Problem
By the time an invasion is officially confirmed, eradication is usually no longer realistic. The signals that could catch it earlier already exist and are already open: citizen-science sightings, climate data, documented introduction pathways. Nobody systematically fuses them. Beachhead watches for the specific combination that should trigger a field survey: a plausible-but-unconfirmed sighting, in a climatically suitable location, with a documented pathway that could have brought the species there.
Constraints
- A retrospective evaluation is worthless if the future leaks in. Every data source has to answer “what was visible on this date?”
- Citizen-science records can be misidentifications, mislocated, or dated decades before they were uploaded.
- Alerting inside a species’ already-established range is the dominant false-positive mode, and a present-day checklist of established ranges would itself leak the future.
- The pathway literature (CABI) is readable but not licensed for bulk redistribution, so it can’t ship in the repo.
System
Four signal agents feed an assessment agent. Occurrence monitors GBIF-aggregated sightings for a watchlist, scores each record’s credibility, suppresses established regions, and clusters records into incidents so an analyst sees one row per emerging cluster rather than fifty records. Suitability builds a climate envelope from native-range occurrences (WorldClim rasters, or ERA5 point queries via Open-Meteo) and asks where the sighting sits inside it. Pathway retrieves documented introduction routes from a provenance-tagged corpus and measures the location’s connectivity: port proximity for maritime pathways, distance to the invasion front for natural spread. Propagule pressure is a fourth signal from clustering. The assessment agent fuses them into a banded likelihood with a full evidence ledger, and a brief agent writes a cited survey recommendation with an explicit statement of uncertainty. A review dashboard lets an analyst confirm or dismiss, and persists those decisions as labeled data.
Decisions
An as-of clock enforced at the data-source boundary, not by convention. Without it the time-lag evaluation is unfalsifiable. The evaluation also reports a strict mode that additionally requires a contemporaneous publication and identification, and the gap between modes is large: much of the apparent lead comes from records whose observation date is early but whose upload happened years later. That gap is the real error bar, and it’s reported.
Disqualifying credibility factors, not additive ones. A photographed, precisely-located, research-grade Vespa velutina record dated 1980 in Europe is impossible, and purely additive scoring rated it the strongest in its cohort.
Small in-repo implementations instead of LangGraph, GeoPandas, Qdrant and PostGIS. The pipeline needs point climate queries rather than raster surfaces, nearest-neighbour over 2,896 fixed ports rather than a spatial database, and deterministic replayable retrieval more than a vector service. Each substitution is a documented swap point.
Report the confound instead of tuning it away. Propagule pressure costs precision on escape-from-confinement species, because decades of sustained pet release look identical to an overwintering population. That’s an open problem in the roadmap, not a hidden one.
Outcome
Using a published methodology (González-Moreno et al., 2025, measuring the lag between citizen-science records and official first detection), run against real GBIF data under the as-of clock: 5 of 7 ground-truth cases flagged early, leads of 12–36 months. The ablation finding that matters more than the headline: occurrence alone raises a lead on 100% of control incidents; adding climate suitability drops that to 25% while increasing both cases detected and mean lead time. Suitability is what makes it work.
Sample size is seven. Every rate is a fraction with a small denominator, and every alert is a lead for field verification, never a confirmed detection.