCF01-1 to -6. ## 3. The documentation - **What helped**: - `labeling-rules.md` §1, the 5-noise rule: it decided -13 and -24, and the reading of the remaining targets at a passage. - §2: "never at midnight", "later of the pose and the curves", "without rest, ends at the next start". These made the old ends of -15, -16 and -29 visibly inconsistent with -9, -10, -21, -23 and -26. - §5: a sentence copied from a model must be reread on each label. The copies of "the evening and the evening" and "and some P" were errors of that kind. - **What was missing**: no rule for a pose that never rests because a monthly drift sits on top of the daily cycle. I took "the level where it then stays for a few hours" (pan and tilt flat until the morning drop). - **Not explicit in the rules**: a criterion for isolated single measurements at 5-6 noises on a long series with many values. I treated a single isolated exceedance with no regime as tail noise (07 and 09/10) and kept a day with seven exceedances (11/10). The user may want to settle this. - **The spanning labels**, read within the range: - SNCF01-5 (F4.3) says "about one measurement in two rejected". Within this range it is 254 of 984 (about a quarter). The regime still holds, so the label is not wrong, but the share is lower here. - SNCF01-1, -2, -3, -4, -6 and -7 match the data within the range: BG.1 and F4.1 are rejected at every measurement; CV.7 667 of 984; F4.4 670 of 984; CV.6 in night runs on 09 and 12-13/10. For -7, the corr_y of F4.6, F4.5, BG.2, P4 and CV.8 rise and P1 drops slowly, and the rotation drops while the tilt rises. ## Difficulties - `precedents.py` failed: Ollama unreachable at its URL. The first call was killed by its timeout, the second reported the connection error. Precedents were read by `grep` in `labeling/auto/` instead. - No weather folder for SNCF01, and no temperature (metrics empty). ## 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`, `labeling/auto/SNCF01.jsonl` (labels around the range), `work/existing.jsonl`, `work/spanning.jsonl`, `labeling/auto/{NORDEND,WILDSTRUBEL}.jsonl` (grep "shimmer"), the station `data/SNCF01/` (exports, camera, metrics, images), and the inventories of FLORIDA, JUNGFRAU, NORDEND, WILDSTRUBEL, LENZSPITZE, NIEDERHORN, CLARIDEN, ALASKA, WYOMING, NADELHORN, SUSTENHORN (gaps).