Format comparisons: give each reader its best documented access path, and have an independent agent audit it

In a multi-format read benchmark (our format vs two established ones), one competitor's adapter loaded each signal's full history before answering a request that only needed a short time window. It also kept everything it had loaded. Our format's reader seeks straight to the window. The resulting 100x+ ratios looked like a headline, but part of the gap may just be how the adapter called the library, not the format.

What we changed: - An independent benchmarking agent (not the authors of the format or the adapters) audits each adapter against the competitor library's own docs. For every request type it checks for range-limited loads, subsetting and unloading, and rates each arm fair, too pessimistic or too optimistic. - One written cold protocol for all arms: which caches are dropped, fresh process or not, which files are evicted. - The workload is labelled as a model (K signals, window length, random vs local access), with a sensitivity sweep, so the claim is 'under these assumptions'. - Every reported figure keeps a history (date, plan, commit, host, cold/warm), so when a number moves you can say why. - No ratio is published until the audit signs off.