Quantify the relative bias from unrecorded Stage-0 transport reissues

Determine the relative bias introduced by successful Stage-0 request reissues whose task- and suite-level identities were not retained, using zero-inclusive Stage-0 attempt counts for all evaluated systems.

Background

Stage 0 is the request-level retry performed before a result record exists. For one retained adapter, the authors know that two such reissues succeeded, but the corresponding task and suite identities were not persisted. Because the other systems lack comparable zero-inclusive Stage-0 records, the authors cannot uniformly attribute or correct these events.

The missing attribution matters because the maximum placement-conditioned leverage of the two reissues is larger than the lower bounds of the nominal top-tier score differences. Consequently, the relative bias associated with this incomplete audit trail remains unresolved and is identified as a requirement for the frozen-v1 protocol.

References

Because the other ten adapters did not retain zero-inclusive Stage~0 counts, the relative bias is unknown.

— IBIB: A Protocol for Measuring Enterprise AI Systems by Serving Route, Not Model Identifier  (2609.10494 - Stenstrom et al., 9 Sep 2026) in Section 4.2, “Sensitivity: what a score difference survives”; Section 6.2, “Reliability and failure evidence”