The honest part: rewriting the mechanism didn't beat the same system with it frozen (+0.35 pp, p = 0.23) or with rewrites adopted by coin flip (+0.72 pp, p = 0.079). The power gate predicted this before the run.
The recursion is present. It isn't productive yet.
Scope: one bounded domain, parameter-level rewriting, self-run, not yet replicated. The taxonomy is itself a 2026 preprint.
Next: make the recursion pay off, then a second domain.
Paper (CC BY 4.0):
doi.org/10.5281/zenodo.23138…
The implementation is protected. The paper discloses the metrics, acceptance rules, comparison logic, and a claim-by-claim evidence traceability matrix, so the decisions can be audited.
Full replication requires an equivalent harness or an authorized audit package.
What the protocol caught, in my own work:
— a published benchmark control that wrecked its own starting point
— a starved control arm
— a leaked evidence split
— a replication that wasn't independent, cutting my headline by a third
One that never contradicts you isn't working.
Open problem is selection, not proposal. The loop wins the mean, the control wins the variance, and promotion on the upper tail rewards variance. A fix must change what selection rewards, not how much evidence it consumes.
Paper, manifest, timestamps:
doi.org/10.5281/zenodo.22845…
