ODVS-Bench - a public domain-pricing benchmark

This page shows the engine losing when it loses - and today it does, on at least one slice below. The WITNESSED table below is empty: no submission has been run against the held-out split yet. Every row in the SELF-REPORTED table, including this engine’s own, is scored on the PUBLIC split, whose answers are already public - a signed log there proves only the bytes a submitter chose to upload, never that the answer wasn’t copied from the key. Treat self-reported numbers accordingly.

The public split

621 rows: 218 recent held-out sales, 3 curated corpus holdout, 400 ledger sample.
Every row’s source is on record as redistribution-permitted in this repo’s provenance registry: com-sales, dnjournal, manual-seed, noncom-sales, open-dataset, parkio-results, registry-reports, repo-curated, sedo-manual, sedo-weekly, venue-annual-reports.
Split hash: 7bfb86fda1478ff509a90c4fca51311e190b85aeb8d611fd4f19bafc32679cf2. A submission log that does not carry this hash is not scored against the current split.
CC-BY-4.0 (rights basis: data/provenance/sources.registry.json, redistribution=permitted rows only)

Witnessed (hidden split)

Run by the maintainer or a steering-group member against a held-out split whose prices were never published anywhere. Only these rows are ever described as a model comparison.

No witnessed submission exists yet. The hidden split itself has not been assembled - it needs rights-cleared, never-published sale prices this repository does not hold.

Self-reported (public split)

Bring-your-own-key runs against the public split above, labelled self-reported because the split’s answers are public.

Submitter	Model	Coverage	Recent held-out sales within-2x	Curated corpus holdout within-2x	Ledger sample within-2x
audit.domains (this repo's own estimator)	champion (rev 4.11, generalization)	621/621	31.65% (Wilson 25.84% to 38.1%), n=218	0% (Wilson 0% to 56.15%), n=3	35.5% (Wilson 30.97% to 40.31%), n=400

audit.domains (this repo's own estimator): Self-reported: computed in this repo by the same code the row is about, on the public split above (3 corpus rows survive quarantine of fabricated ground truth - see FIELD_NOTES.md). ACCURACY.md's published headline is measured on the frozen 792, not this split, so the two are not expected to match exactly even where they overlap.

Submitting a result

Run tools/odvs-bench/run.ts against the public split with your own model endpoint and key - this project spends nothing running your submission.
Score the resulting log with tools/odvs-bench/score.ts.
Open a pull request adding the row and its log under selfReported in data/odvs-bench/leaderboard.json.