Files
aimsight-labeling/auto/analyses/SNCF01_2026-10-04T0000_2026-11-01T0000/analysis.md
T

19 lines
3.1 KiB
Markdown
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
# Requalification of the long labels: SNCF01, 2026-10-04T00:00Z → 2026-11-01T00:00Z (UTC)
Window: 2114 measurement instants (15 min). Rejection counts from `tools/scripts/inventory.py data/SNCF01 --start 2026-10-04T00:00Z --end 2026-11-01T00:00Z` (`work/inventory.txt`); deviations from `tools/scripts/ranges.py --columns corr_x,corr_y --k 5 --min-measurements 6` (`work/ranges_k5.txt`) and `--rolling 1d` on F4.3, CV.6 (`work/ranges_rolling_f43_cv6.txt`); daily medians of `corr_y_mm`, `corr_x_mm` from `summary.py --step 1d --level window`.
| id | Decision | Evidence |
|---|---|---|
| SNCF01-1 | kept | BG.1 rejected at 2114 of 2114 measurements (100 %), no `corr` at all: rejected almost always, Target issue. |
| SNCF01-2 | kept | F4.1 rejected at 2114 of 2114 (100 %), no `corr` at all: Target issue. |
| SNCF01-3 | kept | CV.7 rejected at 1692 of 2114 (80 %), 422 accepted points; the accepted points after the long rejection runs of 25–28/10 jump by about 191 noises in `corr_y` (and 110 in `corr_x`) with no other target following: rejected almost always and its remaining curve false. Target issue. |
| SNCF01-4 | kept | F4.4 rejected at 1737 of 2114 (82 %), 377 scattered accepted points: rejected almost always, Target issue. |
| SNCF01-5 | removed | F4.3 rejected at 988 of 2114 (47 %): not "almost always"; a label of missing points. Its accepted curve has no event of its own: under a 1-day rolling level, only isolated single measurements at 5 to 8.5 of its own noises (no run of two), the ordinary scatter of a noisier target; it is a target, outside the compensation, so its rejections move no other curve. Its share in the drift of SNCF01-7 is already stated there. |
| SNCF01-6 | removed | CV.6 rejected at 454 of 2114 (21 %): a label of missing points. Under a 1-day rolling level, only isolated measurements at 5 to 9.4 of its own noises (one run of two), no lasting episode; a target, so its rejections move no other curve. |
| SNCF01-7 | kept | Drift of the whole month far beyond 5 noises: `corr_y` of F4.5 from −84 noises (start) to +36 (end) of its level, BG.2 −88 to +30, P1 +25 to −72, P7 +27/+47 to −25, CV.8 −9 to +13; daily medians in mm: BG.2 from −0.92 to +2.36 mm, F4.5 −1.36 to +2.67 mm, F4.6 −1.25 to +2.32 mm, P4 −0.58 to +2.15 mm, P1 +0.50 to −2.03 mm, P7 +0.55 to −0.75 mm; `corr_x_mm` P2 +1.46 to +0.43 mm, P3 +0.44 to −0.40 mm, P4 +1.47 to +0.83 mm (noise ≈ 0.05–0.08 mm per measurement). Flattening at the end of the month confirmed on the daily medians (28/10 → 31/10 nearly flat). Kept as it is. |
Note: the rule of the catalogue "a target right at the threshold, the whole batch, with many own rejections" (`families.md`, Target issue) would have kept SNCF01-5 and SNCF01-6; the instruction of this requalification (and `labeling-rules.md` section 1: a missing point is not labeled for itself) prevails: only targets rejected almost always keep their Target issue.
## Sources read
`tools/GUIDE.md`, `tools/DATA.md`, `labeling/families.md`, `labeling/labeling-rules.md`, `labeling/reading-rules.md`, `labeling/stations.md`, `work/existing.jsonl`, the station data through `inventory.py`, `ranges.py`, `summary.py`.