Strict text can wrongfully refuse.
A harmless typo or general non-conflicting description is not necessarily a discrepancy.
DOCUMENTARY CREDIT EXAMINATION · UCP 600
Every document.
Every rule.
One defensible decision.
VERITAS-600 turns documentary-credit evidence into a traceable bank action—locking objective UCP 600 checks, constraining semantic AI, and requiring human authorization.
Deterministic checks locked · Semantic review bounded · Human sign-off required
GUIDED PREVIEWPREDEFINED CASE
RECOMMENDED ACTION
ISSUE MT734REFUSAL
Insurance shortfall $9,600
HUMAN AUTHORIZATION REQUIRED
LC-2026-9104 · predefined evidence · not a live run
01 / THE PROBLEM
Trade documents do not fail in only one way.
AI cannot waive an objective fact.
A fluent explanation cannot make 108% insurance satisfy a 110% threshold.
Documents alone can miss external risk.
Documentary compliance and regulatory exposure are distinct decisions that need distinct evidence.
02 / THE HARD INVARIANT
Interpretation may add context.
It cannot erase a failed calculation.
Objective UCP 600 checks are calculated in Python. Nemotron is used for bounded semantic and lexical reasoning, never to override a locked discrepancy.
OBSERVED COVER108%
LOCKED RULE
REQUIRED MINIMUM110%
DISCREPANT · $9,600 SHORTFALL
03 / TRACE ONE DECISION
Follow one evidence chain.
Nothing is hidden.
Guided demonstration · predefined LC-2026-9104 evidence · not a live examination.
Establish the terms that govern
Accepted amendments become active terms. Pending amendments remain quarantined under UCP 600 Art. 10(c) and 10(e).
OUTPUTActive term snapshot
04 / ACCOUNTABLE ARCHITECTURE
Five owners.
One chain of custody.
Each stage has a distinct responsibility, output, and execution mode. The system does not collapse them into a single opaque “AI decision.”
Meterless / Memory
Resolves active terms and routes scoped context.
Zetaris
Federates operational, customs, and watchlist evidence without ETL.
Python
Calculates dates, thresholds, amounts, and locked failures.
Nemotron
Assesses bounded semantic questions with legal citations.
Policy engine → Human
Assembles a recommendation, then requires examiner sign-off.
05 / THREE POSSIBLE ACTIONS
Honour. Refuse. Hold.
Every recommendation stays pending until a human examiner authorizes the consequence.
06 / INSPECTABLE PROOF
Evidence does not disappear
when the decision arrives.
VERITAS keeps the chain visible from source record to examiner receipt, including execution provenance when the real terminal runs.
- Source digest
- Active-term hash
- Federation receipt
- Decision digest
- Examiner receipt
07 / ARCHITECTURE BENCHMARK
Safer reasoning is not a style choice.
Compare the strict rule baseline, a single-prompt model, and the hybrid system across the evaluation suite.
Strict rulesRigid textMisses semantic tolerance
Single-prompt modelFluent reasoningCan waive objective facts
VERITAS hybridBounded authorityEvidence, rules, semantics, human gate
VERITAS—600
Inspect the evidence.
Then make the call.
Decision support prepares the case. Human authority releases the action.