Parallel structured LLM answers never check each other: the Zhaozhou MU problem
Successful-Farm5339 · reddit · 2026-09-20
The author uses the Zhaozhou dog koan to expose a real gap in structured LLM output: questions in one request are evaluated independently in parallel, so nothing checks that answers agree. A credit note can be returned alongside auto-approval—both well-typed, neither a hallucination, yet jointly impossible.
- Confidence thresholds don't help: this is 'the question was wrong' (mu), not 'the model is unsure'
- Fix: encode policy as constraints over the answer set, run a SHACL checker after the call, abstain or escalate on failure; the checker itself needs auditing (91/120 W3C Core tests, 29/78 OWL-RL rules Lean-certified)
- All six guardrail projects on awesome-jev screen inputs, none validate cross-answer consistency
More from coding & agent
- Open-source testsprite CLI lets coding agents self-test live backend APIs and auto-fix failures — KARYTON618 · 2026-09-20
- OpenEnv sees active community work on infra providers, MCP and WebSocket fixes — ben_burtenshaw · 2026-09-20
- Gemma 26B A4B Aces a C++ Coding Test Locally but Fumbles Tool Calls — HyperWinX · 2026-09-20
- mitsuhiko on the Shared Frustrations of Agentic Software Engineering — mitsuhiko · 2026-09-20
- Comprehensive 55-minute Codex Desktop Tutorial Covers Skills, MCP, TikTok and Blender — aziz4ai · 2026-09-20
- Stalkr adds keyword groups to benchmark your brand vs. competitors, with API and MCP access — marclou · 2026-09-20