Chloride233/dsh-cot-profile

README excerpt

dsh-cot-profile 中文 · English Real-time chain-of-thought trajectory profiling for DeepSeek Harness: watch the reasoning stream as it happens, tally signature wording indicators ( let me / we / let's / I ), classify the current session's trajectory family against built-in baselines, and optionally record per-session measurement aggregates. Read this first: honest framing Wording fingerprints describe the (model × assembly) combination — system prompt, tool schema, reasoning effort — not model identity . The underlying research ( xiaobright/modeltest ) shows the same wording pattern across different models when the interface changes (the V4 Flash counterexample: identical we -heavy, let me -free reasoning under the minimal assembly, at a different ability level). yjh051108/dsh-router-standard maps the same lexicon onto a fault line : along the persona axis, V4 Pro behavior collapses into three bands — spec (collective We , let me ≈ 0), a transition band ( mixed : We / The / Let mixing, unstable), and react (first-person The / Let , we ≈ 0). Neither side is universally "stronger" (maintenance tasks favor the spec side; greenfield builds favor the react side). The author's own correctio…

Category