Limatus

Limatus is a diagnose-and-steer quality loop for AI-generated content — prose today, web pages soon — built to catch generic AI slop and enforce a specific publication or product's own style rules, not just generic patterns.

Press Release

Anthus AI Solutions introduces Limatus, an open-source tool that tells writers exactly where AI-generated prose sounds generic — without rewriting it for them

Teams publishing AI-assisted content face a recurring problem: drafts that are technically correct but sound like every other AI-generated draft — hedge words, empty lead-ins, uniform sentence cadence, unsupported certainty, and a voice that matches no one's actual house style. Fixing this by hand doesn't scale, and letting a model "just rewrite it" trades one set of generic patterns for another.

Limatus is a command-line tool that closes this gap in two separate steps. limatus diagnose reads a draft against a publication's own style profile — its voice, its lexicon, its banned phrases and patterns — and returns a structured list of findings: vague claims, empty lead-ins, uniform cadence, list-shaped prose, unsupported certainty, redundancy, and voice mismatches. It never rewrites anything; findings are signal, not verdicts. A human or agent then marks each finding skip, rewrite, delete, or keep, and only for the findings marked "rewrite" does limatus options generate a small set of constrained rewrite candidates — never a single mandated replacement, and never a fabricated fact, statistic, or anecdote.

"We built this because we were publishing AI-assisted content ourselves and got tired of the same tells showing up in every draft," said the Limatus team at Anthus AI Solutions. "The insight wasn't that AI writing is bad — it's that treating every publication's voice as the same generic 'professional tone' is what makes it read as AI-generated. Limatus makes house style an explicit, versioned, machine-checkable artifact instead of something an editor has to remember and enforce by hand every time."

Limatus is extracted from the editorial engine Anthus AI Solutions built and ran across dozens of real articles on its own Anth.us and Pilobolus publications, and is available today as an MIT-licensed, publication-agnostic Python package at github.com/AnthusAI/Limatus.

FAQ

Is Limatus an AI-content detector? No. Limatus doesn't try to guess whether text was written by a model — it checks text against a style profile you define, regardless of who or what wrote it. It deliberately refuses to accept "detector scores" as part of a style profile; the goal is to enforce your voice, not to pass someone else's AI-detection heuristic.

Does Limatus rewrite my content for me? Not automatically, and not without a human or agent explicitly opting each finding in. limatus diagnose only reports findings — it never touches the draft file. limatus options only generates rewrite candidates for findings you've separately marked "rewrite," and always offers more than one option so the choice of which to accept stays with an editor.

What does a "style profile" look like? A YAML file naming a publication's voice, audience, tone, sentence style, structure, a preferred/avoided lexicon, evidence rules, and five to ten reference samples of real approved prose. It's the same profile used to catch drift, so it's meant to be version-controlled alongside the content it governs.

Why is this a separate package from Papyrus? Papyrus is Anthus AI Solutions' own content and newsroom system, tightly coupled to Anth.us's publishing pipeline. Limatus is the general-purpose diagnose engine underneath it, extracted so any publication or product — not just Anthus's own — can adopt it without adopting the rest of Papyrus.

What's next? A second capability is planned: a usability and accessibility diagnose loop for AI-generated web pages, built on Playwright and axe-core, checking WCAG 2.2 A/AA compliance across themes, viewports, and interactive states — the same diagnose-first, never-auto-fix philosophy applied to markup instead of prose.

Is it on PyPI yet? Not yet — see Status below.

Status

This package was newly extracted from Anth.us's Papyrus content system and is under active development. It is not yet on PyPI.

Installation

pip install -e ".[dev]"

(PyPI installation via pip install limatus will be available once the package is published — see Status above.)

