Request an architecture briefing

Why the record has to be the product

The sentence everything hangs on

Banks are sitting on the largest labor-automation opportunity in their history and cannot say yes to it — because every control they have assumes the worker is a human. Maker-checker, QA sampling, access reviews, change boards: each one breaks when the worker is an AI agent. So risk and compliance veto production deployment, and pilots die in pilot purgatory.

Nobody sells governance of actions — the layer that scopes, gates, evidences, and attests an agent's individual act.

The existing "AI governance" category sells governance of models: inventories, policy documents — paper about the AI. The missing product is a control primitive for the work itself. Wet Ink does not sell agents; agent vendors are a channel, not competition. It is the harness agents must wear — the thing that converts the risk officer's "no" into "yes, under these controls."

The upgrade hiding inside the constraint

Human review forced sampling: QC departments review a fraction of files because reperformance by hand costs too much. A governed machine reviewer changes the economics, and with them the control regime itself — structured rationale on the full population instead of a sample, each file with its own evidence pack. Better assurance and lower cost are the same product here, not a trade-off.

Why the record, specifically

A quality-control department's output is an evidence pack. That makes home-lending review the rare domain where the governance layer is not overhead — the audit trail is the deliverable itself. So the product is not the automation. It is the proof that the automation was accountable: who judged, against which version of which rule, citing which page of which document, at what cost, and whether anyone could have quietly changed the answer afterwards.

Prompts are guardrails for efficiency; the database is the guardrail for integrity.

The kernel treats the model as an untrusted client. Nothing it believes changes what the system of record accepts. The schema enforces separation of duties — the one control every examiner reaches for first: the maker who does the work physically cannot seal it. That is a property of the database, not a policy anyone follows.

The wet-ink moment

The wet signature is banking's most human artifact — the authoritative original, the moment a person put their name to a judgment. Wet Ink takes its name from the layer that keeps that moment in an automated pipeline: every AI action gets its wet-ink moment — a distinct, accountable approval, recorded in a form that outlives the meeting where it happened.

What is honestly true today

The sealing architecture is real — 85 passing proofs stand behind it. The AI judgment layer that will run inside it does not yet exist, and this site says so everywhere the question comes up, in one canonical place:see what runs today. A verification company that inflates its own claims has already failed its own test — claim calibration is not a marketing constraint here; it is the thesis, demonstrated.

Fannie Mae Selling Guide D1-3-01 (the 10% post-closing QC sampling floor; verified 2026-08-17)Fannie Mae Lender Letter LL-2026-04 (AI/ML governance framework, issued 2026-04-08, effective 2026-08-06; see the explainer for verified sources)The Wet Ink repository's own stated proof counts (85 = 33 + 40 + 12)