I build runtime governance and evidence infrastructure for agentic AI — deterministic control paths, measurement discipline, decision evidence, and auditable execution boundaries.

Phionyx treats large language model output as a noisy measurement, not a final answer. The work puts a governed runtime between an AI system and real-world action — and, separately, defines a neutral way to write down what that runtime actually decided, so an outsider can check it. A governed action carries policy + a state transition + an abstain / block / rewrite decision + a replayable evidence record in one inspectable trace. An optional self-claim gate checks development-agent claims against repository evidence, and a runnable control-hardening demo shows what holds, what fails closed, and what stays a documented gap. This sits alongside the emerging agent-governance tooling rather than against it — a capability boundary, not containment.

Four research lines, each with its own artefact and its own claim boundary:

Runtime governance — a canonical pipeline with a state vector, kill switch, human-in-the-loop queue, ethics and safety gates, and a hash-chained audit trail (optionally Ed25519-signed). Its founding axiom: LLM output is a sensor reading, governed before it becomes action — not an oracle to be trusted.

Runtime evidence (AIREP) — a neutral, vendor- and model-independent record format: one hash-chained, canonical-JSON record per AI runtime decision — what was decided, on what basis, and, distinctively, what the evidence does not cover. Checkable offline by two cross-language first-party verifier implementations (Python + Node) that agree byte-for-byte; it depends on no Phionyx code. Experimental — a proposed open format, not a ratified standard. Phionyx currently emits RGE records; no current RGE→AIREP conformance claim is made — RGE is being developed toward AIREP interoperability.

Measurement discipline — the doctrine the rest of this is held to: a governance system that returned safe may have checked nothing. A governance verdict is a measurement of a claim about a system, and a measurement that was not taken must not be reported as one that passed. Seven axioms, a six-value verdict algebra (PASS · FAIL · NOT_MEASURED · INCONCLUSIVE · ERROR · NOT_APPLICABLE), and two prohibitions: an absence is not a result, and a proxy is not the thing. The release includes an audit of my own runtime against it — none of its findings pass, one was contributed by an outside reader, and it is documented rather than quietly corrected. No conformance claim is made about any implementation, mine included. CDE-12 is the companion instrument: criteria for what a system can and cannot record about a control decision and its fate.

Persistent causal traces — Resonatus, an OWL 2 DL ontology with a SHACL validation profile for processes whose causal history must remain inspectable over time: w3id.org/resonatus. An ontological model with verification artefacts — it makes no physical or empirical claim.

phionyx-conformance runs executable scenarios against our own released tooling and publishes the results, including the failures: every current scenario finds a real gap, each result is a frozen machine-readable record, and nothing in it has been independently reproduced yet — it says so itself. measurement-axioms states what should be measured; AIREP states how a decision should be recorded; this repository is the standing attempt to falsify our own claims with published method.

Bridges that export the same runtime evidence into common stacks — each records events as hash-chained envelopes, optionally Ed25519-signed:

• phionyx-mcp-server — MCP trust boundary: descriptor hash pinning + an audit chain over third-party MCP tool calls
• phionyx-pipeline-mcp — the self-claim gate: verifies an agent's own change/test claims against git-diff truth
• phionyx-eval-inspect (Inspect AI .eval logs) · phionyx-langchain-langgraph · phionyx-openai-agents · phionyx-eval · phionyx-letta · phionyx-compliance

• hearthos — bounded-authority household AI: a browser demo with policy gates over every suggested action. A reference application, not a product.
• trace.phionyx.ai — narrative-coherence for game/NPC and storytelling systems: detects character drift and incoherent state before a scene reaches the player.

• LLM output is not truth; it is a signal requiring governance.
• A governance verdict is itself a measurement, and one that was not taken must not be reported as one that passed.
• AI systems need runtime control, not only prompt-level safety.
• Safety, coherence, and telemetry should be structured before a response is released.
• Evaluation must include behavioural stability, not only benchmark performance.
• Human-facing AI should be explainable, auditable, and interruptible.

• Website: phionyx.ai
• The Measurement Axioms: github.com/halvrenofviryel/measurement-axioms · DOI
• CDE-12 (Control-Delivery Evidence): DOI
• AIREP protocol: github.com/halvrenofviryel/ai-runtime-evidence-protocol
• Resonatus ontology: w3id.org/resonatus · DOI
• Trace (narrative + School RPG demo): trace.phionyx.ai · @trace_phionyx
• Substack: phionyxresearch.substack.com
• X: @phionyx_ai
• ORCID: 0009-0002-3718-4010

If runtime evidence for agentic AI is a problem you have, watch phionyx-research to get email updates when we ship new experiments.