Retrospective Out-of-Sample Check: Critical Heat Flux Model Version 1 Against the 2006 Look-Up Table on Three Water Sets Opened After It Was Frozen

This record pre-registers and reports a retrospective check, not a blind test. Version 1 of the programme's water critical heat flux (CHF) model, a seven-parameter pi-group model frozen on 2 October 2026, is scored on three water point sets the programme opened after it was frozen, each by version 1's own pass rule: H1 passes if its relative RMS error is lower than the 2006 look-up table's on the same rows. No model scored was fitted or changed with these points, but each set had already been opened and scored by its own test, so the results carry less weight than a sealed test. The rule was frozen and committed to a public, time-stamped index (Hugging Face commit 4aee05bd) before any version 1 prediction on these sets was computed, then checked and scored once. Results (relative RMS error, version 1 against the table): on the Oettig et al. near-critical series from TU Munich, 22.20% against 10.94% (n = 24), H1 fails; on Chen et al. 2018, 9.43% against 8.85% (n = 30), H1 fails; on version 5's test set with all its rows, 16.33% against 19.40% (n = 82), H1 passes. That pass rests on keeping Wang 2023, which version 5's own test had left out as a repeat of Wang 2018's test section: without it version 1 scores 16.34% against 15.96% (n = 66), and the table leads. Two further reference lines, not verdicts: without Wang 2023's four low Fig. 16 runs, 15.91% against 16.99% (n = 78); without Statham 2016, 17.15% against 20.22% (n = 74). The three pass rules test one model on sets from different laboratories, so the results are not independent, and no programme-level verdict is drawn. Known before this record and reported, not rescored: version 2's regional model and version 3 are void on version 5's set (no point in version 3's domain), and version 2's regional model is void on Chen 2018. Disclosed after scoring, information only: three Wang 2018 points at 22.0 MPa, just below the critical pressure, carry version 1's largest errors (55 to 68% low); without them version 1 would score 11.15% against 19.00% (n = 79). This was not a rule line and changes no verdict. Version 1's training pressures reach 18.0 MPa, and 14 of the 24 Oettig series lie above that. An independent agent recomputed every error figure from the points without the scoring code, and reproduced the whole run byte for byte with the unchanged retro.py. The deposit holds the frozen rule (text and PDF), the driver retro.py, the freeze file with its hashes, the run log, the checks run and the scoring run with every point's measured value and both predictions. Archive SHA-256: 93279ee786e9c48d184d03595a68f5eb61807f350ad09785e666d5ba4a6b1fc9