R-22 · 3 min · Engineering Note
A Verdict Is Not a Sentence
Published Sep 22, 2026
Abstract
A NOMOS verdict — AUTHORIZED, DENIED, or ESCALATED — is never a word a model wrote. It is the result of evaluating a sealed rule set against submitted facts, deterministically, with no generation step anywhere in the path. This note describes how we hold the tooling that builds sealed artifacts to the same standard: every value a model is asked to choose from a fixed set — a severity, a category, a coherence status — now arrives through one validated channel, and any output that cannot be verified is treated as unverified rather than as clean.
R-14 · Note
Confidence ≠ Authority →
Why a model's confidence is not authority, and why a classification needs a defined shape before it can shape what is sealed.
Concept
Two Kinds of Model Output
Language models help NOMOS read policies, triage evidence, and surface contradictions before a policy is sealed. Their output falls into two kinds of work that look alike on the page but are fundamentally different.
- A classification is a choice from a small, fixed, named set: a severity of low, medium or high; a conflict type; a document type; an overall coherence status. It has a checkable answer space.
- A generation is open-ended text: a quote, a summary, a resolution question, a rationale. There is no fixed correct sentence to check it against.
A classification that arrives as free text is a decision without a defined shape. If it is going to shape what gets sealed, it has to be typed and checked before anything downstream relies on it.
One Enforced Path
Every place the platform asks a model for structured output now goes through a single function. It strips formatting, parses the result, and validates it against a schema that names the allowed values for every classification field. Free-text fields remain free text. Anything that fails is rejected with a labelled error and routed to that call site's explicit fallback — it never flows downstream unchecked.
server/lib/model-output.ts
const audit = parseModelJson(output, ConsistencyAuditSchema, 'consistency-audit');
// ConsistencyAuditSchema (excerpt)
conflict_type: enum [threshold_mismatch, owner_mismatch, policy_vs_practice,
definition_drift, missing_owner, other]
severity: enum [low, medium, high]
overall_coherence: enum [COHERENT, CONFLICTS_DETECTED,
GHOST_TERMS_DETECTED, ASSUMPTIONS_DETECTED]
statement_a / statement_b / resolution_question: free textWhere the platform already used forced tool-calling with schema validation — policy extraction and semantic conflict detection — nothing changed. Those were the pattern the rest of the pipeline now follows.
Fail Closed
Validation only matters if the failure path is honest. An audit that could not be verified is now never reported as clean.
- The Coherence Auditor returns INSUFFICIENT_DATA — not COHERENT — when its output cannot be validated.
- The Consistency Auditor, when it cannot validate its result, reports that contradictions were not checked and flags the audit as unverified, instead of reporting a coherent policy.
- Evidence triage and extraction fall back to their existing conservative defaults and require a re-run, rather than accepting a malformed result.
An unverified check and a passed check are different facts. The system now records them differently.
What This Does Not Touch
The deterministic core is unchanged: rule evaluation, sealing, and nomos.can() do not call a model. This work concerns the tooling that helps a person author a policy before it is sealed — the stage where contradictions, undefined terms, and missing owners are caught.
Next: Calibrated Confidence
Schema validation guarantees a classification has a legal value. It does not tell you how likely that value is to be right: a confidence number a model writes about its own output is a self-assessment, not a measurement. We are evaluating typed classification models that return a probability distribution over the allowed answers, and we will adopt one only for classifications — never for the deterministic core — and only after it passes two tests on our own fixtures: the same input returns the same answer, and confidence drops on statements known to be false. We will publish the results.
Conclusion
nomos.can() exists because a verdict should be a value with a defined shape, not a sentence a model happened to write convincingly. The same principle now governs the tooling that produces sealed artifacts.
If a verdict is worth sealing when it leaves the building, it is worth typing before it does.
Cite this
This is a field in formation; this page will be revised. Cite the version date below, not the day you happen to be reading this.
Suggested citation
SafeHaven (2026). A Verdict Is Not a Sentence. Computable Authority Research Library, version of 2026-09-22. https://www.computableauthority.com/research/a-verdict-is-not-a-sentence
BibTeX
@misc{computableauthority-a-verdict-is-not-a-sentence,
author = {SafeHaven},
title = {A Verdict Is Not a Sentence},
howpublished = {Computable Authority Research Library},
year = {2026},
note = {Version of 2026-09-22},
url = {https://www.computableauthority.com/research/a-verdict-is-not-a-sentence},
}