Farfaz Learning

A model can produce a fluent plan that violates one small rule. Rule Replay makes that gap visible before the plan is treated as valid. The prototype uses a fixed fictional game. Meta Llama 3.3 70B (via Cloudflare Workers AI) proposes an action sequence from source rules and a goal. An independent, developer-authored simulator replays each action, preserves the state trace, and returns the exact failing step as a counterexample. The model gets one bounded repair attempt using that counterexample; the simulator checks the repair again. A second failure remains a failure - nothing is ever labeled verified unless it passed the replay. The live demo shows source rules, proposed actions, counterexample and replay side by side. A deliberately authored negative test is labeled separately from actual model responses. Unsupported scenarios cause no model request, unknown actions are refused, and inference outages never turn into a simulated success. Observed honest result: across repeated live runs, Llama's plans failed the replay (collecting on a square with no star, step 4), and the single counterexample-grounded repair also failed at the same step. The product reports not_verified with both counterexamples, exactly as designed: the checking mechanism works; the repair did not succeed in the observed runs. We show this outcome as-is. Scope: a narrow experiment in verifiable planning, not a learning product. A passing trace proves conformance to the fictional fixture only - no real-world truth claim, no child-facing deployment, no learner data, no private code. Live product: https://rule-replay.farfazlearning.workers.dev