I built MCA (Metacognitive Control Architecture), an open-source neuro-symbolic layer that wraps any LLM. Before an answer goes out, MCA asks: is the question itself sound?
5 layers: 1️⃣ Premise validator: an NLI cross-encoder checks the question's hidden assumptions 2️⃣ Semantic entropy: samples the model several times and measures how much the meanings disagree 3️⃣ Z3 logic engine: finds paradoxes and contradictions 4️⃣ Evidence verifier: grounds claims in a retrieval knowledge base 5️⃣ Policy gate: returns PASS / CHALLENGE / ABSTAIN / EXPLAIN_PARADOX
Raw Llama-3 happily answered 89% of the trap questions. Wrapped in MCA, it caught every paradox (25/25).
⚖️ Honest caveats: • MCA refuses more ordinary facts (20–52%) • It adds 4–10× latency • The knowledge base was written with the benchmark topics in mind Calibrating the thresholds is next.