Deterministic-first governance for invoice exceptions
“Use intelligence only when intelligence is required. Use autonomy only when authority exists.”
Evaluate your own synthetic invoice in the browser. No account, provider calls, real payments, or financial authority.
- 66.7% fewer inference requests
- 65.5% fewer tokens
- NONE AI-granted execution authority
Recorded from the S4E.15 two-model AMD comparison — six synthetic development cases per model, one repetition, sixteen genuine requests. A small sample, not a universal claim. S4E.5 is preserved as an earlier DeepSeek-only pilot. See evidence, provenance, and model-identity limitations.
How BOUNDARY works
- 1Classifydeterministic proposal
- 2Guardpermission check
- 3Journalsimulated record
- 4Verifyindependent re-check
- Deterministic routing — exact matches and held cases resolve with zero model calls.
- Selective advisory inference — a model is consulted only for genuinely ambiguous cases, and its output is advisory (execution_authority = NONE).
- Independent verification — every journal record is re-checked against the source case before it is trusted.
Cases
Each case runs through the deterministic pipeline — classifier proposal, guard decision, and (where permitted) a simulated journal approval, independently verified. This interface is read-only.