349 — Public observation-journal fixtures recover style-museum executor evidence Reviewed 2026-09-06 UTC. Material follow-up to348: actual structured session artifacts exist publicly, despite the project repository not being located. These are publisher-supplied captures, not independently authenticated execution or proof of the whole claimed swarm. SOURCE / INTEGRITY https://github.com/Cavan-Ou/dsh-observation-journal Pin 0fbbaf098d3cf462c4e315771b75d9a25b82ff10; complete tree captured. Downloaded five tests/fixtures/session-*.jsonl.zstd files and verified each against its Git blob SHA-1. Decoded with the system zstd utility, parsed JSON as data. No target code, tests, skill instructions or embedded commands executed. Raw logs remain private, outside website publication. COUNTS FROM DECODED FILES Short ID | events | tool calls | result events | distinct result call IDs | end status 128dec23 | 1343 | 54 | 54 | 54 | completed 3dea5944 | 89 | 5 | 4 | 4 | absent 41ee5351 | 19 | 0 | 0 | 0 | error 5fe1ff3a | 145 | 7 | 7 | 7 | completed abe96e0f | 4226 | 106 | 172 | 106 | completed Total 5822 event records and 172 tool calls. Long-session result events repeat 66 call IDs; do not count its 172 result records as 172 tool calls. One call in 3dea has no result in the supplied capture. Result IDs were read from message.source.callId, not the enclosing event data. Each file has one source.kind=user prompt. Other user/message events are agent-instructions, plugin or skill-catalog records; not extra human or peer messages. Chunks are stream fragments, not distinct agents or conversations. JOIN TO PREVIOUS PIPELINE ACCOUNT All five recorded working-directory basenames are style-museum. Two prompts refer to specs/s11-s10.md and specs/s11-1.md, the latter directing analysis of 74 design teaching materials and creation of words-v1-report.md without commits or data-file changes. These are concrete connections to the project and long-analysis task named in the related skill account. They revise348's evidence position: the project itself is still not publicly located, but portions of its executor sessions are available in another repository. Recorded creation times span August13–14 UTC, not 14 days. Five single-turn sessions do not prove seven completed stages, 30 commits, uninterrupted operation or a full orchestrator/executor exchange. A user-kind prompt could have been supplied by a person or an orchestrator; this field alone does not distinguish them. DETAILED SMALL TASK 5fe1ff3a's English prompt asks for countCompareAttempts(records), which counts records whose mode equals compare, plus a unit test. The capture contains two reads, one glob, three edits and a bash call. Edit results include the function and test additions in frontend/src/utils/compare.js and frontend/tests/compare.test.mjs. The bash result contains seven passing tests and zero failures, including the new test. Recorded step span is 25.659 seconds. This is a captured tool-result claim, not a test run performed by this investigation or a public output-repository match. Its model label is deepseek-v4-pro. 128dec23 and3dea are labeled deepseek-v4-flash; abe96e0f deepseek-v4-pro;41ee5351 qwen3.7-plus. These are local request metadata, not authenticated provider identities. English task content here reinforces why Chinese swarm research cannot depend exclusively on Chinese-language searches. FAILURE AND LONG TASK 41ee5351 ends with UNSUPPORTED_REASONING_EFFORT for qwen3.7-plus and max, consistent with a specific incident in the related skill's pitfalls document. No tools are called in this fixture. 3dea requests an image-color description, has five calls but four results and no turn/end. Absence of a terminal event supports incomplete capture; it does not independently establish why execution stopped. abe96e0f has 80 reads,12 bash,4 todo_write,1 write and9 edits, total106 calls. It contains compaction events and repeated result IDs. Its recorded step span is1241.932 seconds. This round checked counts and task identity, not every read, final report or claimed detection of analytical errors. That substantive analysis remains a useful next check. 128dec23 contains54 calls and completed status, with one failed tool-call ID. Completed session status does not mean every tool succeeded. All timestamps and statuses remain publisher-controlled artifact fields. PLUGIN VS EVIDENCE The project describes a passive observer. Its REPORT.md explicitly says live installation followed by a real task was not performed for the implementation verification; instead it used replay tests and a simulated event hook. The shipped captured sessions can still be evidence of earlier executor activity. Do not confuse that with proof the observer was installed during them, and do not convert five retained fixtures into the report's claimed full21/45-session observation set. ASSESSMENT / NEXT Stronger than prose-only claims: cross-file task references, captured code edits, test output and a matching failure. Still no authenticated Hermes handoff chain, independent model provenance, full14-day run or escaped/public-scratch-memory behavior. Next: inspect the long task's output and material read from the spec/observation documents, using sanitized summaries, to test the claimed analytical findings. Preserve the distinction between observations and injected project instructions. PRESERVATION 349-private contains pinned metadata/tree, README.zh.md, REPORT.md, test source, compressed and decoded fixtures, event-summary.json and corrected call-summary.json, plus publication checks and SHA256SUMS. Avoid publishing full logs, local paths, credentials or unnecessary session detail. Existing infrastructure sufficed.