Chain-of-Proof: A framework for verifiable analysis
The CoE is a conceptual framework that defines what makes a analysis artifact reliable, a lot as ACID defines what makes a database transaction dependable. Reasonably than prescribing the best way to construct a analysis agent, the framework specifies the properties its outputs should have. It follows a single precept with two halves: each declare in a analysis artifact should carry a recorded proof chain (completeness), and every chain should genuinely assist the declare it’s hooked up to (correctness). A declare could also be a reference, a reported quantity, a technique description, or a conclusion, that should hyperlink again to corresponding proof, akin to a peer-reviewed paper, an experimental log line, the code that really ran, or the outcomes desk.
A hallucinated reference factors to a paper that doesn’t exist. An unreproducible rating doesn’t reappear when the code is re-run. A misdescribed technique claims one algorithm within the paper whereas the code implements one other. Every is a declare whose chain again to its proof is damaged; the CoE Audit makes these breaks measurable.
