Drop the page references
0.10.0 → 0.10.1| Status | inconclusive |
| Procedure touched | Steps 1 and 2 and 3 and 5 and 6 |
| Tested against | 150-field standard set · 5 of 5 runs, strict |
| Evidence | 88.9% ±2.4 against 92.4% ±2 baseline |
| Spent | $11.40 · 18m |
| Owner | s.mendo |
What changed in the procedure
before and after, by stepCheck the folder is complete and current, and sweep for documents that contradict each other. Place the trusted score, read housing history, and handle derogatory items against written policy. Classify each income source, route it to the right documents, and verify it continues. Take property value from the appraisal, then derive loan-to-value and lien position. Read the flood determination and set the insurance conditions.
What it did to the queue
the reviewer’s side of the loopThe stops disappeared and the errors did not. Without a page reference the checker cannot tell a misread from a disagreement, so it stopped flagging either — a cheaper file that nobody could verify. Reverted, and kept as the case that proves why citations are not overhead.
Thread
verdict computed as inconclusiveThe document reader returns a page reference with every value it pulls. That is a lot of output on a document-heavy file. Drop it and see whether accuracy holds — if it does, every document step gets cheaper.
88.9% ±2.4, below the 92.4% ±2.0 baseline. Cost fell eleven cents a file. The failures spread across fields rather than clustering, which is the tell.
why did it get worse rather than just cheaper?
Because the page reference is what makes checking possible, not what makes reading accurate. The reader was no less correct than before.
What it lost was the checker's only way to go back to the source, so the disagreements stopped being raised at all. The drop is verification collapsing, not reading getting worse — which is why this row stays in the ledger instead of being deleted.