Architecture decisions

@@ -24,7 +24,7 @@ Architecture decisions are recorded commitments — choices made deliberately, f
These three decisions define where automated processing stops and human authority begins. They are the decisions an auditor, a procurement evaluator, or a technical due-diligence reviewer should read first.

**ADR-07 — Data must be classified before the AI boundary touches it.**
Every file that enters the platform is classified by type before any processing occurs. Structured data — spreadsheets, ledger entries, database exports, signed documents — routes to deterministic parsing engines that produce verifiable, repeatable output. Unstructured human text — correspondence, meeting notes, narrative reports — routes to the language model boundary. Nothing crosses the boundary unclassified. This prevents the platform from applying AI reasoning to records that should have deterministic, auditable treatment.
Every file that enters the platform is classified by type before any processing occurs. Structured data — spreadsheets, ledger entries, database exports, signed documents — routes to deterministic parsing engines that produce verifiable, repeatable output. Unstructured human text — correspondence, meeting notes, narrative reports — is ingested through the same zero-AI, deterministic path: no content classification, no entity extraction, and no language-model call of any kind runs inline at ingest. Any AI reasoning applied to that text happens only downstream, reading finished records back out of the ledger — never during the write itself. Nothing crosses the boundary unclassified, and nothing at the boundary itself is AI-derived. This prevents the platform from applying AI reasoning to records that should have deterministic, auditable treatment.

**ADR-10 — Every base-asset ledger entry requires a human checkpoint.**
The platform's [[service-input|F12 input interface]] is the mandatory human-in-the-loop checkpoint for all base-asset ledger commits. No automated process can write a verified record to long-term storage without a human confirming the routing decision at the F12 gate. The design mirrors the control structure of a regulated clearing house: the platform can process, classify, and propose — but the commit belongs to an authorised human.