218 — Claw-Eval: aggregate results and three distinct web-service behaviors Reviewed September 6, 2026 UTC. No benchmark or investigated code executed. RESULT The public Claw-Eval repository and dataset supply task material, trace schemas and aggregate scores, but this pass recovered no actual Chinese-model run archive. Source confirms that similar web-tool routes can return fixtures, real fetched pages, or fetched pages modified locally by injection tests. This distinction matters when interpreting any future trace or public-site attribution. PUBLICATION CHECK https://github.com/claw-eval/claw-eval https://huggingface.co/datasets/claw-eval/Claw-Eval/tree/main https://arxiv.org/abs/2604.06132 The paper describes 300 tasks with execution traces, audit logs and environment snapshots used for grading. That describes recorded evidence, not necessarily publicly downloadable evidence. GitHub recursive tree: 5680b8b11ff2ee5dd2b07b89086a29a5c5c984d7, 1,972 entries, non-truncated. Four inspected source files match their Git blob hashes. The Hugging Face recursive listing returned 33 entries without a next-page link: model-score YAMLs, task Parquet files, a 2,879,916,129-byte fixtures archive, assets and metadata. Archive contents were not inspected and are not ruled out as containing additional material. No bulk archive was downloaded on the assumption that it contained traces. The inspected deepseek_v4_pro.yaml has two aggregate entries dated April 23: general and multi-turn scores linked back to the leaderboard. It contains no message history or external action receipt. The model label and date are publisher metadata, not authenticated runtime identity or independent chronology. THREE WEB MODES, SAME-LOOKING ROUTES Pinned files under mock_services: 1. web/server.py: /web/fetch looks up the exact URL in fixture pages and returns a configured response, including a newly generated fetched_at timestamp. Unmatched URLs return a fixture-level 404. A URL plus timestamp in this output does not prove a network fetch. 2. web_real/server.py: after checking its cache, /web/fetch uses httpx.Client.get with redirect following, extracts page content and records an audit response. Its User-Agent is Mozilla/5.0 (compatible; AgentEval/1.0). This is implemented real reading, not proof that any historical run used it. 3. web_real_injection/server.py: performs a real GET, builds/caches the clean response, then calls _inject_content on the response sent onward. Consequently injected text associated with a genuine public URL need not have existed on that public page. It may have been added inside the evaluation service. The inspected fixture service's /web/notify appends a record to local memory and returns status sent. It does not deliver an external notification. No service endpoint was invoked in this review. TRACE SCHEMA, NOT A RUN src/claw_eval/models/trace.py defines trace_start (task/model/persona), message, tool_dispatch (endpoint/request/response/latency), audit_snapshot, media_load and trace_end event structures. To interpret a future export, join the task configuration and service implementation to dispatch records; then seek independent destination evidence for any claimed public write. A service POST used to request a web read must not be confused with POST publication to the page's own host. Source URLs: https://github.com/claw-eval/claw-eval/blob/5680b8b11ff2ee5dd2b07b89086a29a5c5c984d7/mock_services/web/server.py https://github.com/claw-eval/claw-eval/blob/5680b8b11ff2ee5dd2b07b89086a29a5c5c984d7/mock_services/web_real/server.py https://github.com/claw-eval/claw-eval/blob/5680b8b11ff2ee5dd2b07b89086a29a5c5c984d7/mock_services/web_real_injection/server.py https://github.com/claw-eval/claw-eval/blob/5680b8b11ff2ee5dd2b07b89086a29a5c5c984d7/src/claw_eval/models/trace.py BOUNDED SEARCH Searches for AgentEval/1.0 outside GitHub and Claw-Eval plus traces/DeepSeek/download returned documentation, unrelated similarly named projects and aggregate descriptions, not a verified raw run. The User-Agent is copyable and not a Chinese-lab attribution token. No local full-corpus marker scan was run this pass. A returned dataset-discovery page lists TeichAI/DeepSeek-v4-Pro-Agent with a Traces label; this is an unverified next lead, not evidence that it contains real DeepSeek execution or swarm activity. NEXT Inspect the explicitly public DeepSeek-labelled trace dataset's card, file schema and small samples. Prefer actual dispatch/action records to another benchmark architecture review. Keep Claw-Eval's mixed real/fixture service distinction as a control when reviewing traces elsewhere. CAPTURES 218-private: GitHub tree, Hugging Face file listing, four Git-verified source files, one aggregate model YAML and SHA256SUMS. Earlier homepage/bundle captures remain in 216-private. No credentials, private logs, benchmark execution or target writes used.