cadence

Parse prose, measure the shape of its sentences, then rewrite a draft to match.

Ask a model to write like you and it copies your topics and your favourite words, which are the easiest things to copy and the least of what makes writing recognisable. What actually distinguishes one writer from another is closer to the skeleton: how long a sentence runs before its main verb arrives, how deep clauses embed, whether ideas get joined or subordinated, how often a claim becomes a noun rather than staying a verb.

Those are measurable, which means a style can be a specification rather than a mood, and a specification can be checked after the fact.

The rule that shapes everything

A rewrite may fix structure. It may not invent content.

Where the analysis finds missing information, a superlative with no comparison set, a passive with a suppressed agent, a pronoun with no recoverable referent, that stays missing and is reported as a gap. Filling it would mean fabricating facts, which is the one failure mode this tool is built to avoid.

That promise used to be a line in a prompt plus the model's own report of whether it had obeyed. A model that fabricates is not a reliable witness to its own fabrication, so fidelity.py now checks it mechanically.

Install

pip install "cadence-writer[llm]"   # the import name is `cadence`
cadence download-model

The spaCy model is a separate step because PyPI rejects packages that declare a dependency by URL, and the model is only distributed that way.

Use

cadence analyze notes.txt                    # trees and findings, no rewrite
cadence report notes.txt                     # everything, including fidelity
cadence rewrite notes.txt --backend rules    # deterministic, no API call
cadence profile samples/*.txt                # the specification for a corpus
cadence restyle draft.txt --corpus samples/*.txt
cadence score out.txt --corpus samples/*.txt --source-text draft.txt

import cadence

profile = cadence.build_profile(my_documents, name="me")
result = cadence.generate(draft, mode="profile", profile=profile,
                          task="restyle", api_key=key)

result.similarity          # voice match, 0 to 100
result.fidelity.score      # content kept, 0 to 100
result.fidelity.introduced # anything invented, which should be empty

As an MCP server

pip install "cadence-writer[mcp]"
cadence mcp

That serves the library over stdio for Claude Desktop, Claude Code, or any other MCP client, so a model can measure a voice, restyle a draft, and check fidelity as tools. Nothing leaves the machine except the model call the restyle tool makes. In Claude Code:

claude mcp add cadence -- cadence mcp

Tools: measure_voice, analyze_text, restyle, check_fidelity, score, repair_structure. Each is one library function with its arguments flattened; where a tool takes a corpus, it takes the documents themselves, so the server holds no state a client has to manage.

Two numbers, and what neither can tell you

Voice match compares sentence shape against the corpus: length, embedding depth, clause mix, subordination, function words. It cannot tell you whether the result sounds right.

Content kept compares names, numbers, dates, negations and coverage between the draft and the rewrite. It cannot tell you the meaning survived. Two passages can share every anchor and still say different things.

Both are floors. They catch the failures worth catching and they are honest about the ones they miss.

What the diagnostics find

Findings are defects. Observations are register facts (dates used as attributive modifiers, container-noun inventory) — true of the text, not wrong with it. They are reported separately on purpose.

Layers

syntax.py       PTB tags, dependency trees, constituency projection
repair.py       parse repairs for note-register English (R1–R5)
diagnostics.py  the detectors
profile.py      corpus measurement, divergence, the generation spec
generate.py     constrained generation, and the verify loop
rewrite.py      rules backend (deterministic) + Claude backend
fidelity.py     did the rewrite keep the content, measured not asserted
scores.py       the two numbers, in one call
llm.py          credentials, one call, token usage

On the constituency layer

The dependency parse is real. The constituency tree is a head projection of it, not a native PTB parse. It reproduces PTB conventions that follow from dependencies (flat base NPs, -SBJ/-PRD function tags, PP/ADJP/SBAR labels) and cannot reproduce what only a constituency treebank knows (traces, empty categories, licensed UCP nodes). It is labelled as projected everywhere it appears.

On remove-slop

The remove-slop rule set (slop.py) is not a mode or a checkbox. It runs in the same pass as the structural detectors, and its rules are appended to every rewrite and generation prompt. A dash used as a copula is a structural finding (PUNCT_COPULA) and a slop finding (SLOP_DASH) at the same time; splitting them would have forced the reader to reconcile two reports of one sentence.

Most of the rules are mechanically checkable, which is why they belong here rather than in prompt text alone. "Uniform rhythm is a stronger tell than any single word" is a claim about variance, and variance is what profile.py already computes, so it is measured rather than eyeballed.

Where the two layers meet. A style profile is descriptive and remove-slop is normative, so they can disagree. The resolution is fixed, not a preference:

1. The profile owns structure (length, depth, clause mix, subordination).
2. remove-slop owns surface machine tells (banned phrases, dashes, applause).
3. Where they overlap, the ban wins, and the override is printed rather than left for the reader to notice. Banned features are still measured — the profile stays a complete description — but they are marked NOT A TARGET in the spec and never handed to a generator to reproduce.
4. On rhythm the two agree, with a floor: if a corpus is itself uniform, the spec says do not reproduce that uniformity, because copying it would copy the strongest machine tell there is.

On the parse repairs

en_core_web_sm is trained on newswire and fails on this register in specific, reproducible ways. Five repairs run before any diagnostic, and every one that fires is reported:

• R1 ISO dates and hyphenated compounds kept as single tokens
• R2 the category read is… — head noun misread as a verb
• R2b subject-like arc landing on a nominal head → compound
• R3 one — NN when pronominal, CD when a partitive numeral
• R4 accusative pronoun parsed as a subject (liked him)
• R5 hyphenated premodifier tagging: noun-noun → NN, prefix+noun → JJ

Sentences also carry a coarse parse-confidence signal — a list of structural reasons to distrust the parse, not a probability.

Tests

pytest -q

Nothing in the suite touches the network. Every expectation was verified by hand before it was automated, so a failure means the tool has drifted from an analysis a human checked.

License

MIT.