H.A.R.I. is a deterministic governance layer that sits between your data, your protocols and the people who decide. It never guesses: it answers ALLOW, DEFER — or names exactly what is missing. Football is the flagship vertical; the same kernel applies to other high-stakes domains.
Priority problem: contested decisions you cannot fully reconstruct afterwards — which evidence was seen, by whom, under which protocol.
Do now: pick one review protocol (VAR, disciplinary, return-to-play) and submit it as a use case.
Decision governance module →Priority problem: return-to-play and squad-welfare decisions where protocol, load data and medical sign-off live in different places.
Do now: share your current protocol steps (no player data) for a written gap analysis.
Welfare & squad module →Priority problem: the right angle, at the right second, to the right person — with a record that it happened.
Do now: request the Field Lab brief — role-based delivery on smart glasses, tested with amateur officials.
Smart glasses & field interfaces →Illustrative policy and data — not a real match. What matters is the shape: the kernel never produces a verdict without evidence, and never stops at “no” without saying what would turn it into ALLOW.
Episode 71′ — possible handball. Bound to the episode: camera A, camera B, the review protocol (policy v4.2) and the named official on duty.
Camera B's frame at the touch moment fails its integrity check — the feed dropped two frames. One precondition cannot be evaluated.
The kernel does not guess from camera A alone. It emits a state, not an opinion — and writes to the chain exactly which input failed and why.
Broadcast angle C is an authorized source; it is bound and verified. The kernel re-runs: evidence complete, decision reserved to the named official (DEFER). On their signed call: ALLOW.
Determinism holds for a fixed tuple: policy version, kernel configuration, input schema and the bound evidence set. Change any of these and you get a new, equally reproducible trace — never a silent drift.
Only the kernel emits ALLOW / DEFER / SYSTEM_UNVERIFIED. Modules that read context — welfare, review quality, situational risk — return advisory indications to a named human. None of them produces a ranking or a score that the kernel acts upon.
Same illustrative policy as above. Toggle the preconditions; the kernel re-evaluates deterministically and names what is missing.
{{ outSummary }}
Middleware between AI systems, data, sensors, APIs and the actions they affect. Same input, same output — by architecture, not by promise.
Read the full definition →Not autonomous, not a surveillance system, not a consumer product. It is an institutional layer that preserves human authority.
See every negation →Copyright registered, architecture timestamped (SOV-2026-02-26), 13-point stress test passed, TÜV SÜD pre-audit under way. Field Lab: smart-glasses delivery tested with amateur officials. Stated plainly, nothing more.
Full status and what each claim means →Decision governance on evidence, forensic match review, multi-angle evidence fusion, smart assistant referee, welfare and squad, referee training and VAR quality review. VAR is one module among many. Reads context. Does not profile people.
Each reuses the same deterministic kernel, forensic chain and human-authority guarantees. Stated as potential where they are potential. All domains →
Not a commitment, not a sales call. Describe one decision you would like to govern on evidence; we reply in writing with whether and how the kernel applies, and what a scoped pilot would require.