A specification-grade decision system defined not by marketing claims or hidden assurance doctrine, but by what its public package makes testable.
Most AI systems infer, optimize, predict and generate. In many domains that is acceptable. In some, it is structurally insufficient — where a wrong decision can compromise correctness, auditability, safety, recoverability, legal accountability, or downstream trust. There, the central problem is not model quality. It is governance architecture.
The Verified Normative Agent is a class of decision systems whose verdicts, authority boundaries, evidence obligations, transitions, failure behavior, readmission conditions and public interpretability are all explicitly specified and verifiable from the public artifact itself.
A Verified Normative Agent (VNA) is a specification-grade autonomous decision system whose public package is sufficient to establish that the agent operates under explicit normative authority, with deterministic verdict semantics, bounded jurisdiction, declared validators and gate conditions, controlled state progression, fail-closed behavior, and enough public traceability for implementation, audit and downstream interpretation.
The implementations you build are yours — the package is a specification, not software.
Governing authority comes from explicit normative conditions — not optimization targets, learned preferences, inferred intent, or reward-seeking behavior.
Emits verdicts from a finite, normatively defined vocabulary whose meanings are explicit and stable.
If required validation, closure or verification fails, it closes safely rather than continue by approximation or permissive default.
Scope, jurisdiction and decision objects are declared explicitly. It governs only within that boundary.
Decision progression is conditioned by declared validators, not by opaque internal confidence alone.
Operates through an explicit state model, with declared transitions, blocking conditions and terminal states.
Inputs, outputs, evidence obligations and downstream consumption conditions are specified as normative contracts, not left to interpretation.
An external evaluator or integrator can reconstruct the meaning of the decision path from the public package and emitted artifacts — with no hidden semantic supplementation.
The class constrains the decision system normatively, without prescribing a vendor, programming language, runtime or deployment model.
Specified precisely enough that reconstructions in unrelated paradigms converge on the same verdict — a property anyone can put to the test, not a trophy and not a claim about any single run.
Properties 8–10 — publicly reconstructible, implementation-neutral, paradigm-convergent — are what put the Verified in a Verified Normative Agent: not a claim, a check you can run.
Copied verbatim from the Verified Rehydration Agent specification — the verdict vocabulary, one ordinary transition, and the rule that catches everything else. Not prose describing behavior: entries that decide it.
§6 · Verdict Semantics The normative verdicts are ALLOW, DENY, QUARANTINE, ESCALATE, and READMISSION_PENDING. §12 · Workflow and State Transitions From | Trigger | Guard | To | On failure S0 | startup_complete | V_BOOT_MODE PASS | S1 | S5 or ESCALATE Any state | rollback_detected | V_ANCHOR or V_COUNTER_CONT FAIL | S5 | QUARANTINE + ESCALATE
Read the last row again: fail-closed is a table entry, not a promise — any rollback, in any state, routes to fail-safe. Every agent in the library is written this way; this one is released — inspect it →
The SEMANTICHOR library is written for a reading medium the market does not yet treat as primary: the large language model. Our specifications are terminologically dense, rigidly structured and semantically closed. To an untrained human reader, those properties read as friction.
To an LLM, the same properties are the natural contact surface: a controlled vocabulary, an explicit state machine, decision tables with no ambiguity, input/output contracts with every mandatory field enumerated, and a semantic index built precisely to be traversed by a reasoning engine. We did not write specifications that happen to be machine-readable — we wrote them for the reader that builds from them.
A Verified Normative Agent is not defined by marketing claims, internal process, or hidden assurance doctrine. It is defined by what its public package makes testable — which is why the Verified Rehydration Agent is released in full, free of charge, to inspect.
Inspect the VRA package → · What defines these agents → · Explore the agents →