Engineering Reliability into AI Agent Code Generation. Part II

Automated excerpt

The evidence model: no claim without an artifact Part I established that agent status reports are claims (P2: 75. 8% false-success rates for self-assessing coding agents). The evidence model is the systematic reply: for every kind of claim, define the artifact that proves it, and refuse the claim without the artifact. e1["build. log + exit code == 0"] e2["test. log + exit code == 0"] C5 — The artifact taxonomy. The verifier admits artifacts, never assertions — from any model, including the evaluator.

Selected automatically from source text; not independently written or fact-checked. Read the original for full context.