Articulate

A local writing-quality and AI-tell detection and editing tool. It flags the prose devices and machine-writing tells that make text read as generated, scores how machine-textured a passage is, and (with an LLM backend) rewrites prose to a plain, skilled standard. The core runs standard-library-only with no network call. Detection quality and writing quality are the goals; a detector score is a benchmark and a byproduct, never something the tool optimizes toward, and it is not an evasion tool.

Documentation

Full docs are in docs/: getting started, a thorough walkthrough, the feature reference, the CLI reference, and the boundaries that say what a verdict and a receipt mean and what they never claim. The house voice spec and the guide to series review and your own voice cover the 0.8.0 voice features.

What it does

• Detect. Flags banned rhetorical devices (antithesis including keyword-free parallel-negation contrast pairs, corrective negation, rule-of-three, em-dashes, filler intensifiers, corporate verbs), current frontier-model register, marketing, and email/blog tells, and Williams/Orwell signals (expletive openers, nominalization density, passive voice, adverb density, cadence uniformity, opener repetition). Emits a graded 0-100 texture score and a clean/flagged gate.

• Adapt by register. A profile system (procedure, commit, research, readme, essay, narrative, and more) sets which findings block. Fiction gates nothing; procedures and essays gate strictly. Profiles resolve from --profile, an in-file writing-profile: tag, or the file path.

• Choose a mode or a genre. A writing mode crosses a domain register with an articulation need, such as memo/argue or technical-docs/explain. The genre axis reads narrative and expressive prose by its own convention: literary-fiction, genre-fiction, ya-fiction, memoir, screenplay, poetry. Under a fiction genre, quoted speech is masked out of the device passes so a character's line is never scored as the author's prose; the craft devices report but never block; and a report-only lexicon flags generation artifacts such as the somatic cliche or the "could not help but" reflexive. Screenplay classifies Fountain roles first, so only action lines face the device gate. Poetry reads by the line and drops the craft-device categories from its report. Run articulate modes to list them.

• Edit. judge reads the judgment-level failures a regex cannot see (confident emptiness, vague abstraction, hedging with no position, weak verbs). fix rewrites to the standard, self-checked against the detector. polish loops until five qualities (concreteness, commitment, economy, rhythm, a restatable fact per paragraph) clear a bar. Gated on writing quality, never a detector score. Use the calling model, a configured backend, or mechanical fixes with no model. Every result names its backend and any failed attempts.

• Give the model a voice. The house voice is a published, versioned voice for AI models: answer first, numbers with denominators, "I" only for what the model did in the session, and no claim to a human life. It is opt-in: articulate house on turns it on, and articulate house off turns it off.

• Build your own voice. articulate voice learn --mine builds a profile from your own writing, on your computer. Compare drafts with it, run the authorship interview, and, when you ask, shape your own draft toward it.

• Review a series. articulate corpus reads several documents together and names repeated title formulas, shared scaffolds, even rhythm and missing perspective, with locations and no replacement prose.

Use

Articulate includes no model. A connected client can call edit_plan, write the rewrite with its selected model, and submit it through edit_submit. The model account and conversation policy belong to that client. Local checks and host edits require no publisher-hosted service or separate model API key.

Version 0.9.0 prepares Windows x64 native ZIP and MCPB packages with a Python runtime included. The source plugin remains available for local Claude Code, Codex and portable MCP hosts with installed Python. Native client installation and marketplace acceptance remain unverified.

# lint (exit 1 when blocked under the file's profile)
python -m articulate.cli check path/to/doc.md --gate
python -m articulate.cli check essay.md --profile essay --verbose
echo "some prose" | python -m articulate.cli score

# library
python -c "import articulate; print(articulate.check_text('...', profile=articulate.profiles.load('research'))['gate'])"

# receipt: a re-derivable verdict (Match / Drift / Unverifiable)
python -m articulate.cli receipt doc.md --profile research > doc.receipt.json
python -m articulate.cli verify doc.receipt.json doc.md   # replay; exit 0/1/2

# content-free audit receipt: replayable, but stores no verbatim text (drop or hash
# the matched substring). For a team that must retain a record without the source.
python -m articulate.cli receipt doc.md --redact drop --reviewer alice > receipts/doc.json
python -m articulate.cli check doc.md --content-free --sarif > doc.sarif  # no substrings

# audit: query committed receipts locally (no server), and re-verify they still hold
python -m articulate.cli audit receipts/                 # recorded verdicts, blocked rules
python -m articulate.cli audit receipts/ --reverify --gate   # exit 1 if a source drifted

