235 deals. A model read them. Code checks every claim it made.
This runs the repo's verifier, consistency checks and scorer in Python in your browser, over the claim ledgers Claude Sonnet 5.5 returned in the recorded run of 2026-10-17. Browse any deal, see a receipt for every claim, then try to fool it: invent a quote, cite the rep as the buyer, or change a number, and watch what the checks do.
Downloads about 13 MB on first run: Python 3.13 and pandas from jsDelivr (cached by your browser after that), plus about 0.3 MB of data and code from this site. No model is called.
- The call notes and contacts are synthetic, generated for the 228 open deals of one sales team in Maven Analytics' public-domain CRM data. Deal values, dates (shifted) and real outcomes come from that data; 18 anomaly types were planted, including 7 duplicate deals.
- Sources, in the repo at 617c985: README.md (228 open deals); data/company/deals.csv (235 rows: those 228 plus 7 planted duplicates); anomalies.csv (the 18 types, duplicates as S6). The Scoreboard tab recounts them once it runs.
- The hackathon (Block Convey, Money Talks 2) result is still pending.