The 4-Beat Reflex for Grounded AI

Stop. Search. Cite. Answer. Autonomous AI agents exhibit a persistent failure mode: when presented with a question, they default to fluent, confident answers drawn from training-data priors \u2014 even when a verified, ground-truth source is available. This technical note proposes a 4-beat reflex that interrupts the fluent-answer impulse before the first token is generated. The reflex maps each beat to an operational gate corresponding to one of the four immutable laws of the Law-4 Compliant Framework. v3.0 architecture (this version): production deployment as a two-tier injection pipeline. Tier 1 \u2014 a deterministic, sub-second scorer (kb_tier1.py) that ranks the agent's ideas ledger and recurring-lessons log against the incoming message (token similarity with per-gate time decay, MMR selection, human-decision denylist guard), configurable via an admission floor and suggestion cap. Tier 2 \u2014 semantic hybrid search over the agent's wiki knowledge base (kb_search.py) as fallback. A plugin orchestrates both tiers in the pre_llm_call hook, injecting results before the model responds; session bookkeeping is LRU-bounded with lockstep eviction. Includes the v3.0 README describing the deployed architecture. Companion to Law-4 Compliant Framework v1.3 (DOI: 10.5281/zenodo.17585621).