Ledger

Layers 1-4: claim-level uncertainty quantification and deterministic gate enforcement.

Claim Ledger decomposes every response into atomic claims, verifies each claim against authoritative sources, scores uncertainty across up to four layers, and enforces a deterministic emission gate.

Without it

Without claim-level decomposition, you’re scoring responses as monoliths. One backed claim hides three fabricated ones. You can’t tell auditors which specific assertion failed or why.

Position in the platform

Oracle Foundry (Foundation): Supplies authoritative chunks with provenance, tiering, and SIRE boundaries
SIRE Crosswalk (Post-Foundry): Provides global authority routing and conflict context for cross-oracle claims
Prompt Compiler (L0): Optimizes prompt structure before generation
Claim Ledger (L1-L4): Scores per-claim confidence, entailment, stability, and representation uncertainty
Process Control System (Cross-layer): Tracks drift and process capability over governed output
Forensics Lab (L5): Runs root-cause diagnostics on failed governed responses

Five-stage runtime pipeline

Stage 1 — Claim extraction: Splits model output into atomic factual assertions with claim IDs and claim types.
Stage 2 — L1 calibrated confidence: Converts token logprobs into calibrated correctness probabilities, with separate scoring for critical tokens.
Stage 3 — L2 source entailment: Scores each claim against oracle chunks using embedding similarity + NLI, then applies tier-weighted aggregation.
Stage 4 — L3 stability: Uses multi-draw regeneration and semantic clustering to identify reproducible claims vs stochastic artifacts.
Stage 5 — L4 representation uncertainty: Applies hidden-state probes to detect internal volatility not visible in token confidence.

Deterministic gate logic

Authority boundary check (SIRE exclusions and out-of-scope concepts)
License boundary check for output constraints on licensed material
Worst-claim entailment gate using min(best_score) across all claims
Contradiction detection (numeric predicates, polarity, and NLI where required)
Mode-dependent confidence, stability, and representation uncertainty checks

Governance envelope output

Per-claim audit trail with source attribution, layer scores, and gate reason codes
Complete request provenance: model, oracle, embedding, calibration, and compiler versions
Structured artifact consumed by Process Control System and Forensics Lab for closed-loop remediation