See the verification stack run.
This runs the real prototype logic in your browser — simulate a batch, verify each fish, and certify the lot with a signed record. It is a faithful demonstration on simulated data: it never claims to detect consciousness, and it fails safe.
Dose
Recovery-clock
Echo-Stun
Certify
—
Each run designs a zero-acceptance sample for a 25,000-fish lot, simulates and verifies it, then certifies the lot — exactly as uv run stunassure demo does on the command line. Fail-safe: anything not positively Pass blocks certification.
Four layers, decided conservatively.
Instead of one camera guessing "is this fish awake?", the stack measures the stun physically, enforces the wake-before-death window, optionally reads the fish electrically, and certifies the whole batch with statistics certifiers already trust.
Dose
Was an adequate electrical dose delivered? Field strength × duration vs the cited species spec. If a species has no published spec — gilthead sea bream — dose is UNCERTAIN, never a fabricated pass.
Recovery-clock
Was the fish bled before it could wake? Electrical stunning is reversible — salmon can show recovery from ~44 s. We time the stun→bleed interval against a cited per-species threshold.
Echo-Stun
The bold differentiator: a cheap electrical evoked-response "insensibility index" — the first low-cost proxy for lab-only EEG. Optional; never gates the cheaper tiers.
Welfare-by-Sampling
Certify a whole batch at stated confidence from a small subsample — importing ISO 2859 acceptance sampling, which welfare auditing has never used.
Fail-safe, by design. A false “safe” is the dangerous error in welfare, so missing or un-verifiable evidence is always UNCERTAIN → manual check, never a pass. The worst layer dominates the verdict — we never average a failure away. Heartbeat is not accepted as an insensibility signal (the fish heart is myogenic and beats for minutes after brain death).