The temporal probe is weaker than the design assumed¶

Zhang et al., Test of Time: Rethinking Temporal Signal of Benchmark Contamination (arXiv:2509.00072, ACL 2026), argue that post-cutoff performance decay, the signal the time-partition probe reads, is not dependable evidence of contamination. Their result is that the construction of the items distorts the temporal pattern independently of the source material: model-transformed questions and cloze questions drawn from the very same documents produce markedly different temporal signals.

Their specific confound is controllable here and is now controlled. Every example in both windows comes from one deterministic pipeline with one anchoring rule, and the control window was re-collected as a single uniform window with the cutoff at its own first day so that it is constructed identically to the window it is compared against, rather than as six months each built as their own window. Construction is held fixed; only the date varies.

Their general point survives that fix. A time partition cannot separate exposure from ordinary distribution shift, because a codebase's conventions, reviewers and subject matter all move over eighteen months. So the registered reading changes:

• Min-K%++ on the base checkpoint is the primary instrument. It scores the text directly and does not infer membership from a performance difference across dates.
• The time partition is corroborative only. A gap in the expected direction supports the membership result; a gap on its own is not read as contamination, and its absence is not read as a clean bill. This was already the pre-committed reading of an ambiguous battery; it is now also the registered reason.
• Guided completion was already at its floor in both windows, verbatim and by edit similarity. Two of three probes are therefore doing little, which is worth stating in the report rather than presenting a battery of three.