dsh-grayprint

dsh-grayprint addresses a key challenge for DeepSeek Harness users: objectively evaluating changes in the style of model-generated reasoning text. Many observe subtle differences across model versions but lack tools to quantify them. This plugin provides real-time analysis by examining the openings of paragraphs within reasoning blocks, offering two distinct fingerprint indicators to help users discern whether the text resembles grayprint (earlier version) samples or current version samples, based on evidence rather than direct claims about model capabilities.

Its core capability lies in statistical analysis of paragraph start patterns. It decomposes reasoning blocks into internal paragraphs and checks the first line of each: one fingerprint evaluates the proportion of first-person narration (e.g., I'm planning) versus collective shorthand (e.g., We need answer), yielding a grayprint score; another independently tracks the tendency of Let me (probing) versus We need (planning) openings. Built-in guards, such as detecting Chinese thinking or sample size thresholds, ensure reliable readings. Validation data shows over 94% accuracy in holdout tests, with strong performance on real conversations.

The plugin is ideal for active DeepSeek Harness users, AI researchers, and developers seeking to analyze text style changes objectively. Whether you're exploring model behavior shifts or conducting academic research, it operates locally in the browser for privacy-safe insights. All data stays on your device, and the interface presents statistical details transparently, making the analysis process intuitive and evidence-based.