Usage

limatus scan --draft path/to/draft.md --profile path/to/style-profile.yml
limatus decide --finding-id finding-0123456789abcdef --decision rewrite \
  --note "needs concrete options" --decisions path/to/decisions.json
limatus options --draft path/to/draft.md --profile path/to/style-profile.yml \
  --diagnosis path/to/diagnosis.json --decisions path/to/decisions.json \
  --skill path/to/rewrite-skill.yml
limatus compare --profile path/to/style-profile.yml --baseline path/to/draft.md \
  --candidate path/to/candidate-a.md --candidate path/to/candidate-b.md
limatus apply --original path/to/original.md --working-copy path/to/working.md \
  --options path/to/options.json --finding-id finding-... --option-id option-... \
  --anchor "exact span text"
limatus headline options --job title --working-copy path/to/working.md \
  --profile path/to/style-profile.yml --skill path/to/rewrite-skill.yml
limatus headline options --job subtitle --working-copy path/to/working.md \
  --profile path/to/style-profile.yml --skill path/to/rewrite-skill.yml
limatus eval --manifest features/fixtures/editorial-diagnosis/editorial-corpus/manifest.yml
limatus canary --manifest features/fixtures/editorial-canary/manifest.yml

From Python, the same read-only diagnose step is available without touching the draft file:

from limatus import load_config, diagnose

config = load_config("path/to/style-profile.yml")
result = diagnose(open("draft.md").read(), config=config)

diagnose returns structured findings only; it does not modify the draft.

For a copy-editing agent, the intended loop is: scan (findings only, no steering decisions) → record decisions (skip / rewrite / delete / keep / add, via SDK or a decisions JSON file) → options for rewrite findings only → compare full-draft previews of each patch option (or regression compare after a trial edit) → apply exactly one human-chosen option to a working copy. When the style profile defines a headline pass (when: afterBody), run headline options for title then subtitle on the working copy after body edits are applied — each pass replaces only that YAML field. Nothing auto-applies a compare winner.

Run limatus --help for the full command reference.

Advisory revision verification

After a human explicitly applies a selected option to a separate working copy, limatus verify compares the original and working drafts without changing, publishing, or automatically applying either one. It reports four quality dimensions (specificity, clarity, audience fit, and voice match), penalties for redundancy, unsupported claims, and factual-change risk, plus evidence-backed findings for deleted claims, duplicated ideas, residual boilerplate, and new unsupported or factual claims.

The result includes inspectable weights: each positive quality dimension is weighted 0.25; redundancy is -0.10; unsupported claims and factual-change risk are each -0.20. The default acceptance threshold is 0.05 normalized net improvement. The recommendation is advisory and is accept only when the threshold is met and the working draft has no more unsupported claims than the original. No detector score is accepted as input or emitted in output.

Portable style profiles

A profile is a versioned YAML or JSON document. The profile below is complete: required fields describe the publication voice and evidence policy, while checks, rules, and density make the optional controls explicit.

schemaVersion: 1
publicationKey: example-publication
voice:
  name: Practical engineering voice
audience: Engineers evaluating tools and operational risk.
tone:
  - Conversational but precise.
sentenceStyle:
  - Prefer active voice and concrete nouns.
voicePatterns:
  - Lead with the practical stake.
structure:
  - State the claim, then its evidence and limits.
lexicon:
  prefer: [inspect, verify, latency]
  avoid: [game-changing, leverage synergies]
evidenceRules:
  - Cite primary sources and label uncertainty.
referenceSamples:
  - id: sample-one
    title: Approved sample one
    url: https://example.com/articles/one
    path: reference-samples/sample-one.md
  - id: sample-two
    title: Approved sample two
    url: https://example.com/articles/two
    path: reference-samples/sample-two.md
  - id: sample-three
    title: Approved sample three
    url: https://example.com/articles/three
    path: reference-samples/sample-three.md
  - id: sample-four
    title: Approved sample four
    url: https://example.com/articles/four
    path: reference-samples/sample-four.md
  - id: sample-five
    title: Approved sample five
    url: https://example.com/articles/five
    path: reference-samples/sample-five.md
checks:
  informationDensity: true
  uniformCadence: false
rules:
  bannedPhrases: [at the end of the day]
  bannedIntensifiers: [very]
  bannedPatterns:
    - pattern: "\\bseamless\\b"
      message: Prefer a concrete description of the integration.
  contrastCap: 2
density:
  minWords: 400
  minLexicalDensity: 0.45
  maxGzipRatio: 0.35

Reference path values are resolved relative to the profile file, not the shell's current directory. Keep five to ten approved samples beside the profile (or use paths such as ../samples/approved.md). YAML and JSON use the same field names; features/fixtures/editorial-style-profile/portable-profile.json is a checked-in JSON example.

