EMPIRICAL EVIDENCE / 3 CASES / 3 ARCHITECTURES

The benchmark
behind the verdict.

A live comparison of what each reasoning architecture can see, protect, and prove in documentary credit examination.

BUNDLED DEMO REPLAY

ARCHITECTURE COMPARISON

Not all correctness is the same.

Each lane exposes a different failure mode—not just a score.

VERITAS-600

100%Hard invariant preserved
Neuro-symbolic synthesis

Single-prompt LLM

40%Semantic judgement without locks
Silo blindness remains

Strict rules

40%Exactness without commercial context
Wrongful dishonor remains

CASEBOOK

Open the evidence.

SELECTED CASE / GOLDEN

Insurance 108% Shortfall ($9,600) + Curable Demurrage

Governing ruleUCP 600 Art. 28(f)(ii) & Art. 16(c)(iii)
Strict rulesREFUSAL

Caught the numerical threshold but cannot explain the operational remedy.

Single-prompt LLMAPPROVE_HONOUR

Treated 108% as ‘close enough’ to the mandatory 110%.

VERITAS-600ISSUE_MT734_REFUSAL

Python locks the 110% rule; the system produces the lawful refusal and re-presentation advice.

Provenance is explicit. These featured cases are a clearly-labelled bundled replay until a server benchmark is available.

Trace a decision