Pre-registration, Version 6: Repairing the Water CHF Blend Where Version 4's Curve Turns Up, Blind-Tested Against the 2006 Look-Up Table and Version 5

Abstract

Pre-registration of a blind test, frozen on 10 October 2026 before any CHF value of a test article was seen. Version 5 of this programme, a local blend of the 2006 look-up table, the version 4 π table and a corrected table, passed its blind test against the table on 42 water points (doi:10.5281/zenodo.23282663). Run afterwards on data already seen, it overpredicted a 7.1 m tube at 15 to 18 MPa by up to 57%. The cause is the shape of version 4's local critical heat flux (CHF): it falls with quality, reaches a minimum near 0.6 and rises again, and a long tube's critical flux stops at that minimum. Version 6 blends the three local curves on a quality grid with version 5's weights, gives version 4 no weight wherever its curve rises with quality, and finds the critical flux from the tube inlet. On 1600 development rows, each predicted with its laboratory family left out, version 6 scored 7.6% relative RMS error and 3.6% median absolute error, against 17.6% and 4.0% for version 5. A further correction, learned from version 6's own held-out errors on 23,867 rows, was tried and failed its pre-stated rule; it is reported, not adopted. Version 6 is not better everywhere: on the McMaster runs of the version 5 test it gives up median accuracy, and on the seen Oettig runs every method, version 6 included, overpredicts. H6: on the pooled primary points, version 6's relative RMS error is lower than the 2006 table's. H6b: the same for the median absolute error. H6c: on the primary points where version 5 and version 6 predict more than 5% apart, a set fixed by the predictions alone, version 6's relative RMS error is lower than version 5's. The test articles come from two laboratories outside the NRC KM-0011 file: Previti et al. (1966, SORIN Saluggia, EUR 3113, 245 runs) and Würtz (1978, Risø, Risø-R-372, 71 runs). A TU Berlin dissertation at low pressure (Ruan, about 1993) is admitted only if it passes a stated check. The test articles lie mostly where the two versions agree, so H6c may be void, and the record states that version 6 is untested at 14 to 18 MPa in long tubes. An independent AI audit found no critical problem; its four major findings, all in the protocol text, were fixed before freezing. Version 1.1 adds Addendum A1, written after version 1.0 was published and before any value of a test article was read. It changes nothing in version 1.0. In critical heat flux an overprediction is the dangerous error, and version 6 was judged on averages; on data seen before the addendum it shrinks version 5's share of predictions more than 10% too high (7.4% against 10.9% on development rows) but not the table's (6.0%). The addendum adds a conservative bound, version 6 times 0.8104, set on the development rows with each laboratory counting equally so that about 1% of an unseen laboratory's points lie above it (1.16% on average when each laboratory is held out, at a median margin of about 19%). Local, condition-dependent margins were tried and dropped: they did not predict which laboratory runs high. The rule written before the last round adopted a 5% target (factor 0.8720); the 1% target is the author's later choice, made on development data. No bound is claimed above 20 MPa, where every method overpredicted the seen near-critical runs. H6d: at most 1% of the pooled primary points lie above the bound. H6e: version 6 has a lower share of predictions more than 10% high than version 5. H6f: the same against the 2006 table, which the development data suggest is more likely to fail than pass. A second independent AI audit found no critical problem; its findings, including a correction of the times in the development log, were fixed before deposit. Every frozen file of version 1.0 is carried over byte-identical. Archive SHA-256: b6e4faa69a3e1759d80b7ba69b9566d1656a8cf3092cfd012b797cfba2392667