Falsifiable Release Gates for Self-Improving Systems

This collection of arXiv papers (July 16, 2026) documents both emergent failure modes and mitigation strategies for agentic LLM tooling and self-improving runtimes. Key findings: (1) Falsifiable release gates show a practical method to require machine-verifiable acceptance suites and preserve standing invariants for self‑improving agent runtimes (mitigation pattern). (2) A compaction failure in an agentic coding tool (Claude Code) records partial, timed‑out command output as durable, propagating false positives across sessions and models — an integrity/persistence conflation that can poison ML