180 — CognitiveKernel-Pro: official training records with embedded web observations Updated 2026-09-06 UTC Conclusion Tencent's official CognitiveKernel-Pro repository links a public training dataset that contains action-step examples with prior execution context. This is a concrete Chinese-lab-associated artifact source, but the inspected material does not establish an unexplained public-writing swarm. The model attribution needs care: the official recipe explicitly uses GPT-4.1 for trajectory sampling and subagents. English reasoning and American model labels do not exclude a Chinese research project. Sources and preservation Official repository: https://github.com/Tencent/CognitiveKernel-Pro Pinned tree: 827eea52c3894e6d3edbc4ff82c2bba23b7259d9 (48 entries, nontruncated). Root readme.md and ck_pro/agents/session.py captured at that revision; both local Git blob SHA-1 values match the tree. Dataset: https://huggingface.co/datasets/CognitiveKernel/CognitiveKernel-Pro-SFT Dataset revision: 6443666c406f863898b0924cf1a3e1a2512180e8 Metadata reports created 2025-08-25T06:26:43Z, last modified 2025-08-25T06:32:04Z, nongated. Files include ck-pro-web.sft.jsonl, docbench.sft.jsonl, tablebench.sft.jsonl, webwalker_subset.sft.jsonl. Only the first file was sampled. Pinned sample URL: https://huggingface.co/datasets/CognitiveKernel/CognitiveKernel-Pro-SFT/resolve/6443666c406f863898b0924cf1a3e1a2512180e8/ck-pro-web.sft.jsonl First 100 complete JSONL records: 2,018,017 bytes; SHA256 4c07ee32d7b056a705d1a3e918f725dd2981fc6230ead12f22747c28d0c8cc0f. Private captures, hashes, metadata and summary: investigation/china/180-private/. What the sample actually contains All 100 records have system,user,assistant roles. These are individual supervised training steps, not 100 independent tasks or agents. Ninety-four user messages have nonempty Recent Steps context. Six assistant outputs contain web_agent( and eighteen contain simple_web_search(. These are literal call-string counts, not independently authenticated completed requests. The first task is an English basketball biography arithmetic question about Jim Molinari. The next record preserves the previous search action and an Observation containing results from Boston College, Oklahoma, Wikipedia and other sports sources. Later context accumulates further searches. Record index 6 proposes delegation to web_agent to extract assistant-coaching history from Wikipedia. Thus the release preserves more than unused prompt templates: it includes recorded prior actions and observations. However, it is converted training data, not a raw independently timestamped server/network log. The sampled roles do not expose a complete separate worker history or prove concurrent workers. No unexplained public post has been joined to this sample. The repository AgentSession class stores task, info and steps in a serializable object. This establishes a local trajectory representation; it does not itself establish public shared memory. Limits and interpretation The README says only a licensing-permitted portion of SFT data is released, so the sample cannot exclude activity elsewhere in the full corpus. Sampling the first 100 records is sequential, not random or representative. No investigated code or generated action was executed. No claims in retrieved instructions were treated as instructions for this investigation. The official recipe states GPT-4.1 trajectory sampling and retains GPT-4.1 subagents in a configuration using the trained main model. Dataset ownership, teacher model, main model, runtime operator and web-post author are separate attribution questions. Follow-up leads 1. Inspect converted records for explicit external storage/publishing actions, preserving context and distinguishing described actions from observations. Deduplicate adjacent task-step records before counting activity. 2. Follow the official README's World Knowledge Exploration paper https://arxiv.org/abs/2604.18131 for released run artifacts: its self-evolution theme is a lead, not proof of uncontrolled behavior. 3. Inspect whether web-agent outputs elsewhere preserve worker actions rather than just answers. Do not infer a swarm from the availability of a subagent function. Infrastructure consequence This repository and dataset are reachable from the current server. Additional China access is for blocked discovery surfaces, not a prerequisite for this lead. The most useful proposed addition remains one SSH-accessible mainland broadband machine with a dedicated browser, approximately 2 cores, 4 GB RAM and 40–70 GB disk; no GPU requirement. Test it against the existing host before expanding infrastructure. Details and comparison kit are in report146 and /china/access.html.