cove
Agentic Chain-of-Verification (CoVe 2.0)
Verify and refine the most recent response (or a /cove <text> argument) through a
three-phase, open-book self-verification pipeline. This upgrades the original
closed-book CoVe by grafting CRITIC's tool-interactive critiquing onto it — exactly the
extension the CoVe paper proposes in its own conclusion: "equip CoVe with tool-use,
e.g., to use retrieval augmentation in the verification execution step."
Why open-book
The CoVe paper verifies claims using only the model's own knowledge (closed-book). The CRITIC paper finds that "exclusive reliance on self-correction without external feedback may yield modest improvements or even deteriorate performance." So CoVe 2.0 verifies high-risk claims against external evidence (web search) and reserves confident rewriting for claims that evidence actually grounds.