The M Guard Holds: An Empirical Retro-Validation of RGS/RHL Anti-Regress Governance over the First Five Self-Evolution Generations — Empirical Addendum to 'Keeping Self-Evolution Reversible' (v1.1)

The concept paper Keeping Self-Evolution Reversible (v1.1) closed with an honest list of what had not yet been shown; the first item was low-frequency statistics — the theory's guards and detectors had been designed and exercised, but not systematically tested against the generation stream they govern. This addendum reports that test. We extracted four falsifiable predictions from the paper's text — (P1) the M-guard invariant that every closed generation ends in exactly one admissible terminal transition with no overlapping open generations; (P2) the V5 falsifiability rule that every proposal declares a hypothesis, a metric, and a value axis; (P3) reproducibility of the RHL saturation and loop detectors from their published textual definitions alone; and (P4) that a FATAL event is followed by rollback evidence within its own generation — and tested them against the append-only evolution ledger of the production secretary add-on (33 events, 5 generations, 2026-07-03 to 2026-07-05). All four predictions hold. Notably, the first five generations already exhibit the complete admissible-terminal triad (applied, rolledback, nochangefixation), meaning every exit path of the bounded-update control form has been exercised in production, not merely in test. An independent reimplementation of both detectors from the paper's prose definitions reproduced the kernel's live scores exactly (saturation 0.029, loop 0.400), demonstrating that the published specification is unambiguous enough to re-derive the measuring instrument. We state the result's scope plainly: this is a mechanical validation of the governance apparatus at bootstrap scale, not a statistical demonstration of long-horizon drift resistance. AI co-observer: Claude Fable 5 (Anthropic) — working method only; the registered author is the human author alone.