BOUNDARY — Governed AI, Measured

BOUNDARY is a working web application for finance and procurement teams: a governed, deterministic-first approach to handling synthetic invoice and purchase-order exceptions. Principle: Use intelligence only when intelligence is required. Use autonomy only when authority exists. An independently defined purchase-order fixture and supplier evidence feed input validation, deterministic classification, an independent guard, a simulated SQLite journal, and read-only verification. Exact matches resolve with zero model calls; missing evidence and amount discrepancies result in HOLD. Where interpretation is useful, AI advice is constrained and never grants payment authority (execution_authority = NONE). All data and actions are synthetic. Try BOUNDARY lets judges choose a scenario, edit invoice values, and inspect a newly calculated decision and evidence trace. The six-invoice batch and four-path comparison are explicitly labeled offline simulations. Public visitors cannot trigger paid inference. The AMD evidence is real and separately recorded: DeepSeek-V4-Flash and Google Gemma-4-31B-IT were tested through AMD Token Factory. Across 16 genuine requests on a limited synthetic workload, LLM-first used 12 calls/2,930 tokens versus BOUNDARY's 4 calls/1,010 tokens—66.7% fewer calls and ~65.5% fewer tokens. These are workload-specific inference reductions, not universal cost or accuracy claims. Serving-model identities remain unverified. A Gemma-only code-fence fix preceded an outage; Gemma was restored October 16. Evolus extraction and DeepSeek inference succeeded in an earlier partial integrated trial. On October 16, a separate live trial completed both Gemma paths using the same validated Evolus evidence, with source-consistent advice and execution_authority=NONE. The earlier trial remains partial. Built with FastAPI, SQLite, Docker, and Railway, BOUNDARY turns evidence, AI interpretation, and operational authority into independently inspectable boundaries.