Architecture

A system with structure — not just fluent output.

FineSchema does not rely on fluent generation as proof of reasoning. It separates interpretation from verification, so every judgment can be inspected.

Pipeline

Seven pipeline layers, one inspectable judgment.

Click a layer to see its input, output, and a worked example.

Guarantee

Every output must show its work.

  • the claim being evaluated
  • the evidence used
  • the schema applied
  • the contradiction detected
  • the uncertainty retained
  • the final justified state
VerificationOutput · API-ready
type TruthState =
  | "VERIFIED"
  | "LIKELY_TRUE"
  | "CONTESTED"
  | "WEAKLY_SUPPORTED"
  | "LIKELY_FALSE"
  | "FABRICATED"
  | "UNKNOWN"
  | "UNCERTAIN"
  | "BLOCKED";

type VerificationOutput = {
  input: string;
  truthState: TruthState;
  confidence?: number;
  verified?: string[];
  likely?: string[];
  uncertain?: string[];
  contradictions?: string[];
  missingEvidence?: string[];
  finalJudgment: string;
  recommendedAction?: string;
};
Evaluation protocol · in preparation

No benchmark numbers yet. This is what we will publish.

We will not quote accuracy we have not measured. The protocol below defines how FineSchema will be evaluated — publicly, with failure cases included.

01planned
Emotion inference cases

Curated dialogue scenarios with context, scored on hypothesis plausibility and calibration of stated uncertainty.

02planned
Historical claim verification cases

Constructed case files with controlled evidence sets, scored on truth-state assignment against expert annotation.

03planned
AI action verification cases

Proposed agent actions with rationales, scored on defect detection — overgeneralization, proxy discrimination, evidence gaps.

04planned
Baseline comparison

Against a naive classifier, an LLM judge, and a rule-only gate — same cases, same scoring.

05planned
False positive / false negative review

Every disagreement manually reviewed; both over-blocking and under-blocking reported, not just accuracy.