How verification works

Definite separates doing the work from deciding whether the work is right. Agents produce; the deterministic engine verifies; humans resolve.

Verdicts

Every check resolves to exactly one of three verdicts, emitted only by the deterministic, versioned rules engine:

• PASS — the rule held against the bound inputs.
• FAIL — the rule was evaluated and did not hold.
• CANNOT_VERIFY — the engine could not establish the result. Missing, late, disputed, or low-confidence input always lands here.

Fail-closed is a hard invariant. There is no assumed pass, no materiality threshold that clears a CANNOT_VERIFY, and no model output that upgrades a verdict. Identical inputs plus the same rule versions always reproduce the same verdict.

What agents may and may not do

Agents map source data to concepts, investigate exceptions, draft corrections and explanations, and assemble artifacts. Their classifications carry confidence and provenance, and anything an agent produces is either deterministically re-checked or routed to a human. Agents never emit verdicts, never gate an outcome alone, and their event telemetry records which model acted, when, and on what — check-stage events can never carry a model identity.

Receipts

Each check writes a receipt: the rule ID, version, and citation; the verdict; typed references to the inputs (opaque IDs plus sha256: content hashes — never raw values); the previous receipt hash, forming a per-tenant chain; and a signature over the exact payload bytes. Receipts are append-only and independently verifiable with Definite's published keys, with no dependency on Definite staying available. See security and independent verification for the byte-level procedure.

Behind receipts sits a per-tenant Merkle audit ledger with signed checkpoints, inclusion and consistency proofs, and externally witnessed checkpoint exports, served by the ledger and proof APIs.

Evidence and lineage

Results stay traceable to their sources. Run inputs are immutable snapshots with content hashes; artifacts expose lineage from each result back through transforms to source records with locators (sheet, cell, row, field, or page), observation times, and snapshot hashes. The console's Evidence surfaces render these traces, and control-testing workpapers annotate the underlying evidence line by line.

Exceptions and human review

FAIL and CANNOT_VERIFY findings open exceptions. An agent classifies the root cause — timing, misclassification, or a genuine error — gathers the relevant records, and drafts a correcting entry or explanation. A human reviews, decides, and re-runs. Reviews are recorded with a named, authenticated actor; they never rewrite the original verdict, and anything reviewer- or auditor-facing requires a human to adopt it.