v0.7.0

Cadence v0.7.0

Released October 26, 2026. Scores change from v0.6.0.

Highlights

• A new signal: vocabulary variety. AI prose repeats words less than people
  do. Text of 40 words or more gains up to 21 points when its moving-average
  distinct-word ratio passes 0.93.
• Score the clipboard from any app. In the desktop app, copy text and press
  Cmd+Shift+. (Ctrl+Shift+. on Windows and Linux).
• Modern human text in the benchmark. 80 passages written before ChatGPT's
  launch now guard against rules that only flag modern writing.

Accuracy

The rule was tuned on development data, then scored once on a held-out set
that was frozen before the work began:

detector	AI caught	human left clean
v0.6.0	2 of 35	30 of 35
v0.7.0	5 of 35	30 of 35

Recall is still low; most AI text without the stock tells grades A. That
held-out set also showed more human false alarms than earlier sets. Uniform
rhythm and triads cause most of them, and one triad match counts "even" as a
list item. Those fixes wait for a new frozen set, so this result stays honest.
The exact reports are in benchmark/reports.

Upgrade notes

Scores from 0.6.0 and 0.7.0 are not comparable, and results now carry
detectorVersion: "0.7.0". Results also gain metrics.vocabVariety and
breakdown.structural.variety. If you store scores or gate on a cutoff, pin
the version:

npx --package=cadence-deslop@0.7.0 cadence-install-skill --agent kimi
npx --package=cadence-deslop@0.7.0 cadence-install-codex

Replace kimi with zcode, claude-code, opencode, or grok as needed.

Desktop installers

System	File
macOS (Apple silicon and Intel)	.dmg
Windows	.msi or .exe
Linux	.deb, .rpm, or .AppImage

The builds are not code-signed yet. On a Mac, open System Settings → Privacy &
Security and choose Open Anyway. On Windows, choose More info → Run
anyway. Cadence_universal.app.tar.gz is a zipped copy of the Mac app that
most people can ignore.

v0.6.0

Cadence v0.6.0

Released October 21, 2026. Scores are unchanged from v0.5.0; no detector rule
changed.

Highlights

• A desktop app for macOS, Windows, and Linux. It opens the score page in a
  native window and works offline. Paste text, drop a .txt, .md, or .html
  file, or open one with Cmd/Ctrl+O. Download the installer for your system
  from this release.
• Every score names its detector. Results carry detectorVersion, so a
  saved score says which rules produced it.
• Releases publish from a tag. npm packages now come from GitHub Actions
  with provenance, through npm trusted publishing.

Desktop installers

System	File
macOS (Apple silicon and Intel)	.dmg
Windows	.msi or .exe
Linux	.deb, .rpm, or .AppImage

The builds are not code-signed yet. On a Mac the first launch is blocked: open
System Settings → Privacy & Security and choose Open Anyway. On Windows,
choose More info → Run anyway.

Agent installs

npx --package=cadence-deslop@0.6.0 cadence-install-skill --agent kimi
npx --package=cadence-deslop@0.6.0 cadence-install-codex

Replace kimi with zcode, claude-code, opencode, or grok as needed.
Run npm installers from your project, not from inside the Cadence checkout.

Accuracy

Unchanged. On the 0.5.0 held-out set the detector catches 3 of 34 AI samples
and leaves 31 of 32 human samples clean. A new held-out set,
benchmark/heldout-0.6.json, is frozen and unscored until the next rule
change. See the benchmark.

v0.5.0

Cadence v0.5.0

Released October 16, 2026. Extension-store submissions remain separate from
the npm package and GitHub release.

Highlights

• Scores change. The triad rule now fires only on exact three-item lists, and
  three- or four-sentence texts can lose points for flat rhythm.
• The benchmark grew to 122 regression samples, plus a 66-sample held-out set
  that was never used for tuning.
• On held-out text the detector catches 3 of 34 AI samples (8.8%) and leaves 31
  of 32 human samples clean (96.9%). The 0.4.0 detector caught none of them.
• Node 20 or newer is required.

Upgrade notes

Scores from 0.4.0 and 0.5.0 are not comparable. Lists of four or more items
no longer add triad points, so some human prose scores lower. Short text with
uniform sentence lengths can now pick up to 15 points. If you store scores or
gate on a cutoff, pin the version:

npx --package=cadence-deslop@0.5.0 cadence-install-skill --agent kimi
npx --package=cadence-deslop@0.5.0 cadence-install-codex

Replace kimi with zcode, claude-code, opencode, or grok as needed.
Run npm installers from your project, not from inside the Cadence checkout.

