Verification infrastructure for home lending
The proof trail is the product.
The kernel seals every judgment it records — who judged, against which version of which rule, citing which page of which document, at what cost — into a record no one, including us, can quietly change.
checker — distinct from the maker, enforced by the schemathe exact version live at judgment time, captured in the sealdocument and page — every extraction must say where it readderived by the kernel from token counts — never self-reportedappended to a tamper-evident chainprior-entry hash carried forward — append-only
Exact numbers
Product status
What runs today
Built and proven now
- An append-only, tamper-evident chain holds every sealed judgment.
- The schema separates the actor who does the work from the actor who approves it — a maker physically cannot seal its own work.
- Every judgment carries its provenance: who judged, which rule version, which document and page, at what cost.
85passing proofs hold all of it.
Not yet built
The AI judgment layer. The kernel stores predicates but does not yet evaluate them; no model call leaves the process; no document lifecycle exists. No loan file is under review today, and there are no customer deployments. We say so plainly, because a verification company that inflates its own claims has already failed its own test.
finding recorded reviewer A insufficient evidence
superseding pass reviewer A not accepted — same actor as the original finding
superseding pass reviewer A not accepted — same actor as the original finding
seal blocked until a distinct reviewer resolves the findingThe mechanism
Five steps, one record
Claim
A named actor claims the file. From that moment, the record attributes everything done to it.
Evidence
Every extraction must say which document and page it read, and which model run produced it. The schema refuses unanchored evidence.
Checks
Deterministic checks run against versioned rules. The database refuses a verdict citing a rule version that does not exist.
Gate
The database refuses a seal from the actor who did the work. Approval requires a second, distinct actor — a schema property, not a policy.
Seal
The seal appends the result to a tamper-evident chain and captures exactly which rule versions were live. The chain catches an edited history at the exact position.
Interface preview
The console a sealed record gets
Every Wet Ink deployment inherits the same console design language: a review queue on the left, the sealed record on the right, and the schema's refusals rendered as first-class states — never hidden. What follows is a design preview, not a product capture.
sealed record
maker + distinct checkercaptured at judgment timedocument and page, every extractionprior-entry hash carried forward — append-only
Illustrative schema states from the sealing architecture that runs today; the review layer that would fill this queue does not yet exist — see what runs today.
Applications
Where it applies in home lending
Loan file review runs on the same judgment shape at every stop in the mortgage pipeline — first-mortgage purchase, refinance, HELOC, and home equity files alike. Six places it decides money:
Origination document review
The same judgment shape at the front of the pipeline: documents in, rules applied, every conclusion anchored to its source page.
Underwriting conditions
Condition cleared, by whom, against which guideline version — attested, not just marked done.
Closing and funding checks
The last look before money moves: settlement documents, title commitment requirements, escrow figures — sealed with the evidence behind them.
Post-close QC
The worked example: full-population reperformance instead of the 10% sample.
Servicing transfers
What the boarding team reviewed, under which rules — in a form the next holder can check.
Audit and disclosure response
When the examiner asks what your AI did, the answer is a record built to be checked — not a memo.
The economics
Full population, not a sample
Lenders sample because human review is expensive. Fannie Mae's Selling Guide D1-3-01 sets the floor: a minimum of 10% of the loans a lender originates or acquires, selected randomly, for post-closing QC review. The incumbent category — mortgage QC software and outsourced audit services — manages that sample. Wet Ink built its architecture for the other 90% too: full-population reperformance, every file with its own evidence pack. That is a design basis, not a delivered customer result — the review layer that will produce it does not yet exist.
Dated pieces
Writing
Disclosure on demand: what LL-2026-04 actually requires
Fannie Mae's AI governance framework took effect on 2026-08-06. What it obligates, and what an answer built to be checked looks like.
Keys to the vault
Everyone asks whether the model can be trusted. Wrong question. Trust is a property of the access, not the worker. Scope the key.
Keys to the walk-in
One platform credential, every client site. Fifteen years of kitchens said locks are the architecture — the tool inherits the human.