Home / Responsible AI

Responsible AI boundary

We support human judgment — we don't replace it.

StunAssure is a decision-support and risk-management tool. It identifies observable risk indicators and process deviations, explains every alert, and keeps welfare experts and operators central to every decision. Here is exactly what that means.

The boundary

What we do — and what we do not claim.

What StunAssure does

  • Flags observable welfare-risk indicators and process deviations.
  • Explains each alert: what, when, and which data supported it.
  • Routes uncertain cases to human review.
  • Produces audit-ready evidence for operators, vets, and certifiers.

What it does not claim

  • It does not determine whether a fish is conscious.
  • It does not guarantee humane slaughter.
  • It does not replace stunning equipment or experts.
  • It does not publish a fixed "accuracy %" without validated evidence.

Fail-safe by design. In animal welfare the dangerous error is a false "safe." Lower-cost tiers are deliberately more conservative — they produce more "uncertain — check manually" outputs rather than ever falsely certifying a fish as insensible.

Design principles

The rules we hold ourselves to.

Practical before fancy

Usable in real farms and vessels — offline-first, simple alerts, low operator burden.

Fail-safe

Uncertain cases trigger manual review, never false confidence.

Explainable

Every flag shows its reasons and supporting data; no black-box verdicts.

Expert-validated

Welfare indicators are reviewed with domain experts, not asserted by software alone.

Species-configurable

Profiles adapt thresholds to species and local conditions.

Evidence-based

Visible signs alone may be unreliable; we combine them with process data and expert validation.

Known limitations

What we are honest about.

Stating limits plainly is part of responsible practice.

  • Video and behaviour alone cannot prove unconsciousness; physical signs may not match neurological evidence.
  • Performance depends on data quality, species variability, and field conditions (wet, low-light, occluded).
  • Early tiers rely on conservative rules and human review while validation data is built.
  • Independent validation against stronger indicators (with experts) is required before any strong welfare claim.
Honest about prior art

We are not the first AI welfare-monitoring system.

Camera-based welfare-monitoring systems already exist for slaughterhouses. Our contribution is specific: fish-specific stunning verification, low-cost deployment, explainable risk scoring, expert-reviewed audit evidence, and multimodal fusion (video + sensors + process) rather than camera-only inference.