Results

Joint first place: Polymanes and FinVeritas. Each answered 665 of the 680 held-out cases correctly (ACC 97.79%), and each is wrong on the same number of cases under every rule. The table lists Polymanes above FinVeritas under the ordering this board has applied throughout the Cup — accuracy, then calculation error rate. That settles the listing, not the result.

Ranked by ACC, then by calculation, extraction and structural error rate in that order, then by team name — the ordering this board has applied to the development phase throughout the Cup. Teams on an identical accuracy are joint at that rank and share a number; the order printed within such a group comes from the calculation error rate, which decides how the rows are listed and not how the teams placed. 5 groups of teams share an accuracy to the second decimal, including first place.

Every team is scored on its last valid upload before the deadline, not its best: the test phase returned an acceptance receipt and no score, so the held-out cases could not be used to choose among a team’s own candidates. 79 uploads were accepted from 23 teams, and none arrived late. Every submission parsed; no team lost a case to format.

ACC within each rule. The three rules are not equally represented, so these columns do not average to the headline. DQC_US_0117 is where the field separates: 54 cases spanning 98.15% down to 5.56%. DQC_US_0126 is 464 of the 680 cases and near-perfect for almost everyone, so it moves the headline very little. DQC_US_0015 is the rule whose answer is always the negation of the extracted value; no team exceeds 91.98% on it, and every team scores lower on it than on DQC_US_0126 — the reported value and its scale, not the rule, were the hard part of this task.

Rows marked Merged are one system that submitted from more than one account, ranked once. Polymanes uploaded from two accounts under the same name; DOVE uploaded as 666, 668, DOVE, QuantCat. Every upload involved carried the same answers, so merging changed no score. A team affected by a merge that believes it is wrong should contact the organizers.

Scored by openai/gpt-5-mini, prompt finmr-judge-v1, evaluation version 1.1, over all 680 held-out cases, under the same frozen configuration the development board uses. Gold file SHA-256 bd036288f315ed439a1235a526dc01d2ccd86b7cc4e28c9168ef681f35c53218. Submissions closed 2026-10-25 23:59 AoE.

The board above ranks by the official judge, while most of these rows are measured with the deterministic one — which is why the rule-based baseline appears twice with different figures. The official judge accepts 7 points more of the same extractions, 3.97% extraction errors against the deterministic judge’s 10.88%, because it reads values that mean the same number but are written differently. Compare like with like: a deterministic figure is a lower bound on the official one.

Reproducible from the starter kit: baselines/rule_baseline.py, with --mode extract and --mode negate for the other two. None of them is told which data-quality rule a case belongs to; the rule-based baseline picks its method from where the concept sits in the calculation linkbase.

Read the sign-flip row as a floor, not a method. It negates whatever it extracted and reasons about nothing, so a submission scoring near it has learned nothing whatever its rank says. On development the relationships earn their keep — 25.00% against the shortcut’s 15.29% — but on practice they do not, because a third of that set is the one rule whose answer is always the negation. Both numbers are published as measured.

The do-nothing row’s 4.52% calculation-error share is the 15 practice cases whose reported value genuinely is zero: answering "0" clears the extraction gate for those and fails at the calculation step. Every other case fails earlier.