Accuracy
Lares Pecoris answers only from named, published veterinary sources, and it shows you which source every statement comes from. This page reports how the live product behaves when it is tested adversarially — and states results only.
Verified against the live system
We keep fixed sets of hostile test questions and run them against the same production system that serves every user — not a separate lab copy.
The most safety-critical set targets withdrawal periods and food-chain rules — the figures that decide whether meat or milk from a treated animal may enter the human food chain. It holds 50 items, of which 40 decide a food-chain outcome. In the run of 23 July 2026 against the live system, all 40 food-chain items were correct: every withdrawal figure matched the summary of product characteristics it was drawn from, and every question with no supporting source was declined. Release is gated on this — if a single food-chain figure is wrong, the affected species is held back and does not ship.
A second, standing set of 39 adversarial questions — 24 with a correct, source-backed answer, 10 that must be refused because no named source supports them, and 5 built to bait a wrong number — was last run on 15 July 2026, with zero fabricated numbers and every unsupported question declined, including a prompt-injection probe.
We report the outcome, not the questions themselves. Publishing the test items would let them be gamed, which would make the result meaningless.
What testing has changed
An earlier version of this page said that when the product errs, it errs toward silence and never toward an invented number. Testing proved that wrong: we found a case where it stated a cattle withdrawal figure that did not appear in the source it cited. We fixed the cause and built the food-chain gate described above — a check that now blocks a species from release if any withdrawal number fails to match its source. We publish this because a verification set is only worth as much as the failures it is allowed to surface, and this one surfaced a real one.
We also treat over-caution as a defect. In the 23 July 2026 run, three cattle dosing questions were declined even though the dose appears in the source — the product held back where it could have answered. That errs toward silence, which is the safe direction, but it is still a defect: it is logged, counted, and named here until fixed.
Every answer is sourced
Each answer carries verbatim citations to the specific named sources it draws on. You can see the source behind every statement and check it yourself. When no named source supports an answer, the product says so and declines, rather than improvising one.
What the answers draw on
As of 23 July 2026, the English-language corpus comprises 1,211 named sources: UK Veterinary Medicines Directorate product information for companion animals, the complete equine set, and cattle; manufacturer summaries of product characteristics from the NOAH compendium; and clinical guidance from bodies including AAHA, WSAVA, AAEP, ESCCAP and WOAH, alongside peer-reviewed open-access veterinary literature. Every source is named, and every answer points back to the ones it used.
No live counters
The figures on this page are point-in-time and dated. We do not display moving dashboards or real-time scores. When we re-run the test set or the corpus grows, we update the date.