Establish the provenance of the historical repaired-1B lineage

Establish which numerical reduction path produced the historical parent checkpoint’s recorded embedding-table norm and determine whether the historical repaired-1B lineage differed from the reproduced lineage in any way not detected by the available pathology checks.

Background

The original recipe-C repair checkpoint and its warm-start parent were lost when the infrastructure holding them was retired. The paper therefore reports results from a reproduced run using a surviving sibling of the lost parent.

A discrepancy exists between two methods for computing the embedding-table norm: the historical record reports approximately 635.2, whereas one library reduction reports approximately 630.13 and an explicit sum-of-squares computation reports approximately 635.16. Because the historical logs and parent checkpoint no longer exist, the paper cannot determine which reduction path generated the original figure or completely rule out an unobserved lineage difference.

References

One thing remains genuinely unresolved. No logs survive from the historical run or from the job that produced its parent, so we cannot establish which reduction path produced the $635.2$ figure recorded for that parent at the time, and therefore cannot fully exclude that the historical lineage differed from the reproduced one in some way the pathology gate would not have caught.

— Keyword Harnesses Fail Open: A Cheap Diagnostic Ladder for Tool-Use Claims in Small Language Models  (2610.02142 - Santillana, 1 Oct 2026) in Section 3.5, paragraph “L12: The repair's original checkpoint and its parent are lost; every repaired-1B number here comes from a reproduction,” item (d)