In multi-agent dialogue, a hallucination crosses an operational control boundary when another agent accepts it and downstream decisions inherit the error. We formalize this cross-speaker transition as unsupported endorsement and introduce SocialCritic, a learned receding-horizon controller that ranks and repairs candidate action sequences with separately trained utility, norm, and epistemic advantage critics. Across 692 test episodes spanning LIGHT, DealOrNoDeal, SOTOPIA, and template-family-held-out Persuasion-Conflict over 10 generation seeds, SocialCritic exceeds SAFE-Dialogue by 6.9 points in goal completion and 7.9 points in actionable accuracy while reducing endorsement by 44.9% and norm violations by 42.7%. The public-benchmark aggregate preserves the actionable-accuracy gain (74.9% vs. 66.8%), and trace analysis shows that endorsement falls 77.5% versus 63.5% for initial injection relative to LLM-only. Leave-one-benchmark-out critics retain an actionable-accuracy advantage on every target (minimum 5.2 points), while an H=1 controller without partner rollouts preserves 71% of the full controller’s advantage with 18% lower latency. At the default operating point, batched local scoring requires 1.7 external round trips and 748 ms per turn, enabling propagation-aware simulation, safety evaluation, negotiation, tutoring, moderation, and red-team workflows.

Inference-time search can substantially improve LLM reasoning when tasks admit deterministic verification, but existing methods largely refine single trajectories and lack a reliable mechanism for composing partial solutions across candidates. We propose contract-checked graph editing: represent each candidate as an interface-typed reasoning DAG and validate every nontrivial edit with a deterministic structural gate (acyclicity, namespace closure, schema validity, terminal constraints) before invoking the verifier. The gate certifies runnability only and emits auditable rejection reasons; semantic correctness is determined solely by the verifier. Instantiated in Genetic Inference Search (GIS) with Qwen2.5-32B-Instruct under strictly matched token budgets (8K tokens), contract-checked grafting increases verifier-runnable recombination from 41.2% to 92.8% and improves accuracy over rStar (+6.1 on MATH, +9.1 on MATH L5) while using 42% fewer verifier calls. The same operators transfer across outer loops (beam, best-first, MCTS) and to structured generation and code, outperforming execution-guided beam search on Spider (+2.8) and improving multi-file code generation on HumanEval-MF (+9.2).