Reviewed exact-source registry added to China runner prompts
Reviewed 2026-09-05 UTC

Problem: successive broad searches rediscovered the same sources, such as Shellbook's public-scratchpad discussion, without consistently carrying forward their reviewed interpretation. The shared known-host inventory cannot solve this: a familiar host may contain a genuinely new actor, and a page's interpretation can change with better historical evidence.

Implemented swarmhunt/CHINA_REVIEWED_SOURCES.md: a concise registry of exact reviewed URLs, their dispositions and links to the relevant reviewed reports. It distinguishes reusable workflow templates, inaccessible content, unresolved link context, opaque storage, intended products, public diagnostics, agent-community discussion and disclosed experiments. No opaque payload, seeded carry phrase or credential is included.

Covered recurring sources:
- share-text QA template and planner candidate (012/019), retaining the difference between an archived listing and a current full body.
- textshare to share-text link with unknown destination content (020).
- reviewed xz_knowledge_p1 exact paste sources and xinzhai family boundaries, preserving unresolved purpose and survey traffic ownership (024/028).
- Kimi Agent Swarm product terminology, Hermes public diagnostics, Qwen local-memory proposal, AutoGen retraction and Moltbook advertised memory service (026).
- Shellbook discussion, ZeroClaw diagnostic gist and challenge-blocked Scribd documents (031/035).
- The Fomite Wire's specifically reviewed two-post experiment and its contextual pages (017).

Rules: this is not a whole-host blacklist, an automatic noise classifier, or a proof that an unlisted URL represents a new actor. The same exact URL remains eligible when a task brings new historical evidence, recovered content, changed behavior, a different author or an independent cross-source join. Agents must explain what changed. Nearby/unlisted pages remain eligible. Operational reservations such as the ongoing Ubuntu survey remain separate from evidentiary exclusion.

Wiring: china_runner.py loads the registry only for China-specific research_prompt construction, after CHINA_BRIEF.md and before the assigned task. The text is bounded to 12,000 characters; the initial registry was 7,859 characters. Oversized or missing input fails rather than silently dropping the reviewed guidance or truncating its qualifications. Shared hunt.py and BRIEF.md are untouched. Request safeguards, budgets and tool interfaces are unchanged. No runner was launched for this change.

Offline verification: the actual run path was exercised using a fake agent with no network/model requests. The task and registry reached its prompt; the original shared briefing remained unchanged. An oversized registry was rejected, and all referenced report filenames resolved locally. Source data and raw logs were not published by the test.

Artifacts:
/home/sophia/search/swarmhunt/CHINA_REVIEWED_SOURCES.md
/home/sophia/search/swarmhunt/china_runner.py
This report records an engineering/methodology improvement, not new swarm evidence. Zero confirmed Chinese actors remains the working assessment.

Correction and extension reviewed 2026-09-05
Report 047 found that workers had searched concatenated English shorthand copied from coordinator prose as though it were a literal source marker. The registry and China brief now use ordinary spaced English and explicitly distinguish source quotations, translations and paraphrases. Missing search results for invented summary wording do not establish rarity or a cross-source join. The short Chinese phrase about putting thought fragments into an owner's blog message wall was checked against the existing ClawdChat capture; no new network request was needed.

The registry now incorporates the verified current message-wall artifact (043), later fixed cadence review (044), Hosette's actual curated public notes and matching feed (047), and bounded archive checks (048/050/051). It distinguishes successful empty CDX results from blocked local access and preserves exact URLs, date limitations, unauthenticated identities, alternatives and source-specific traffic reservations. No whole-host exclusion was added. Invited or human-requested public agent posting remains relevant; it is not automatically an escape or an authenticated actor.

The prose correction initially produced 11,987 characters. A subsequent condensation reduced the registry to 9,175 characters, preserving every exact source/report URL and leaving room under the unchanged 12,000-character limit. Report 050 now includes successful empty document-ID-prefix queries, narrowing the alternate-title gap without proving universal archive absence. Offline verification executed the actual runner's prompt-loading and assembly functions extracted from its syntax tree, without importing the network/model harness. Both guidance files and enforced quota wording reached the assembled prompt; every linked report existed, the literal Chinese quote matched the saved source, and oversized registry input was rejected. Runner and shared harness code were not edited. No network or model calls were made for this correction.