AI output is not authority.
LLM output is a noisy measurement, not a final answer. Phionyx governs the space between model output and real action: safety gates, bounded decisions, audit trails, replayable runtime evidence.
Phionyx does not make the model deterministic. It makes the runtime evidence path deterministic.
What each profile addresses
Three runtime profiles.
Each profile addresses a distinct governance challenge — not a monolith. Each does one job; together they keep AI output accountable to evidence.
The notary
Evidence Runtime
A signed, replayable record of every AI decision — tamper at any link is detected. The record, not the model's word, is the account of what happened.
The boundary
Abstention & Boundary
When the system can't evidence an answer, it hedges, asks, defers, or refuses — with a calibrated confidence, not a guess.
The gate
Safety Gate
Fail-closed gates between model output and action, with escalation to a human. The model proposes; the runtime decides what may proceed.
What changes
The difference is what happens before action.
Without Phionyx
The model's suggestion can become the system's behaviour. No deterministic record of scope, safety, approval, or replay.
With Phionyx
Every transition passes verifiable gates and produces a replayable audit envelope. The model stays probabilistic; the path is reproducible.
Verifiable artefacts
Real-world demonstrations
Applications
Safety Gate + Evidence Runtime in practice
HearthOS
A bounded-authority household assistant. The system proposes; the responsible adult decides. Every decision is recorded.
See HearthOS →Governed runtime in practice
Trace — School RPG Demo
A school-safe RPG demo built on the same governed runtime, with each turn bounded and recorded.
Open the School RPG demo →A model saying “fixed” is not evidence.
It is a claim.
Runtime evidence for agentic AI. Tool calls · agent claims · gates · audit chains · replayable runs.
See the runtime evidence protocol →