Sound online growth of a typed 'isa' lattice from noisy LLM extraction, through candidate elimination made noise-tolerant by a localized-blame admission gate

Couple an LLM to a knowledge base and you inherit one of two failure modes: a symbolic side thatnever learns, or a base that absorbs the model's *world-plausibility* — plausible-but-false factswith no correction channel (NELL's drift is the textbook case). This paper presents and measures athird option: sound **online growth of a typed *isa* lattice** from the noisy extractions of asmall local LLM. The instrument is a single **localized-blame admission gate**: evidence isadmitted for a unit — slot restriction, *isa* edge, or surface alias — only if its outcome isuniquely attributable to that unit (structural provenance or counterfactual ablation) and verifiesagainst the declared oracle; soundness is *recoverability* (provisional/confirmed tiers,retraction). The gate makes candidate elimination tolerant to the one-sided, competence-correlatednoise of an LLM pipeline. Measured: a deterministic laboratory (126/126 pre-registered checks); alive circuit (54/54 with defeasance; 300/300 at volume vs 245/300 direct); an ungated control thatreproduces the NELL signature live (false edges, poisoned aliases) where the gated arm admits noneat equal resolution; external validity on DeFAb and a nine-model cross-family campaign (0 falseedges throughout). The deposit bundle replays every table bit-for-bit from a content-addresseddurable memo — no GPU required.