Methodology audit: known markers and known hosts must not gate second-actor discovery
Reviewed 2026-09-05 UTC

Audit finding: the shared swarmhunt/BRIEF.md is strongly centered on the known thread A. It lists specific markers and datasets, asks agents not to re-investigate known hosts, and tells them to fetch and run signature_scan before recording. These are understandable discovery preferences for that original objective but would be an unsuitable gate for finding an independent actor on an already-known surface.

Important qualification: the China runner was already substituting CHINA_BRIEF.md for the shared brief. It did not pass the entire original A-marker list as its current top-level briefing. CHINA_BRIEF.md already allowed English/overseas activity and warned against limiting searches to Western names. Therefore this audit does not establish that the completed China wave was driven exclusively by the shared A-only brief.

Residual anchoring risk remained: the China briefing says to use is_known_host, while that tool's description emphasizes new places; signature_scan recognizes historical A/B/noise patterns; and record_finding defaults to thread A if the caller does not specify otherwise. None of those tool outputs can establish an independent actor's absence or origin. These are prompt-design risks, not a measurement of how often past agents made the mistake.

Implemented correction:
- Revised the existing swarmhunt/CHINA_BRIEF.md with a language-neutral opening and explicit methodological clarification, avoiding a redundant second briefing file.
- A known host may contain a second actor. Inventory membership is context, not a page/author/cluster exclusion gate.
- The A component of signature_scan is a known-cluster classification/join aid, not a requirement for an interesting lead. The scanner also has B/noise patterns; it is not literally an A-only implementation. No match means only no regex match.
- Behavioral evidence is primary: actual dated sequences, coherent research tasks, cross-surface unusual identifiers and independently verified public actions, evaluated against alternative explanations.
- Chinese text, a Chinese/cloud IP, English text, unfamiliar markers and “not A” do not establish Chinese attribution. Separate observable behavior, independent-agent hypothesis, and operator/model attribution.
- Candidates must explicitly use thread='?' rather than inherit record_finding's A default. A matching known marker can itself be a quotation or copy requiring context.
- Raw English phrases on global websites/GitHub and Chinese 智能体/多智能体 searches remain in scope. Ambiguous 代理 is not an AI-agent gate.

Prompt wiring now uses a runner-local research_prompt function combining the revised China brief and the assigned task. The runner no longer overwrites hunt.BRIEF. Shared BRIEF.md and hunt.py were not edited. Other harness runs retain their existing briefing. Request safeguards, budget logic, tool interfaces and manifests were not changed by this correction.

Offline verification checks the exact prompt passed to a fake agent, confirms the methodological instructions occur in it, confirms the original hunt.BRIEF remains the shared file content, and confirms the task title/body remain present. No model/API requests are needed. This change affects subsequently started runner processes; it does not retroactively rewrite earlier reports or alter an already-loaded process.

Sources reviewed locally:
/home/sophia/search/swarmhunt/BRIEF.md
/home/sophia/search/swarmhunt/CHINA_BRIEF.md
/home/sophia/search/swarmhunt/hunt.py (signature_scan, is_known_host, record_finding)
/home/sophia/search/swarmhunt/china_runner.py (prompt construction)
Related evidence review: 008-capability-hypotheses.txt and 025-language-and-attribution.txt.

No new swarm artifact or attribution is claimed. The correction improves the discovery method while preserving the working assessment of zero confirmed Chinese actors.