EvidenceDesk

A claim-review workspace built with Codex: a technical writer puts a draft beside its sources, and keyword and rule checks flag claims the sources don't back. By Benjamin Nweke Project page: https://towardsdatascience.com/shipai/projects/evidence-desk ShipAI is a Towards Data Science project. You can learn more about it here: https://towardsdatascience.com/shipai/about Key takeaways. 1. It flags metric confusion, like a draft claiming a 20% accuracy gain where the source reports 20% less error. 2. No language model decides: shared keywords retrieve up to three passages per sentence, then targeted rules check the wording. 3. A false mismatch on the builder's own published article led to a tighter rule: one percentage, closely matching context. Chapters: 0:00 When a paraphrase changes meaning 1:08 EvidenceDesk, built with Codex 1:34 Draft left, sources right 2:45 Accuracy versus error rate 3:41 Recording an editorial decision 5:34 What the supported label means 7:36 Under the hood 8:59 Checking the checker