遇见数据集

Glass Ionomer (formerly VeriFlow) — Measurement Record 2026-09-03: GCAB Blinded Holdouts and Frontier-Model As-Filed-Figure Assertions

收藏
Zenodo2026-09-02 更新2026-10-01 收录
官方服务:

资源简介:

Frozen measurement record for Glass Ionomer (formerly VeriFlow), a claim-assurance and release-gating engine for AI outputs (github.com/AMRISHERE-DEV/Veriflow, engine v1.1.1). Contents: (1) General Claim Assurance Benchmark (GCAB): 18-case development pilot, blinded holdout-1 (25 cases) and blinded holdout-2 (25 cases), with SHA-256-pinned frozen harnesses, evaluation policies v1 and v2, sealed-constructor inputs and gold labels, scored results and the complete run log (failures preserved unrepaired). (2) The SEC as-filed-figure measurement of frontier language models: protocol v1.1 frozen before any query plus a v1.2 addendum recording every departure; run 1 archived; run 2 = 306 questions x 4 models = 1,224 answers, dual-graded offline under engine v1.0.0 and v1.1.1 (zero bucket differences), with a per-row receipt for every answer and Wilson 95% intervals on every rate. (3) Frozen public-domain corpora with hash manifests: SEC EDGAR companyfacts and filings, ClinicalTrials.gov records, DailyMed SPL labels. SHA256SUMS.txt covers every file. The record corresponds to git commit b8326b5, tag record-2026-09-03, of the measurement repository. Scope note: all cases were author-constructed and author-run with procedural (agent-sealed) blinding; no external party has audited this record. Supplements the whitepaper DOI 10.5281/zenodo.22233268.

提供机构:
Zenodo
创建时间:
2026-09-02
二维码
社区交流群
二维码
科研交流群
商业服务