Selective filtering is the design principle stating that execution experience must be critically abstracted and filtered before being committed to persistent procedural memory. Unfiltered failure-driven adaptation risks three primary failure modes: overfitting to ambiguous task specifications by creating arbitrary default conventions, memorizing external evaluator quirks rather than genuine task contracts, and over-abstracting heavy, restrictive procedures that harm general performance across varied workflows. Harness scaling is fundamentally a problem of abstracting and protecting consequential execution boundaries rather than accumulating rules from every observed failure trace.