Attempts
Each one is a hypothesis about the eight-step review, the set it was tested against, and what it did to the queue. 4 of 6 reached a verdict; 2 are still open.
A hypothesis with a set to test it against. Nothing spent yet.
Running. No verdict until every run is in.
Graded, but the interval overlaps. The change may have worked; the evidence cannot say.
Raised the underlying model and how hard it reasons. No change to the procedure itself.
Step 4 was assembling the debt picture and building the qualifying payment in one pass. Split so each half can be checked on its own.
Removed the page reference the document reader returned with every value, to see whether accuracy held. It did not.
Separated above baseline and promoted.
Separated below, and reverted. These rows are the most useful in the ledger.
None
Two of the six reached a verdict. That is the normal shape of this work — an attempt that comes back inconclusive has still told you the change was not worth the runs, and the discarded column is where the most transferable lessons sit.
The bars are calendar days from drafting to a decision, not machine time — the runs themselves take between eighteen and forty-one minutes. What takes days is deciding what to test and reading the result honestly, which is the part no extra compute shortens.
| attempt | procedure | status | evidence | spent | bundle |
|---|---|---|---|---|---|
| Add title review, and show the checker the obligations | Steps 4 and 7 | kept | 96.4% ±1.2 | $26.28 | 0.12.0 |
| Widen what the checker can see | Steps 5 and 6 | grading 8 of 12 | 96.9% ±1.6 | $17.44 | 0.12.1-draft |
| Move to a newer model | No procedure change | inconclusive | 93.8% ±1.9 | $10.60 | 0.11.0 |
| Split obligations from the qualifying payment | Step 4 · Obligations and DTI | inconclusive | 93.1% ±2.1 | $11.55 | 0.10.0 |
| Drop the page references | Steps 1 and 2 and 3 and 5 and 6 | inconclusive | 88.9% ±2.4 | $11.40 | 0.10.1 |
| Prefer the lender statement on small disagreements | Step 4 · Obligations and DTI | drafted | — | — | 0.12.1-draft |
Verdicts are computed from the interval, never entered. An attempt whose band overlaps the baseline reads inconclusive however much the point estimate moved.