Real recorded sales from Indiana Sales Disclosure Forms in the counties on the Michigan line, dated on or after 2026-06-01. Gold labels are the county assessor's own validation decision and the recorded price. A model trained before the freeze date cannot have seen these answers.
Composite = 40% subject-validity skill · 30% comp-validity skill · 30% valuation. Skill is Youden's J (true-positive rate + true-negative rate − 1): calling every sale valid, or every sale invalid, scores 0. Raw accuracy is shown beside it so you can see why. Valuation is scored only on sales the assessor ruled valid, the only ones whose price is ground truth.
| # | Solver | Composite | Validity skill | Raw accuracy | Comp skill | Valuation | Median error | Parse |
|---|---|---|---|---|---|---|---|---|
| 01 | baseline:rulesbaseline | 24.0 | 27.4 | 51.3 | 17.0 | 26.6 | 25.1% | 100% |
| 02 | baseline:majoritybaseline | 10.0 | 0.0 | 69.3 | 0.0 | 33.3 | 20.7% | 100% |
| 03 | baseline:naivebaseline | 0.0 | 0.0 | 30.7 | 0.0 | 0.0 | 9526.2% | 100% |
The sale as filed with the state: form field names, the price in the state's implied-cents layout, assessed values, checkboxes, buyer and seller mailing city and state, and up to six earlier sales from the same assessor neighborhood. Names are replaced token-by-token with keyed pseudonyms (a shared surname stays shared, so related-party signals survive), street lines are dropped, and the assessor's verdict, reason text, and validator name are removed. The underlying records are public; nothing here is secret, and anyone can rebuild the split from the source file.
One split, 150 items, freeze 2026-06-01. Of the counties requested, Elkhart and Steuben had no rows in this release, so items come from LaPorte, St. Joseph and LaGrange. Assessor validation is a human judgment, not ground truth in the absolute sense; it is the label the county actually used. The rules baseline reads only form checkboxes. Rebuilds after a new freeze date produce a new split with a new hash; old scores stay on the old split.
2026-06-012026-08-21Elkhart, LaGrange, LaPorte, St. Joseph, Steuben1500.307https://www.stats.indiana.edu/topic/sdf.aspSDF_2026_20260922.zip9ba9f8eb6ff73161b03f393768dd24906fc02c481e07a562a1e1129065d4f8499aa0b50c8df06cf9b3e7878afcb6d0b37938c3f367efa66bba80a04a309851e4689bce11f454af0947b1b3335a8548e743d0b821cf4eb1fb64a81cbeed552e790008f25e1b0414c20f51886f7605454e4bce7245f4a118af385970beef4e2b3b454python3 -m prbench.live verify # recompute every hash + re-score every stored response