Pre-registration, Version 5: A Local Blend of the 2006 Look-Up Table, the π Table and a Corrected Table for Water CHF, Blind-Tested on Laboratories Outside the Table's Database

Pre-registration of a blind test, frozen on 9 October 2026 before any CHF value of a test article was seen. After the version 4 freeze (doi:10.5281/zenodo.23254782), an exploratory round scored three ways of predicting critical heat flux (CHF) in water tubes on 24,156 water rows, leaving out one laboratory family at a time: the 2006 look-up table, the version 4 π table, and C, the table times a correction learned in the π coordinates. The table was best in short tubes and for subcooled CHF, C at middle lengths, and version 4 in long tubes. A local blend weights the three by their held-out errors near the point, with no boundary chosen by hand. It scored 19.9% relative RMS error and 10.6% median absolute error against 36.5% and 11.6% for the table. These figures come from data already seen, and they flatter the blend, because the table was built largely from the same rows. The test articles come from three laboratories outside the table's database: Moon et al. (2005, KAERI), Statham (2014 thesis, McMaster), Statham and Novog (2016, McMaster), and Wang et al. (2018 and 2023, Xi'an Jiaotong University). H5: the blend's relative RMS error on the pooled primary points is lower than the table's. H5b: its median absolute error is lower than the table's. Both are void with fewer than 10 primary points. A test-article choice withdrawn before freezing is stated in full: three CIAE conference papers, dropped when an audit found that one of them had already been seen. Predictions at fixed conditions are published, together with a known weakness at high quality inherited from version 4. The deposit holds the binding protocol, the test code with hash and once-only checks, the held-out errors that set the blend weights, the published version 4 files, the controls, four audits by separate AI agents with the changes made in response, the development scripts and logs, both source searches, and a consistency pass of the record. The test articles will be obtained after this record is public. Archive SHA-256: 3cf6d178e03639b99fcf638cf20959f175909ec496797af116f6592629ca15d2