LIVE COORDINATOR REVIEW — FIRST SEARCH WAVE
Updated 2026-09-05 10:33 UTC. Interim, not final.

Working assessment: no confirmed second actor. The 72-task wave is running with 24 workers. Completed machine reports are stored under swarmhunt/runs/CN26*.md and remain unverified unless specifically reviewed here or in a dedicated report.

Reviewed low-confidence lead: http://zblog.nciaer.com/?id=2
Agent CN26005 flagged a comment because it fell in the priority date window. Coordinator fetched the actual public page: HTTP200, displayed comment author/body both “消防培训” (fire-safety training), displayed publication 2026-05-06 00:51:48. The page is a default Z-BlogPHP guestbook. No research-data link, distinctive test sequence, coordination pattern or independent join is shown. This does not meet the candidate threshold; ordinary promotional comment is an adequate explanation. Its existence is verified, its alleged relevance is rejected.
Raw evidence: pastebins/data/zblog.nciaer.com/china-review-id2.html; request metadata alongside. SHA256 670e4bf1f75a17fc5de781c6382dc2d7dbb342ba0d2b4bb6ffc1e2dfcb647959. Displayed date is a website claim, not an independently archived date.

Useful coverage leads (not actor findings): Chinese DokuWiki and PukiWiki instances, paste.ubuntu.org.cn public recent/RSS listing, and share-text.org multilingual public-note pagination. Initial bot reports mostly inspect present-day pages; that does not cover May–August historical content. A dedicated archive pass is now checking share-text.org's older public listing.

Native search reliability: a small Baidu reachability test succeeded earlier, while fleet queries subsequently encountered captchas. Reachability is query/time dependent, not a universal reachable/blocked property of a domain. No challenge was solved or bypassed.

Logging caveat: the inherited tool harness uses thread-local task IDs, but the agent library can dispatch tools in different threads. Some first-wave shared-log events consequently have task='?'. Per-task final reports and tool transcripts still exist; first-wave findings counts by task prefix can undercount. The next-process harness fixes context propagation and adds a sanitized per-task event ledger. Do not treat an absent task-tagged finding as proof that its tools did nothing.

Operations notes: the published spend figure is measured OpenRouter key-usage delta, not a guaranteed total-cost accounting of other services. The $20 first-wave guard is periodically checked and cannot cancel already in-flight API charges. Existing AWS corpus is being used for one additional scan; the instance was not purchased or terminated in this session. Its billing status must be checked separately from OpenRouter spend.

Runtime update 2026-09-05 10:49 UTC: interrupted the initial runner after 21 completed reports to apply bounded network queues/timeouts and shared failure cooldowns. Completed files and all cached pulls were preserved; unfinished in-memory task transcripts were not checkpointed and those tasks restart. Resumed original manifest skips existing reports. Next-wave supervisor continues waiting for completion. Fixed baseline usage is shared across resumed/follow-up processes, so the $20 session guard does not reset.