Files
aimsight-optimizer/archive/batches/2026-11-05/improve/validation.md
T

2.3 KiB

Validation

Verdict: keep

Answers

  1. The detectors load: no skipped line in benchmark-before, S1, S2 or after; the benchmark rerun at HEAD is identical to benchmark-after.
  2. No station goes back: benchmark-before, -S1, -S2 and -after are identical, line for line, on all 11 stations and every family.
  3. The code stays simple and the gain justifies it. S1 replaces the median gap with a general measure, the cycle per clock hour. It adds no constant and removes 286 false No data instants (0 % correct) on SUSTENHORN. S2 reuses the existing "inside heat shimmer" filter for Poor visibility episodes of point 3, as a one-line change. It drops duplicate events, so the per-instant score does not move, but the same instants are no longer reported twice. Neither change is fitted to the range: no value is cut to its bounds.
  4. Nothing forbidden: the six detectors are unchanged in number and family, and detect keeps its signature and episode keys. Only the standard library is used and no constant was added. Station names appear only in the docstrings and thresholds.md as provenance, never in a rule. The MISSING line of thresholds.md is updated.

The range

SUSTENHORN, 2026-11-05 → 2026-11-06:

  • No data: correct 0/286 (0 %) → no instants flagged (the false day of No data is gone).
  • Every other family on SUSTENHORN, and all of BREITHORN and LENZSPITZE: unchanged.

Total: No data 0/286 → none. The other families are unchanged: Displacement 33/33, 33/36; Obstruction 31/66, 31/37; Thermal displacement 513/513, 513/519; Visibility 20/35, 20/34.

Non-regression

No line changed, on any station or family, between benchmark-before and benchmark-after (nor in benchmark-S1 or benchmark-S2). The Total block is unchanged: Displacement 5746/6345, 5746/20444; No data 75/836, 75/100; Obstruction 1044/4061, 1044/3718; Target issue 19080/22627, 19080/19467; Thermal displacement 10897/16142, 10897/14828; Visibility 3412/7207, 3412/4442.

Reservations

  • No data, S1: the cadence step can only rise, so a station whose hours are only partly filled sees a longer step and needs a longer hole before it reports one. This costs nothing on the reference truth (No data 75/836 and 75/100, unchanged). Watch for missed short holes on stations with an irregular cadence.

Tools

none