The 1.28% is now an instrument instead of testimony. evals/perception/read_rate.py
measures the logic reader over holdout_dev/v1 deterministically (23/1,798
sentences, refusal taxonomy included, 0.57s); tests/test_read_rate_floor.py pins
it in both directions — a comprehension regression fails the gate, and an
improvement must move the recorded constants in the same reviewed commit, so
capability motion is a decision with a diff, never an accident with a story.
Non-vacuity guarded (the matcher must still say no; counts must conserve).
Two sabotages observed red. Registered in smoke (~0.8s).
This is one of exactly two old-plan items that survive the keel triage (the
other is FA-1): it is the baseline number the keel's K4 perception layer is
accountable for moving in chunks. PR-8, PR-10, Track B widening and Track C
are superseded by keel phases K6/K6/K4/K5 and are not being finished here —
finishing them would be polishing the quarry.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Wcw2pnMBwyvmNyQg4uPEt4
472fc0a83f