The preceding paper in this series typed truth for an embedded observer but deferred the difference between true-by-luck and knowledge to the present paper. This paper supplies that anti-luck treatment with a method-relative, two-sided, aptness-robust construction. It denies a global knower—no physically realized subsystem combines complete truth access, warrant integration, modal comparison, and infallible authority—while affirming local and distributed knowledge. An Epistemic Specification separates the actual belief-forming method, taken up to an admissible method-grain, from the family of admissible inspections; fixes the doxastic bearer as a raw token whose content is set by alignment; and distinguishes belief from truth-directed acceptance. Anti-luck provenance is decomposed into five lineages; the actual-world truth connection is given by claim-type-indexed grounding modules, kept type-distinct from the method’s modal performance. Discrimination safety conjoins actual-world responsiveness with a positive margin over all relevant non-true worlds—false or non-apt—which screens off rare-random-guess methods. The Kernel–Safety results come in exact and approximate form: when a relevant non-true world is unresolved by the actual protocol, the method forms the target belief with equal probability in both worlds, so the discrimination margin is non-positive; when the transcript laws lie within total variation δ, the margin is at most δ, a bound that applies to finite, noisy, and single-shot records. Warrant, anti-luck connection, and undefeated standing are separated, with explicit non-vacuity and method-individuation constraints. The framework’s knowledge predicate is a technical, framework-relative standing for empirical and physical knowledge, not a final analysis of the ordinary concept. Knowledge requires neither a global knower nor infallible self-certification, but more than truth, local warrant, agreement, or recovery success.
Tomoyuki Uchida (Thu,) studied this question.