A False Floor: Refuting Our Own Information-Limit Claim in a Lethal Identification Task

The quantity at issue is the false-eat rate: the share of held-out deadly specimens that a rule wrongly passes as safe to eat. It is always read together with the rule's edible yield, the share of edible specimens it passes, because a rule that passes nothing has a false-eat rate of zero. We had concluded, and believed, that when deciding from a photograph whether a mushroom is safe to eat, this rate has a floor set by the information in the image rather than by the method: on frozen I-JEPA ViT-H/14 embeddings of field mushroom against Destroying Angel, an AUROC of 0.94 coexisted with a false-eat residual near 5% that roughly 15 methods did not remove, and the archived count behind it, 7 of 81 on one split, has a one-sided 95% lower bound of 4.13%. Such a floor is a claim about every image-based rule, so a single rule at equal or higher yield whose upper bound falls below that lower bound refutes it, while any number of rules failing together cannot establish it. This paper refutes that claim on a different fatal pair, honey mushroom against deadly Galerina. The claim is our own, was never published, and is archived in anonymized form. At an empirical 80% edible yield, a common MLP probe on frozen DINOv2 ViT-L/14 embeddings passes 0 of 150 held-out deadly Galerina marginata as safe to eat in each of ten random splits, an exact one-sided 95% Clopper–Pearson upper bound of 2.0%. SigLIP 2 and MetaCLIP 2 also pass 0 of 150 in every split, but because image-text training can carry species names in captions, the refutation rests on DINOv2, which has no text or label channel. On the pair the floor was measured on, the evidence sits at the floor's own lower bound and decides nothing. A second result concerns positive claims: BioCLIP 2 and 2.5 reach AUROC 1.00 on every hard pair, but their model cards declare GBIF, the source of every image here, together with taxonomic label supervision, so we disqualify them on provenance. Why this matters beyond mushrooms: a claim that an image cannot carry the needed information tells a field to stop working on that modality, and our own claim shows how a weak baseline can manufacture one. What is new is a public, checkable case of an information-limit claim refuted by its own authors, together with two rules that carry over to any evaluation of this kind: such a limit can be refuted by a single rule under a different representation but never established by failures, and on benchmarks assembled from a public biodiversity aggregator, a domain model's score counts as evidence about the photograph only after its declared training sources and labels are audited. This paper is a research vehicle and not a foraging aid. Preprint; not peer reviewed. Code and derived feature caches: https://osf.io/d3ft6/?view_only=a7785622ad2e4e1995f2c830359447b3