Evaluation

0/123

Unsafe cards: 55 published cases and 68 real label photos, live code

Every one of 55 published poisoning cases and 68 real label photos, run on the code that is live now: 44 cases specific and correct, 10 of 12 documented misidentifications set right, every other case safely held for confirmation.

0/68

Real label photos: unsafe cards

68 photos of real Indian pesticide labels from product pages, annotated before any run. Two independent vision reads and a Tavily brand check: 49 read correctly, the rest held for the doctor to confirm from the bottle.

107 vs 0

Drug doses with no source: general model vs Antidote

A general model asked the same 55 questions directly wrote 107 drug doses with no source, in 47 answers. Every clinical sentence on an Antidote card is a guideline's own words with its page.

83.6%

Cards consistent with the authors' management

Two independent judges compared each card with what the treating doctors actually did: 83.6% and 98.1% consistent.

100%

Lethal-error sentences blocked

10 of 10 by rules alone

100%

No-antidote labels routed to a stop

20 of 20 chemical groups without a specific antidote

100%

Non-pesticides kept away from antidotes

10 of 10 fertilisers, cleaners, medicines, fuels

100%

Organophosphates kept on the antidote path

10 of 10; target 100, these must not be falsely stopped

96.9%

WHO 2019 class agreement

31 of 32 registry chemicals match the parsed WHO table

100%

Stop card on no-antidote labels (full pipeline)

20 of 20, Nemotron 3 Super composition with registry gates

100%

No atropine or oxime in those cards

20 of 20; the headline refusal

92%

Citation integrity

80 of 87 clinical quotes found on the cited guideline page

3.1625

Ledger total (USD)

all Token Factory calls for this run

2/2

Class-only labels with cholinergic signs, guided correctly

held-out set 2 only; the rule was written after set 1 and tested on cases it had never seen

83.3%

Real label photos read correctly when the chemical is legible

42 photos where the chemical name is readable at photo resolution; 72.1% across all 68

1/68

Same labels, NVIDIA Nemotron Nano 12B VL alone (no cross-check)

NVIDIA Nemotron Nano 12B VL alone, served on our own Nebius Serverless Endpoint: 57 correct, 92.9% when legible. Its single misread is why Antidote requires two independent reads and a brand check.