What is known.
What must be earned.
This prototype begins with a strict evidence boundary. It connects Numbering Nature to its strongest technical and visual precedents without claiming that related projects, metrics or datasets are the same thing.
Three states prevent one false story.
A three-city, three-day challenge
The official programme names Amsterdam, Laval and London, visual AI developed by the Amsterdam team, and a 24–26 August 2026 hackathon at AMS Institute.
B++ is technical lineage
B++ is the strongest public method precedent, but no primary source says Numbering Nature is identical to B++, uses package v2.4.0, or shares its exact schema.
The comparison denominator
Exact sites, active sensor-hours, model bundle, taxonomy, rights, protocol comparability and a blocked human-labelled holdout are not public.
From a detection to a defensible decision.
Reveal where and when observed urban insect activity differs with habitat and weather, show how uncertain that pattern is, then recommend the next measurement that would reduce uncertainty most.
- 01
Evidence contract
Keep project observations, historical context, model outputs, simulated paths and unknowns as visibly different evidence classes.
- 02
Comparable activity
Deduplicate confirmed encounters and divide by valid active sensor-hours. Never present frame counts as population abundance.
- 03
Blocked validation
Hold out whole city, site, device and time blocks; report calibration, abstention and long-tail performance alongside aggregate accuracy.
- 04
Decision endpoint
End on the next camera, expert review or habitat measurement most likely to reduce decision uncertainty.
Do not transfer the paper’s accuracy.
The 2024 paper reports a YOLOv8-based image pipeline trained across 2,584 species. The audited B++ package v2.4.0 instead combines GMM motion detection, Hungarian tracking, topology checks and hierarchical ResNet classification. Those are materially different systems, so the paper’s reported accuracy does not validate the current package or the workshop data.
Open the evidence behind the argument.
Full source annotations, data contracts, modelling plan, B++ audit and residency runbook are versioned with this application. Public iNaturalist records in the atlas are contextual candidate-taxon priors, never project detections or independent validation.