Fresh sources and host transfer

This overview explains how C23 tests the best-artifact task on fresh generator families and separately recorded hosts. Fresh input families and a new host test different forms of transfer, and neither implies a universal performance ranking. Imagine a bounded task with three valid artifacts, scored three, two, and one under a declared smaller-is-better toy ranking. The true best candidate in this bounded set is C. Exactness and rank are both visible. If the reference omits C, it names B as the winner. A method returning C may then appear to disagree with the reference even though it found the better permitted artifact. The problem lies in the reference's incomplete task coverage. Rebuild the reference to cover every candidate it claims to score. Keep the original defective freeze in the correction trail and bind the corrected result to the actual enumeration bounds. The example does not claim that a bounded enumeration covers an unlimited search space. C23 keeps the seven methods and best-artifact task while changing to previously unused Yosys generator-family functions. The first incomplete task freeze remains recorded; the corrected bounded reference must cover every candidate it claims to score. On the retained Windows table, source-packed ANF narrowly leads the screened paths. The later unchanged Linux table changes the ordering: compiled screened CM leads in aggregate. Both host results are complete, and their narrow aggregate winners differ. Both host lanes retain the same canonical artifact contract, even though their timing rankings differ. Source-family freshness and machine transfer are different checks. The retained report includes both Windows and Linux results, yet narrow headroom and individual regressions keep production promotion false. A correct selected artifact and a new host do not authorize training or a default change. A new source family and a new machine are independent changes. Place comparisons in a two-by-two grid so it is clear which combinations were actually tested. One filled cell does not establish the result in the others. Two implementations can prepare and screen data differently while owing the same canonical artifact. Compare both reconstruction and the complete ranking fields. Different internal timing orderings do not redefine the caller's contract. A narrow aggregate lead can change across hosts while every returned artifact remains exact. Preserve both host records and individual regressions. Neither a new host nor a correct artifact alone authorizes training changes or production promotion. Preserve each host's exactness and timing results. Keep individual regressions and the unchanged non-promotion decision visible.