Overview
The ultimate goal of this project is to move beyond passive stylometric analysis toward active tonality derivation — the ability to express an author's writing style in a quantified, communicable form that can:
- Describe a tonality to a human or AI in precise, measurable terms
- Guide a generative AI to embody that tonality
- Validate whether the generative AI is actually reproducing it
Problem Statement
Today, you can prompt an LLM with "Write this in the style of Nevil Shute" — but this is a black box. The LLM may or may not capture the actual stylistic fingerprint of that author, and there is no mechanism to measure conformance.
pystylometry can already extract a rich set of stylometric features: syllable distributions, prosodic patterns, entropy measures, lexical density, punctuation behavior, sentence rhythm, and more. But the missing layer is:
How do we synthesize those features into a form that is useful as both a generative prompt and a conformance target?
Proposed Capabilities
Phase 1: Tonality Report
Generate a structured, human-readable stylometric profile of an author based on a reference corpus. This would include:
- Dominant rhythmic patterns (e.g., "predominantly 1-2 syllable words, low variance")
- Sentence length distribution and rhythm
- Lexical density and register
- Prosodic fingerprint (syllable pattern repetition, stress regularity)
- Entropy signature (predictability, syntactic complexity)
Phase 2: Prompt Generation
Translate the tonality report into a structured LLM prompt — a style guide that instructs the LLM how to write in that tonality, grounded in measurable features rather than vague impressionistic labels.
Phase 3: Conformance Scoring
After generation, run the LLM output through pystylometry and compare its feature vector against the reference author's. Produce a style conformance score that answers: Is the generated text actually writing like Nevil Shute?
Why This Matters
- Makes stylometric analysis actionable, not just descriptive
- Enables iterative prompt refinement — if conformance is low, adjust the prompt and regenerate
- Opens the door to style fidelity benchmarking across LLMs
- Bridges computational stylistics with modern generative AI workflows
Acceptance Criteria
Related
Overview
The ultimate goal of this project is to move beyond passive stylometric analysis toward active tonality derivation — the ability to express an author's writing style in a quantified, communicable form that can:
Problem Statement
Today, you can prompt an LLM with
"Write this in the style of Nevil Shute"— but this is a black box. The LLM may or may not capture the actual stylistic fingerprint of that author, and there is no mechanism to measure conformance.pystylometry can already extract a rich set of stylometric features: syllable distributions, prosodic patterns, entropy measures, lexical density, punctuation behavior, sentence rhythm, and more. But the missing layer is:
Proposed Capabilities
Phase 1: Tonality Report
Generate a structured, human-readable stylometric profile of an author based on a reference corpus. This would include:
Phase 2: Prompt Generation
Translate the tonality report into a structured LLM prompt — a style guide that instructs the LLM how to write in that tonality, grounded in measurable features rather than vague impressionistic labels.
Phase 3: Conformance Scoring
After generation, run the LLM output through pystylometry and compare its feature vector against the reference author's. Produce a style conformance score that answers: Is the generated text actually writing like Nevil Shute?
Why This Matters
Acceptance Criteria
Related