Accuracy, stated plainly

Recall stays low. The score flags prose that uses the common tells. It is not
an authorship test, and most AI text that avoids those tells grades A. Run
node benchmark/bench.mjs --corpus benchmark/heldout.json to reproduce the
held-out figures. See the benchmark.

Validation before publishing

Run the release checklist on the exact commit to be tagged.
The package smoke check installs from the actual npm tarball. It does not prove
discovery in a live agent session.

v0.4.0

Cadence v0.4.0

Released September 20, 2026. Extension-store submissions remain separate from
the npm package and GitHub release.

Highlights

• Install Cadence as a Grok Build CLI skill at ~/.grok/skills/cadence.
• Pack the detector for a Grok Bot /workspace folder, then save a private
  skill in the app. This is not a Grok Bot marketplace listing.
• Export a local xAI function-calling bridge with npm run build:grok.
• grok.com chat still cannot load local skill folders.

Upgrade notes

Detector scoring is unchanged from v0.3.0. The new surfaces add hosts, not
rules. Manifests share version 0.4.0. The unrelated README Auto Update
marketplace entry keeps its own version.

Run npm installers from your project or another directory, not from inside the
Cadence source checkout.

Grok Build

npx --package=cadence-deslop@0.4.0 cadence-install-skill --agent grok

Or from a checkout: npm run install:grok. Start a new grok session and use
/cadence. See the Grok guide.

Grok Bot

From a checkout:

npm run grok-bot:pack

Copy the packed folder to /workspace/Cadence on the Bot computer, score a
file, then send SAVE_SKILL.md. See Grok Bot.

Other agent hosts

npx --package=cadence-deslop@0.4.0 cadence-install-skill --agent kimi
npx --package=cadence-deslop@0.4.0 cadence-install-codex

Replace kimi with zcode, claude-code, or opencode as needed.

Validation before publishing

Run the release checklist on the exact commit to be tagged.
The package smoke check installs from the actual npm tarball. It does not prove
Grok Build discovery or a live Grok Bot run.

Cadence v0.3.0

Cadence v0.3.0

Released September 10, 2026. Extension-store submissions remain separate from
the npm package and GitHub release.

Highlights

• Install Cadence as a Codex skill once per user or per project. It bundles the
  existing rules with the voice profiles and deterministic detector. No manual
  AGENTS.md copying and no npm download for each score.
• The same install-once workflow now targets Kimi Code CLI, ZCode Agent,
  standalone Claude Code, and OpenCode. It uses the host's selected model and
  leaves provider settings untouched.
• Rewrites measure before and after, preserve markup, and report remaining tells.
• Inspect exact score contributions and the documented scoring formulas.
• Use the expanded benchmark reports as regression checks, not authorship proof.
• Oversized inputs fail visibly rather than leaving old scores in the UI.

Upgrade notes

The detector API still exports analyze() and the same CLI names. The returned
breakdown is additive. breakdown.total can exceed 100; score remains capped
at 100. Unicode ellipses now affect sentence splitting and can change rhythm
scores. Contraction-style negation pivots are also scored differently than in the
prior release. See the changelog for the complete list.

The analyzer rejects text over 5 MiB. The CLI checks raw files and streamed stdin;
URL and extracted text checks happen after retrieval/extraction. Do not treat this
as a download/decompression security boundary.

All Cadence integration manifests now use 0.3.0. Chrome and Gemini previously used
0.1.0. The unrelated README Auto Update marketplace entry keeps its own version.

Codex installation

Available from this checkout now:

npm run install:codex
# Or install for one repository only:
npm run install:codex -- --project /path/to/project

After v0.3.0 is published:

npx --package=cadence-deslop@0.3.0 cadence-install-codex

See the Codex guide for paths and usage.
Existing modified skills are never overwritten; back them up outside the discovery
tree before upgrading.

Other agent hosts

npm run install:kimi
npm run install:zcode
npm run install:claude
npm run install:opencode

The npm command is
cadence-install-skill --agent <host>, with kimi, zcode, claude-code, or
opencode as the target. See the agent guide
for the exact discovery paths and invocation syntax. ZCode project installation
is a UI import, not a guessed .zcode/ project directory.

The Codex integration and multi-agent follow-up are now merged on main.
The npm package includes these integrations. Model-provider setup remains separate.

Validation before publishing

Run the release checklist on the exact commit to be tagged.
The package smoke check installs from the actual npm tarball, verifies the bundled
version and voices, and scores Markdown without a model. The optional Codex check
verifies actual skill discovery with an isolated home directory. Neither is a
substitute for reviewing a model-produced rewrite.

