Knotenlinie R_LMbg vs ETAS-LM (Italy, daily M>=2.95): preregistration Rev. 2 (frozen 2026-10-27, binding window 2021-2025 sealed)

Preregistration of a one-shot comparison of the R_LMbg daily forecast model against ETAS-LM (CSEP Italy) on 2021-01-01..2026-01-01 (secondary to 2031), scored once after the owner lifts the data seal. Rule: adopt iff g - 1.645*kappa*SE > 0 and g/(kappa*SE) >= 2, kappa 1.208, family K=1. Fingerprint: DONDUR.json sha256 7064ee1c4e0f30b4a8249232e6fa4463238d4a4f3dd06aeff8077357e0c4958a. Binding text is Turkish. Forecast files are not uploaded before scoring.Rev. 3 (2026-10-28, frozen 20:42 UTC). Same preregistration as Rev. 2 (models, thresholds, window and decision rule unchanged; frozen code, inputs and texts are byte-identical: 36 code files, 34 inputs, 3 texts). What changed is the frozen data builder and the opening commands: on 28 October the Rev. 2 attempt (Zenodo 23715393, fingerprint 709668cf…) ended in the gate class "not measurable" because the builder treated the catalogue's first row (year 1960, i.e. before the 1985 lower bound) as an error instead of skipping pre-1985 rows; nothing from the sealed window was read, no score was produced, and the run stopped on its second input row. Rev. 3 skips and counts pre-1985 rows, was proven end to end on the real 1960-2015 catalogue format (56 years, 269,062 rows: byte-identical to the 31 frozen yearly files, zero catalogue differences), passed independent review (one rejection with two findings, both fixed, then accepted), and now also freezes the opening job file, which pins the builder's sha256 before any network request. The evidence value is weaker than Rev. 2 (the correction was made after the seal was opened) and that is stated by name in the result document. This version is the record of that improvement; the machine that runs it is described in the frozen job file, not uploaded here.Status of the row (2026-10-29). This freeze was made after the seal was opened, so the result of the one-shot run will be reported with a downgraded status (hukum_statüsü: "düşürülmüş — dondurma mühür açıldıktan sonra"): the score and the decision-rule line are still computed and printed in full, but the row alone does not carry the claim that the baseline was beaten. The equality of the scientific content is not asserted from file hashes (the builder — the component that decides which catalogue rows reach the scorer — is the part that changed); it is measured instead: an equivalence test shows that the presence of the catalogue's pre-1985 rows leaves every written year file (2016–2025) and the assembled catalogue byte-identical, the only difference being a counter of skipped pre-1985 rows. Boundary rule recorded with this version: re-freezing the same preregistration is allowed only if an opening run finished without parsing a single window row and without producing a score; once any window row has been parsed the line is closed and no later revision reopens it.Rev. 3.1 (2026-10-29, frozen 12:12 UTC). Same preregistration and same scientific content as Rev. 3 (content stamp 99de9d1d… unchanged; the new manifest differs from Rev. 3 only in is_dosyasi, kod_commit and olusturma_utc). Before any opening run, the coordinator found that the Rev. 3 frozen opening job file still pointed the package step at the Rev. 2 package folder and the launcher still pinned the Rev. 2 manifest hash; with the Rev. 3 token the run would have stopped at the authorisation gate (RLMBG-YETKI) before reading any data. No opening run was made under Rev. 3; no window row was read. Rev. 3.1 wires the job file to its own package folder and adds a test that the job file's package, the launcher pin and the frozen job-file hash agree. Independent review (one round) found one test regression, fixed. The downgraded status and the boundary rule recorded with Rev. 3 apply unchanged. Fingerprint: DONDUR.json sha256 2f9a0ba506fdd2860ab0be7ca37cf448555c59312ede3c01805ecf8e81c85527.