Framed in Jul 2025, [aswritten-as-first-user discipline](https://aswritten.ai/narrative#Claim_AswrittenFirstUserEvalRigor) established the foundational bar: a pre-registered battery run in full with numbers reported counts as validated; informal smoke-tests do not. The deliberate revert of the merge — keeping dev as a clean \"before\" baseline — was the first concrete act of that discipline. By Jul 2026, [the single-run timing condition](https://aswritten.ai/narrative#Claim_IterateBatchesResolvesSingleRunCondition) was resolved: the final full comparison runs once, against the converged pipeline, after queued fixes (retry-repair, quote-suffix elision) and the prompt-cache dividend land — running earlier would invalidate results at full re-run cost. Simultaneously, [ADDENDUM-A3](https://aswritten.ai/narrative#Claim_PreRegAddendumA3) operationalized the pre-registration: every case record now carries its pipeline SHA, making mixed-version masquerade structurally impossible. Days later, [anchor-5 replay validation](https://aswritten.ai/narrative#Claim_ReplayMethodologyValidated) closed the methodology loop — May-era transactions replayed identically to fresh production, confirming the measured 0%→100% citation-offset gap reflects pipeline improvement rather than environmental drift.
Witness phrases
verbatim A pre-registered battery, run in full with the numbers reported, is the bar for 'validated'; 'the smokes passed' is not.
verbatim The control matters more than the appearance of being finished.
verbatim resolves Scarlet's 'only if we only need to do this once' condition
verbatim Every case record carries the pipeline SHA it ran against so mixed-version results can never masquerade as one run.
verbatim the measured gap (0% → 100% position-exact citation offsets) is pipeline improvement, not environmental drift
verbatim The replay methodology is self-validating.
verbatim "The smokes passed" is explicitly named as an insufficient bar — tension between the appearance of being finished and the discipline of a full pre-registered run.
verbatim The single-run constraint cuts both ways: it resolves Scarlet's condition but also means premature execution (before queued fixes land) forces a costly full re-run — timing is a hard dependency, not a preference.
verbatim The self-validation claim for replay methodology rests on one empirical anchor (May-era vs. fresh production parity); the 0%→100% gap interpretation as pipeline-not-drift is load-bearing and could be contested if environmental conditions differed.