The tarball check covers all five install targets. Native client behavior and
model/provider compatibility must also be checked in the clients you use; a
filesystem install test is not proof of an agent-generated rewrite.
npm run check:kimi adds optional native Kimi user/project discovery checks
without starting a model session. The analogous Codex check remains available.

Experimental work and limits

The Kaggle notebook includes dry/smoke/train controls and downloadable artifacts,
but full GPU training has not been validated for this release. Its current split
sorts by source then slices rows; it does not guarantee source-group isolation.
Dry mode reuses sample fixtures, so its generated writeup is not trained-model
evidence. Do not publish those metrics as an adapter benchmark.

External dataset documentation is a research checklist, not approved ingestion.
No new dataset or trained model is included in the npm package. A Cadence score
alone cannot validate truth, preserved meaning, or writing quality.

The optional .cadence.json configuration from issue #18 is deferred. Use request
instructions and the detector's --max flag for a score gate in this release.

Cadence v0.2.0

Cadence goes multi-surface: the same writing skill now runs in Claude Code,
a regular Claude conversation, Codex, Gemini CLI, and DeepSeek, alongside the
Chrome extension and the npx detector. A ninth voice ships, and a bare
/cadence now guides you to the right command.

Added

• Guided /cadence. Invoked with no command, it now shows a menu that routes
  you by intent; when a command needs a voice and you haven't named one, it lists
  all nine and lets you pick.
• More input formats. The detector now reads .html (scored as visible text),
  .docx (parsed from the Word ZIP), .epub (a book's XHTML chapters), and live
  http(s) URLs — on top of .txt, .md, and .pdf. PDF extraction was also wired
  into the CLI itself.
• Repo scan. Point cadence-deslop at a directory to score every prose file
  under it, ranked worst-first, with an average and the worst offender. Skips
  node_modules, dist, .git, and the like.
• --prose-only (Markdown-aware scoring, skipping code/quotes/tables) and
  --max <n> (a configurable exit gate).
• Three more voices: seminar (a demystifying literature-seminar register),
  dispatch (narrative science-journalism), and column (a calm analytical essay).
• A ninth voice, kin — a parent's unsparing letter to a child, every truth
  anchored in the body, landing on a short imperative. Learned from an epistolary
  memoir register.
• A Chrome extension (extension/) — score any text for AI tone right in the
  browser, by popup or right-click. The detector runs locally and is generated from
  the same deslop.mjs so it can't drift (a test enforces it).
• Codex support. The skill now ships an AGENTS.md beside SKILL.md, so the same
  folder drives Codex as well as Claude Code. The detector is already portable —
  npx cadence-deslop runs in any shell — and the AGENTS.md carries the writing laws
  and voices a Codex agent needs to recast and write.
• Gemini CLI extension. An installable extension under integrations/gemini/
  (gemini-extension.json + GEMINI.md). Symlink it into ~/.gemini/extensions/ and
  the voices and writing laws load every session; the detector runs via npx as usual.
• DeepSeek skill under integrations/deepseek/ — a markdown Skill for DeepSeek's
  drawer, with the writing laws and voices. The detector runs via npx in a terminal.
• npm run build:claude-skill. Packages a self-contained cadence-skill.zip
  (SKILL.md + scripts + voices, paths rewritten) for upload to claude.ai's Skills panel
  — so Cadence works in a regular Claude conversation, not just Claude Code.
• Docs: a consolidated MANUAL (install, workflows, full reference), a repo-scan
  tutorial, and CI that re-scores the project's
  own docs on every push and fails if any drops below grade A.

Cadence v0.1.0 — write like a human, prove it

First public release of Cadence — a Claude Code skill that makes AI agents write like a human, in a voice you choose, with a detector that shows its work.

Highlights

• /cadence skill — 5 commands: write, learn, recast, deslop, voices.
• De-slop detector — deterministic, zero dependencies. Scores prose 0–100 and names every AI tell (uniform rhythm, hollow-confidence words, triads, negation pivots, cliché openers).
• 5 seed voices — counsel, reckoning, measured-academic, plain, punchy.
• Pure-Node PDF extraction for learning voices — no pypdf, no native deps.
• Standalone CLI — npx cadence-deslop draft.txt (also --json, --strict).
• 21 tests, its own landing page, and a logo.

Install

Plugin: /plugin marketplace add wuisabel-gif/Cadence then /plugin install cadence@cadence
Detector only: npx cadence-deslop draft.txt

Full notes in CHANGELOG.md.