Observation
Authorized inputs are received from connected systems — AI providers, federations, leagues, clubs, sensors, VAR infrastructure, smart devices — through governed API connections. No uncontrolled execution.
Platform / 1.1
What is H.A.R.I.
The canonical definition, the functional layers, and where the platform sits in a high-stakes AI environment.
Canonical definition
H.A.R.I. is a complete, modular AI nervous system architecture — not a concept, not a single model, and not a consumer product. It is designed to supervise, coordinate and safeguard complex AI-driven environments in a human-centric, privacy-preserving way.
01The problem
High-stakes AI environments — football governance, clinical research, stadium safety, critical infrastructure — share a structural failure mode: the systems making consequential decisions are probabilistic, opaque, and irreproducible. The same input does not always produce the same output. Decisions cannot be replayed. Audit trails are partial, scattered across vendors, and frequently broken at the boundary between a model's output and the action that follows.
This is not a research problem. It is a governance problem. Regulators, auditors, insurers, and the institutional principals responsible for outcomes do not have a deterministic layer to point to. Without one, accountability collapses upward — to whoever happened to deploy the model — or sideways, into model providers who cannot be held to the institutional standard.
02The approach
H.A.R.I. is not another model. It is a deterministic middleware layer between AI systems, data, sensors, APIs, and the consequential actions they affect. It does not generate. It does not classify. It does not score. It governs.
For every action that crosses the layer, H.A.R.I. produces one of three canonical outputs — ALLOW, DEFER, or SYSTEM_UNVERIFIED — together with a hash-chained, signed record of the inputs that produced the decision. The same inputs always produce the same output. Any decision can be reconstructed from the chain. Irreversible actions require explicit human authorization within an identified scope.
This is what makes the platform institutional rather than experimental: the layer between intent and consequence is reproducible, signed, and reviewable.
03Functional layers
H.A.R.I. is organized as a small number of architectural layers, each with a single responsibility. The boundaries between layers are explicit so that scope, audit, and validation can be reasoned about per layer.
Validation
Inputs are validated against policy constraints. Missing, malformed, or untrusted inputs route the decision to SYSTEM_UNVERIFIED rather than producing a guess.
Deterministic governance
The deterministic kernel evaluates the validated inputs against the active policy and emits one of three canonical outputs. Same input, same output, every time.
Forensic logging
Inputs, policy version, output, and provenance are written to a hash-chained log with ECDSA P-256 signatures. The chain is the audit trail.
Recovery path
DEFER and SYSTEM_UNVERIFIED are not failures. They are pause states with a defined escalation path: a named human, within scope, resolves the gap and the decision either proceeds or does not.
Human authority preservation
Irreversible actions require a human authorization step. The Time Sovereignty Layer ensures that the authorizing human is identified, in scope, and authenticated — before the action proceeds.
Auditability
Any prior decision can be reconstructed deterministically from the chain. Reviewers and regulators receive the same answer every time. Audit is a property, not a process.
04Technical foundation
The platform is built on conservative, well-understood technical primitives. The institutional posture is that boring components in a verifiable arrangement are stronger than novel components in an unverifiable one.
- PostgreSQL as the system of record for the forensic chain and policy versions.
- Hash-chained audit trail linking every decision to its predecessor. Tampering is detectable in O(1) per record.
- ECDSA P-256 signatures on every chained record. Standard, audit-grade cryptography.
- Deterministic kernel with three canonical outputs. No probabilistic component participates in the decision; probabilistic inputs (model outputs, sensor data) are treated as evidence to be evaluated by deterministic policy.
- Forensic timestamping via OpenTimestamps. The current sealed-core proof is registered under ID
SOV-2026-02-26.
05Where H.A.R.I. sits
H.A.R.I. is institutional middleware. It is not a replacement for the AI systems, data sources, or operational tools an organization already runs. It is the deterministic layer between those systems and the consequential actions they participate in.
For the kernel internals and the formal definition of the three canonical outputs, see Core architecture. For the forensic chain and timestamping, see Determinism & forensic traceability.
Back to top