# SARIF for CI (GitHub Code Scanning, Azure, reviewdog)
python -m articulate.cli check src/**/*.md --sarif > articulate.sarif

# LSP server (inline squiggles in VS Code, JetBrains via LSP4IJ, Neovim). Stdlib
# only, no dependency. Point your editor's LSP client at:
python -m articulate.lsp_server

# benchmark (regression-gated corpus) and MCP server
python -m articulate.bench
python -m articulate.mcp_server

Editor setup

The LSP server speaks standard LSP over stdio, so any LSP client can drive it. A minimal Neovim registration:

vim.lsp.start({ name = "articulate", cmd = { "python", "-m", "articulate.lsp_server" },
  filetypes = { "markdown", "text", "tex" } })

For VS Code, a thin client that launches the same command as a LanguageClient is all that is needed; no server code lives in the extension.

Scientific and mathematical writing

academic/prove and science-writing/explain target hard technical exposition: stating the idea before the formalism, keeping a roadmap, defining each symbol once. The proof mode does not rewrite by default, because a wrong change to a quantifier order or an inequality direction changes a theorem; it routes to --judge, and --fix is opt-in. On a .tex file the editor masks every math span before a rewrite and splices it back byte for byte, so a formula is never altered.

The boundary is fixed and load-bearing: a clean gate, a low texture score, or a Match receipt means the prose was screened under a named ruleset. It says nothing about whether the theorem is true. A clearly written proof can still be false, and Articulate never checks the mathematics. Correctness comes from referees and proof assistants (Lean, Coq, Isabelle), never from this tool.

Editing from a host or the command line

In an MCP host, fix, judge and polish use sampling only when the client advertises it. Otherwise they return an edit plan for the calling model. The host reads the plan's instructions and masked text, then calls edit_submit with the original text, its rewrite and the plan ID. Articulate checks protected spans and returns the accepted text, gate changes and a host receipt. No second model account is needed. Sampling and host edits follow the host's privacy policy.

articulate plan notes.md --goal fix > plan.json
articulate submit notes.md rewrite.md --plan PLAN_ID --model HOST_MODEL
articulate fix notes.md --backend none --out edited.md
articulate polish notes.md --backend ollama --json

Replace PLAN_ID with the ID in plan.json; the host writes rewrite.md from the plan's masked text. The CLI's auto selection tries configured Anthropic, the Claude CLI, configured OpenAI-compatible endpoints, Ollama and deterministic editing in that order. A failed backend is recorded before the next is tried. See the backend configuration reference.

Privacy

The detector and --backend none never call a model or the network. Anthropic and the Claude CLI send text to their provider; an OpenAI-compatible endpoint receives text at its configured address. Ollama receives text at its configured server. Host editing and MCP sampling share text with the calling host, which may use a remote model. Set ARTICULATE_LOCAL_ONLY=1 to allow only loopback Ollama and deterministic editing; other editor backends are refused before a connection. Plan payloads and edit results contain document text; handle them as source material. A content-free audit receipt keeps no verbatim text: it drops the matched substring and the exact offsets, keeping only which rule fired, its tier and category, and the line. A team can retain and replay a record without storing the sensitive source. Content-free is not zero-leakage: which rules fired and the line remain, which for a closed-vocabulary rule narrows the flagged word to that rule's small public candidate set. A personal voice profile is built only from files you name, holds aggregates and no sample sentence, and stays in your local voice store until articulate voice delete --all. The house-voice brief that enters a model's context is the same published text for everyone. The hash mode keeps a sha256 for an equality check against a known string, so it is dictionary-reversible for those closed-vocabulary rules; use drop when the flagged word must stay secret. An auto-filled reviewer (from $GITHUB_ACTOR) records CI attribution, and a named human sign-off needs an explicit --reviewer. Committed receipts live as long as the repo, with no automatic expiry (bounded retention is a later self-hosted tier). The full documentation covers each surface and the boundaries in depth.

Status

Pre-1.0. The core detector, profile system, writing modes (including the science modes for proofs and technical exposition), the genre axis (fiction, memoir, screenplay, poetry), the editor injection boundary, per-span mixed-authorship verdicts, a sub-threshold "unverifiable" calibration, binary fail-closed input guards, the benchmark, the editor layer, the CLI, the LSP and SARIF surfaces, receipts, the content-free audit receipt, and the MCP server are built into this one package. Version 0.9.0 is prepared for release as articulate-writing; the changelog records what each release added. Optional adapters can use a model you supply; no model is bundled. This release retains the published v0.5.2 detector rules and fingerprint while extending packaging and rewrite guards. It excludes withheld development-branch scanner changes. Retention does not establish fairness, and accepted rewrites can still change meaning.