# Modification **Verdict**: changed ## Diagnosis BREITHORN (one measurement every 4 h), LENZSPITZE (5 min), SUSTENHORN (bursts: three measurements a minute apart every 15 min, a median step of 1 min), 2026-11-05T00:00Z → 2026-11-06T00:00Z. | Family | labelled | found | flagged | wrongly | |---|---|---|---|---| | Displacement | 33 | 33 | 36 | 3 | | No data | 0 | — | 286 | 286 | | Obstruction | 66 | 31 | 37 | 6 | | Thermal displacement | 513 | 513 | 519 | 6 | | Visibility | 35 | 20 | 34 | 14 | - **SUSTENHORN D1, No data 00:00 → 24:00 (and since 11-02)**: extra, the largest defect on the range (286 flagged instants, none labelled). `no_data.py` takes the median gap (1 min, the gap inside a burst) as the cadence, so every 13-min gap between two bursts counts as a hole. → S1. - **SUSTENHORN D8/D9, D13/D14, D17/D18**: duplicates. The same instants appear twice as Poor visibility: once from point 3 of `visibility.py` ("none", greyed) and once from the diffuse rule of point 8 ("reduced"). The two do not merge because their levels differ. The scanner removed the greyed copies, but the labels cannot choose between two varieties of the same family. → S2, which keeps the point-3 episode. - **SUSTENHORN S1, Obstruction 05:45 → 08:45**: missing. The same 11 targets are rejected for one or two measurements of a burst at 06:01–06:02, 06:32, 07:31–07:32, 07:45–07:46 and 08:31–08:32. The single-measurement cases elsewhere in the day are Brief obstruction (D3, D4…). A run of two falls between Brief obstruction (one instant) and Poor visibility (RUN_MIN = 3), so neither rule catches it. I tried a fix and dropped it (see Not fixed). - **SUSTENHORN D3, D4, D6, D7, D11, D12, D15, D16, D19 (Obstruction), D2, D5, D10 (Thermal), D9, D14, D18 (Visibility)**: correct. - **LENZSPITZE S1, Visibility 08:35 → 09:55**: missing. The error_value stays at its usual level (median 0 to 1 MAD) except at 09:05–09:10, which is two instants at 3 to 4 MAD, one fewer than RUN_MIN. The curves move a bit more (median deviation 0.03 → 0.06–0.10 px) and no target is rejected. The families document says that agitation without a rise of error_value is not a turbulence, so the detectors' silence is a defensible reading. See Limits. - **LENZSPITZE D1–D9**: correct (same as the labels). - **BREITHORN D3, Displacement 08:00**: extra, one instant. COLI22 sits displaced from 20:00 to 04:00 (error_value ×4, not rejected, inside the obstruction D2). The return at 08:00 is measured against the displaced level and counted as a jump ("the jump ended a dip", point 2 of `displacement.py`). Small nuisance, not fixed. - **BREITHORN D1 (Displacement) / D2 (Obstruction), 00:00 → 04:00**: alternative reading, the scanner's doubt. The same dip of POUI22 appears in two families. Both labels are kept and score as found. ## What changed - **S1** (commit 7e3043d), `scripts/no_data.py`: the cadence is now the larger of the median gap between instants and one hour divided by the median number of instants per clock hour (counting only hours that have one). It used to be the median gap alone. No constant changed. Docstring and the `MISSING` line of `thresholds.md` updated. - **S2** (commit 5438b0b), `scripts/visibility.py`, `detect()`: a diffuse-nuisance episode (point 8) that lies entirely inside a Poor visibility episode of point 3 is dropped, as one inside heat shimmer already was. No constant changed. Docstring (point 8) updated. ## Why - S1: when a station measures in bursts (several measurements a minute apart, then a pause), the median gap is the gap inside a burst. Every pause between bursts then counted as a hole, which reported a whole day of No data on a station that was measuring normally. Measured per hour, the cadence is the burst cycle. A missing burst is still a hole. For a regular cadence nothing changes, because both measures agree, and the per-hour count only ever raises the step. - S2: the same instants were reported twice in one family (two Poor visibility events, one per level), because the two rules of the detector both fired. The episode with the rejection-based level is kept. ## Before / after on the range Events command, SUSTENHORN (BREITHORN and LENZSPITZE unchanged, no `skipped` line): ``` - 2026-11-02T00:02Z → 2026-11-06T00:00Z No data kept | No data | all sensors | absorbed: No data 11-02T00:17, … (one per 15 min) - 2026-11-05T16:45Z → 2026-11-05T16:47Z Visibility kept | Poor visibility | 11 sensors - 2026-11-05T18:45Z → 2026-11-05T18:47Z Visibility kept | Poor visibility | 11 sensors - 2026-11-05T23:00Z → 2026-11-05T23:02Z Visibility kept | Poor visibility | 12 sensors (the greyed Poor visibility events at the same times remain) ``` Range score, per family: - No data: correct 0/286 → — (no instant flagged) - Displacement, Obstruction, Thermal displacement, Visibility: unchanged (33/33 and 33/36; 31/66 and 31/37; 513/513 and 513/519; 20/35 and 20/34) ## Benchmark No line changes: `benchmark-S1.txt` and `benchmark-S2.txt` are identical to `benchmark-before.txt` on every station and family, and in the `Total` block. ## For the user to decide - BREITHORN-2026-11-05-D1 (Displacement) and D2 (Obstruction), 00:00 → 04:00: the detectors put the dip of POUI22 in both families. A parked vehicle (Obstruction) is likely, given the history of this station, but only the images can say. The benchmark is unchanged either way. ## Not fixed - SUSTENHORN-2026-11-05-S1 (Obstruction, 05:45 → 08:45). I tried a rule that treats two high instants in a row, salient against the instants around the pair, as a Brief obstruction (a passing shorter than RUN_MIN). On the range it caught only 08:31–08:32 here (06:01–06:02 misses salience by one sensor: 11 against 3 × 4; 07:31–07:46 is two pairs next to each other). Elsewhere on the day it chained the Brief obstructions of 18:30–20:32 into one event through the one-hour merge, so Obstruction "correct" on the range fell from 81 % to 46 %. Dropped. - BREITHORN-2026-11-05-D3 (extra Displacement, 08:00, one instant): the return of a target from a dip (inside an Obstruction that ended at the previous measurement, 4 h earlier) is counted as a jump. Fixing it would mean letting a Target obscured episode absorb the return that follows it, a new host rule for one instant. Not worth one more rule. ## Limits - LENZSPITZE-2026-11-05-S1 (Visibility, 08:35 → 09:55): no rejection, error_value at its usual level except at 09:05–09:10 (two instants), curves only slightly more agitated. Telling a loss of visibility from agitation without a rise of error_value takes the images. The scanner judged it from a figure at a coarse scale. - SUSTENHORN, the periodic rejection of the same half of the targets (11 out of 22) for one to four measurements, day and night: whether it is something passing (Obstruction), the image (Visibility) or the station itself takes the images. The labels mix the two families on the same mechanism (Obstruction for the label S1, Visibility for D9/D14/D18). ## Tools - `batches/2026-11-05/improve/tools/instants.py `: for each instant, the number of sensors, sensors rejected, sensors "high" (rule 1 of `visibility.py`), the median error_value in MAD and the deviation of the corrected curves. Worth keeping as a workflow tool: it shows directly why a visibility or obstruction rule did or did not fire.