Measurement Known-positive control, OLCI 2026-09-08, no credentials: Atlantic belt 0.5298 % of observed pixels flagged algae against clean gyre 0.0022 % — a 241× contrast. Land returns 0 observed pixels. Grid definitions read from the files themselves: MSI 175 001 × 111 112 at 0.00018° (20.0 m N–S, 19.3 m E–W at 18.5 N); OLCI 45 201 × 14 001 at 0.0025°.

Measurement Failed the same known-positive control outright: zero detections in the Great Atlantic Sargassum Belt, with detections appearing only where a coast was nearby. Four rounds of fixes produced four distinct artefact populations rather than a working instrument — land mask rasterised in the wrong CRS, NDVI computed on wrongly-offset reflectance, haze and sun glint, and shallow-water seabed reflectance. A fixed FAI cut fires on anything lifting NIR against a red/SWIR baseline; that is the wrong instrument, not an untuned threshold. It is retained only as a cross-check and is never the answer.

Measurement flag_nfai publishes its meanings against values that include 32767, which is also the array’s _FillValue — a pairing that must not be trusted on sight. Read twice on the 2026-09-01 MSI file: flag=1 → 86 882 605 px, index at fill (no observation / cloud / land); flag=0 → 12 298 467 px carrying the −0.5 algae-free sentinel; flag=32767 → 39 365 px with index 0.0015–0.5795, i.e. algae — of which 12 501 px were themselves fill and are dropped. The reader re-derives and re-checks this on every file it opens and skips the window loudly if a control fails, because a silently reordered pairing would report the entire ocean as sargassum.

Measurement Shore-normal profiles at 200–1 000 m for every segment, reproduced 2026-09-09. Depth 1 km offshore: Bávaro 7.3–12.2 m, median 11.1 m, 0 of 11 segments reach 30 m; La Romana 6.7–42.6 m, median 22.1 m, 5 of 25 reach 30 m. Eight segments are refused rather than estimated — the routine will not answer finer than its grid cell.

Measurement Along the BPB-002 shore normal at 100 m steps, 100–1400 m, GEBCO reads 5, 5, 5, 5, 21, 21, 21, 21, 35, 35, 35, 35, 35, 8 m — four identical readings, a step, four more, then a 27 m drop while still heading offshore. A sea floor does not do that; the profile is an artefact of a cell wider than the question. Arbitrated independently by the imagery: green (560 nm) bottom reflectance on S2B_19QEA_20251230 stays 2–4× background out to ~900 m, which through GEBCO’s claimed 35 m would be attenuated to ~0.7 % and is impossible — consistent with GMRT. The two grids nonetheless correlate at r = 0.9975 across the EEZ window and disagree about the 30 m isobath on only 1.7 % of points, so the basin-scale mask does not depend on which is used.

Measurement 44 segments, 13 146.9 m total. Bávaro: 12 × 301.2 m = 3 614.4 m, normals bearing 26.4°–81.5°. La Romana: 32 segments 295.8–313.6 m = 9 532.5 m, normals 90.5°–246.3°. Every length and bearing is computed from the geometry, not asserted.

Measurement 136 Dominican beaches. Every field is an OSM tag, a count of OSM features within 400 m, or a point-in-polygon result — nothing is written from knowledge. Coverage is therefore uneven and is published as such: province 100 %, surface 43 %, parking 32 %, access type 18 %, lifeguard 4 %, entrance fee 1 %. Fields that do not exist in OSM — activities, wildlife, water conditions, best season — are left empty rather than reconstructed, because writing them would mean publishing 136 unsurveyed claims.

Measurement Over ten days the Bávaro box was observed on 1 of 10 MSI days and 0 of 10 OLCI days; the La Romana box on 0 of 10 and 0 of 10. Worse, the La Romana box as specified (18.396–18.410 N, −68.944–−68.900 W) is 100 % land in the producer’s own mask: along −68.92 W the water CLS will report stops at 18.361 N, roughly 4 km south of the box. That box can only ever return 0 km², and 0 km² on a map reads as a clean beach. The boxes are not silently moved — the La Romana extent is an assumption inherited from the specification, and renumbering a segment invalidates every observation filed against it. The ingest logs an error naming each unobserved box instead.

Measurement 147 of 393 detections in one stored run, and 96 of 293 in another, sat at positive elevation on both DEMs. They are not scattered inland: every one is within 1.1 km of the Natural Earth coastline, median 0.0 km, p90 0.32 km. A 1:10 000 000 coastline carries roughly a kilometre of positional slop — about 17 Sentinel-2 pixels at 60 m — so surf, wet sand and shoreline vegetation stay inside the water class. A depth cut removes them as a side effect; the underlying fault is the coastline dataset and is recorded as its own defect rather than absorbed.

Measurement The same corridors on the same day returned 3.91 km² (MSI) and 556 km² (OLCI) — a factor of about 40, all of it resolution. Reported area is flagged-cell area, not raft area: at 278 m a single 50 m windrow flags a whole 0.077 km² cell. Runs are stored per source so one cannot displace the other, and the legacy risk bands (0.3 / 3 / 12 km²), tuned on 60 m Sentinel-2 patches, are not applied to OLCI.

Status The particle model now exists. An OpenDrift ensemble with GSHHS stranding and windage sampled across 1–3 % runs 4 800 particles over a 72 h horizon in under 20 s, and it passes a known-positive control: 400 particles released at 17–20 N, 58–55 W travel −96 km east in 72 h, 100 % westward at 0.375 m/s — the North Equatorial current, out of measured forcing. So prob_llegada is computable, and /arrivals?engine=opendrift returns it. It is not what this page serves. The default engine is unchanged, because a probability that exists is not a probability that is right: no arrival from either engine has been checked against an observed landfall, so calibrated stays false and the acceptance criteria of VBK-SAR-01 §6 — Brier, ROC-AUC, ETA error — remain unevaluated. That needs the drone, which has not flown.

Status The measurement today is an explainable colour index with a per-image Otsu threshold, and what it emits is coverage of a brown-object class. It is not a trained sargassum classifier and is described that way to the customer. The only flight on record is synthetic. Four flights per day over the 44 segments is the plan that turns this row into a measurement, and it has not started.