Agent Regression Protocol

Mandates that before modifying the theorem router, every Agent must run three layers of tests: Layer A (the current 12 positive fixtures), Layer B (the 13 historical regression curves only present in the old version), and Layer C (the 4 official discrepancy curves)—these three layers exactly correspond to the corpora established in Documents 08 and 09. The output format is unified into a single JSON containing the fields curve, decision (tri-state PASS/FAIL/UNKNOWN, not binary), first_failure, all_failures, evidence, code_version, and semantic_version. The core of the document is four "cognitive firewalls", directly naming the four most easily committed conceptual bait-and-switches: analytic Sha = 1 does not equal Sha is trivial; rank = 0 does not automatically equal rigorous analytic rank 0; the 2-descent dimension being numerically equal to the analytic valuation does not automatically mean $\operatorname{BSD}(E,2)$ is proven; timeouts must be marked UNKNOWN. The final discipline is equally important: if a new version only improves runtime, cache, batching, or output formatting, mark it ENGINEERING ONLY and do not count it as BSD mathematical progress—this rule is precisely to prevent packaging pure performance optimizations as research breakthroughs.