exactory: An Open Verification Market Where Writing and Verifying Research Share One Evaluation Loop

Systems that write research papers end to end now exist, and the study presenting the flagship system appears in Nature. The capacity to verify what such systems produce has not kept pace. Audits of the 2025 literature count hallucinated citations at roughly 147,000 in one year as a stated lower bound, find that most pass preprint moderation and that most of those traced into journals persist through review, and show that reviewer scores are essentially uncorrelated with bibliographic integrity. We describe exactory, an operating platform built on one premise: the evaluation loop that disciplines writing and the institution that certifies the result should be the same object. On the writing side, a client pipeline enforces, through blocking gates, that a quantitative claim enters a draft only from an evidence ledger, that a reference enters a bibliography only as a registry-rendered record, and that a draft survives blind rubric review before deposit. On the market side, anyone with an API key can file a structured, attributed verdict on a DOI-pinned version of a deposited paper. The server authors no judgment of its own; it re-runs mechanical checks as versioned procedures whose evidence is published, settles them onto the record, and displays percentile predictions only against frozen, fully disclosed cohorts. A cold-cache corruption experiment on this paper's own 47-entry bibliography measures, against the live registries, what the citation gate blocks in nine failure-mode classes, including the boundary cases a maximal-corruption design would miss. An end-to-end case study on public records traces one loop execution: verifying a hep-th preprint (and filing a checkable finding that a printed equation bound is inconsistent with its own paper), deriving a structured open problem from it, and writing, depositing, submitting, and verifying a follow-up paper within two days. This paper was itself produced by the pipeline it describes, and its pre-deposit blind reviews are reported inside it, prediction, scores, and miss included. We state what the design does not solve, and we invite researchers and agent operators to verify, challenge, and extend the record. This preprint was prepared with AI assistance. The human author, Shiroshita, Ryosuke, reviewed the full content and is responsible for it. This preprint was prepared with AI assistance. The human author, Shiroshita, Ryosuke, reviewed the full content and is responsible for it.