HELOC File Review
bundle · underwritingBundle
| Version | 0.11.0 → 0.12.0 |
| Runtime | smart-sdk v3 |
| Default model | opus-5 unchanged |
| Reasoning effort | xhigh |
| Repository | /repo/agents/underwriting |
| Signed by | s.mendo · Sep 3, 12:03 |
| Status | live in production · batch-0903-am |
Topology
The root delegates document extraction to a peer that returns typed values with page citations and never raw text, and step-boundary verification to a peer that sees only the worksheet. Neither peer can see what the other sees — that isolation is what makes the checker’s agreement meaningful.
Agents 3
Executes the locked eight-step procedure on one loan folder against one policy card and produces the code-validated final record. Delegates document extraction to document-reader and step-boundary verification to checker.
Reads one loan document in a fresh context and returns typed extractions with page citations — never the document text. Used for document-heavy steps so the primary context stays small.
Independent step-boundary checker. Sees only the worksheet and state files and the acceptance criteria it is given — never the loan documents or the transcript. Re-derives key numbers from recorded entries and flags anything a stranger could not re-derive.
Granted obligation_lines in 0.12.0. Without it the checker could not re-derive the qualifying payment, which is what “calculation mismatch” meant on eight of 150 fields.
Tools 29
Grouped by what they do. Nine of the 29 are income calculators — one per income type — which is why the raw count looks alarming and isn’t.
Skills 14, was 13
The step skills map one-to-one onto the locked procedure. If a step has no skill, the run has no instructions for it — so the mapping is rendered in full, including the gaps.
The manifest and staleness checklist, applicability gate, and cross-document conflict sweep.
Working the credit report against written policy: trusted-score placement, housing history, derogatory handling.
Classify each income source, route it to the right documents and computation, and verify continuance.
Assemble the debt picture from each debt’s true source and construct the qualifying payment.
~Edited: the checker is now handed the raw obligation lines alongside the recorded entries.
Collateral is consumed, not derived. Property value, LTV and CLTV, lien position from structured sources.
The flood determination read and the insurance conditions.
Title commitment read: vesting, exceptions, and lien position confirmation.
Derive the supportable line, choose the outcome, draft conditions and reasons.
On a counteroffer or decline, pull the Chase adverse-action reasons and match every reason to a recorded finding.
Prohibited-basis characteristics never enter reasoning; documented continuance analysis only.
The HELOC fraud-tell pattern library — income, credit history, title, and process-gaming signals.
How to read and apply the policy card grid; never improvise a threshold.
The completed-record exemplar — good worksheet entries, step statuses, and final record shape.
How to write a referral that routes a finding out of the review.
Policy cards
| dti-thresholds | 43% standard, 45% with reserves | unchanged since 0.9.2 |
| collateral | LTV and CLTV limits by lien position | 1 amendment proposed |
| adverse-action | Chase reason table, 2026 revision | unchanged |
| reserves | Months required by occupancy and score band | unchanged |
Thresholds live in policy cards, never in skill text. A skill that improvises a threshold is a defect regardless of whether the number is right.
Rollback
retired is not deletedRolling back returns 0.11.0 to live and retires 0.12.0. Decisions already stamped with 0.12.0 are left exactly as they are.
Audit chain
Every decision resolves down to a page in a document and up to the experiment that authorised the bundle which produced it. This chain is the reason the platform can be reviewed by someone who was not in the room.
Any cleared file resolves this way — the chains above are three of 1,004, and every cleared row in a batch links to its own.