DappGo

Market Claim Verifier checks extracted financial claims against SEC filings and returns a corrected draft with citations. Headline result: Baseten's corrected-gold held-out text split reached 90.2% (240/266), with false support at 1/139 among non-supported claims; benchmark cost was ~$0.02 per claim and the live-text planning band is 6–25 s per Baseten check. The problem is one we hit in our own publishing at DappGo: a post can quote a real number with the wrong period, unit or comparison. In the retained Apple example, “Apple pulled in $416.2 billion in fiscal Q4 2025” uses the annual figure; the quarter was $102.5 billion. A fluent summary does not reliably catch that, and numbers like this move money. How it works: Meta's Muse Spark reads the post, extracts financial assertions and selects an SEC tool such as lookup_fact, compute_change, history_rank or list_periods. The tools return XBRL facts with evidence IDs, units, periods and filing links. Python, not the model, binds each claim to its evidence, performs the arithmetic and returns supported, contradicted or insufficient_evidence. A correction changes only a checked assertion with a program-bound value and citation. The model proposes; the program decides. For context, the same held-out split's no-tool Baseten baseline was 21.1% (56/266); the all-claim baseline was 20.7% (74/357). The live site also wires Fireworks and the Meta Model API. The CLI accepts text, images, native-text PDFs, audio and video; Meta OCR/ASR turns media into claim text, while financial verdicts still require SEC evidence. What you can try: hackathon.dappgo.com offers 3 stories × 5 packaged inputs, saved results with no password, and live checks on Baseten, Fireworks or the Meta Model API with the team name as the judge password. Public history reopens full results. The MIT repository includes the CLI, a frozen evaluation set, per-claim traces, and an offline verifier that rebuilds committed outputs. Nothing in this repository pre-existed the hackathon; existing DappGo product code is outside this submission. Limits, stated plainly: benchmark accuracy is not accuracy on every post; real-world coverage is reported separately; two gold labels were corrected after the run and both scorings are published; the primary scope is US-GAAP SEC XBRL filers; and extraction can miss claims.