Add title review, and show the checker the obligations
0.11.1 → 0.12.0| Status | kept · in production |
| Procedure touched | Steps 4 and 7 |
| Tested against | 150-field standard set · 12 of 12 runs, strict |
| Evidence | 96.4% ±1.2 against 92.4% ±2 baseline |
| Spent | $26.28 · 41m |
| Owner | s.mendo |
What changed in the procedure
before and after, by stepRead the title commitment for vesting, exceptions, and lien position.
Assemble every debt from its true source and build the qualifying payment.
What it did to the queue
the reviewer’s side of the loopClosed the calculation-mismatch stops at step 4 — debt-to-income failures fell from 8 in 150 to 1. Autonomy rose from 79% to 86% on the first batch. It did nothing for conflicting extractions, which are still 2 of the 7 files waiting.
Thread
verdict computed as keepDebt-to-income keeps failing for “calculation mismatch”. The checker can only see the total the underwriter wrote down, never the individual obligations behind it, so it cannot rebuild the qualifying payment the way a person would. Give it those lines. And add the title step, which has been missing since the baseline.
Two changes. Step 7 now reads the title commitment, which no version has done since the baseline — that is the gap behind the failed build on 0.11.1. At step 4 the checker is handed each obligation line alongside the recorded total.
Graded 12 runs against the 150-field standard set, strict. 96.4% ±1.2 against a 92.4% ±2.0 baseline — the intervals separate, which is the first real movement since the baseline. Debt-to-income failures fell from 8 in 150 to 1, and the one that remains is a genuine document conflict rather than arithmetic.
It costs seven cents more per file, all of it the extra context the checker now reads.
has it held up on live files?
Partly. Autonomy went 79% to 86% on the first batch, the largest single move so far. But blind review of files that cleared on their own reads 95.1% against the 96.4% graded — expected on live files, and worth watching rather than announcing.
Of the 7 files still waiting, 2 are conflicting extractions, which this attempt did not address.