# Every decision, against a documented rule, cited to its source documents

*HOW IT WORKS*

> Most AI in this category asks you to trust an output. Sei is built so you do not have to: the rules are yours and written down, our triple-check layer stands between the regulation and the decision, and everything the agent concluded stays on the record with the evidence attached.

## Trigger Points move the loan, not manual clicks

A decision starts because the loan moved — a document arrived, a payment was missed, a fee changed, a clock began. Every AI agent runs with a goal and doesn’t stop until the goal is reached or no longer valid

- **[The Loan Graph](/platform/loan-graph)** — One live model of the loan that every desk reads and writes. Shared state enables airtight decisions, every step along the way.
- **[The Orchestrator](/platform/umos#orchestrator)** — Holds the loan-level goal, decomposes it into desk-level goals, resolves what blocks what, runs what can run now, parks what is blocked and wakes it when the block clears — and escalates to a human when a desk hits its confidence floor.

## The Rulebook

The rule set an agent reasons against. Yours, not a generic model’s idea of mortgage lending.

### Agency and investor guidelines

- Fannie Mae Selling Guide
- Freddie Mac Seller/Servicer Guide
- FHA Handbook 4000.1
- Your investor overlays

### Consumer protection and lending regulation

- TILA, RESPA and TRID
- ECOA and Fair Housing
- HMDA
- FDCPA, TCPA and UDAAP

### Your own operating rules

- Underwriting guidelines and credit policy
- Servicing policies and SOPs
- QA rubrics, including the edge cases that live in reviewers’ heads
- Call scripts and disclosure requirements

## Triple Check

Three checks stand between a regulation and a decision. Each one can stop the work, and each one is logged.

### 1. The agent reasons against the Rulebook

Not against a model’s memory of mortgage lending. The agent works from the encoded rule set — the agency handbooks, your investor overlays, and your own SOPs — for the step it is performing.

_On failure:_ A file that does not satisfy the rule is not cleared. It becomes a condition, written in plain language, with the rule it failed named.

### 2. An independent checker validates the output

A second, separate model re-checks the result against the same rules before it leaves the system. Different maker and checker.

_On failure:_ Disagreement between the two is not resolved quietly. The item is held and surfaced rather than shipped on the first answer.

### 3. The result is confidence-scored

Every finding carries a confidence score. Anything above your threshold clears automatically; anything under it routes to your own underwriter as a named exception, with the reasoning and the source attached.

_On failure:_ Under threshold is not a failure state, it is the designed one. Your reviewers spend their time on the exceptions instead of on the whole file.

## The Evidence Trail

Every decision carries the reasoning that produced it and a citation to where the evidence came from.

### Cited to the source, not summarized

A finding on a loan file points at the document and the page it came from. A finding on a call points at the timestamp. A reviewer checks the citation rather than taking the answer on trust.

### The reasoning is kept, not just the outcome

Every agent decision is logged with its full reasoning trace, providing full transparency.

### Built to be handed over

Records are immutable and retained to your policy, and the logs export in a format that is examination-ready.

### It covers the conversation too

Underwriting, closing and QC leave a trail, and so does every borrower call, email and chat — scored against your rulebook with the moment cited.

## When it is unsure

Anything below your confidence threshold routes to your human-in-the-loop as a named exception.

The reviewer sees what the agent concluded, the rule it was working from, the confidence score and the source document, and makes the call themselves.

---

_Source: [https://seiright.com/how-it-works](https://seiright.com/how-it-works) · Sei AI_