The checks mapping can enable or disable individual diagnose checks. If it is omitted, every check is enabled. rules is optional and defaults to no custom rules. density is optional and defaults to minWords: 400, minLexicalDensity: 0.45, and maxGzipRatio: 0.35. Unknown control names and invalid values fail profile validation before a draft is analyzed.

Optional editorialAim is a non-empty string describing how the publication balances priorities (for example, cite-first versus ordinary-language clarity). When set, judge and rewrite-option prompts include it so steering can reflect that tradeoff.

Optional compare controls candidate ranking hard disqualifiers. If compare is omitted, compare.hardConstraints defaults to unsupported_claims_increase (candidates that add unsupported claims cannot rank first). Set compare.hardConstraints to an empty list to disable hard disqualification on those axes. Only constraint names listed in the profile are applied; unknown names fail validation at load time.

Limatus is style-focused, not an AI detector. Profile documents must not contain detector scores or detector-related fields; the loader rejects those keys, including nested keys.

From the repository root, run the portable fixture against a checked-in draft:

limatus diagnose \
  --draft features/fixtures/editorial-diagnosis/sloppy-draft.md \
  --profile features/fixtures/editorial-style-profile/portable-profile.json \
  --output /tmp/limatus-diagnosis.json
cat /tmp/limatus-diagnosis.json

The command leaves the draft unchanged and writes validated diagnostic JSON. The same command works with a YAML profile by changing only the --profile path, for example features/fixtures/editorial-diagnosis/style-profile.yml.

Offline editorial eval corpus

limatus eval runs a deterministic, network-free regression over checked-in drafts listed in a YAML manifest. Each row names a draft path, expected finding kinds (expectKinds), optional maxFindings, and option-safety checks so diagnose output never embeds rewrite payloads. The default smoke corpus lives at features/fixtures/editorial-diagnosis/editorial-corpus/manifest.yml.

limatus eval --manifest features/fixtures/editorial-diagnosis/editorial-corpus/manifest.yml

CI runs the same command on every pull request (see .github/workflows/ci.yml).

Aggregate metrics

After each row, limatus eval prints kind-level precision and recall by comparing diagnosis finding kinds to each manifest row's expectKinds (gold positives for that draft). True positives are expected kinds present in the diagnosis; false negatives are expected kinds missing (those still fail the row); false positives are extra diagnosis kinds not listed in expectKinds (they do not fail sloppy mustFail rows, but they lower precision). On mustPass rows with empty expectKinds, every surfaced kind counts as a false positive. Micro-averaged ratios use the usual TP/(TP+FP) and TP/(TP+FN) formulas; when both numerator and denominator are zero, the metric is reported as 1.0 (a clean negative with zero findings is perfect).

When the manifest includes an options list pointing at checked-in options JSON, the summary also reports unsupportedClaimRate: the fraction of rewrite option candidates with factVerificationRequired: true. Those fixtures are validated with the same option-safety forbidden-key checks as diagnosis output. The summary line judgePromptVersion=… pins the editorial judge prompt revision used elsewhere in the product, even though offline eval never calls a live judge.

Turning a reject into a fixture

When copy-edit is rejected or a false positive is confirmed in production:

1. Copy the draft (or a minimized excerpt that still reproduces the behavior) into features/fixtures/editorial-diagnosis/ — for example must-fail/ for drafts that must surface specific findings, or beside the corpus under editorial-corpus/ for clean negatives.
2. Add a mustFail or mustPass row to the manifest with id, path, expectKinds, and for failures a non-empty expectTerms list that appears in the diagnosis JSON. Set maxFindings: 0 on mustPass rows that should stay clean.
3. Run limatus eval --manifest … locally until the new row passes.
4. Commit the draft and manifest change together.

Agents should follow the same loop when editorial diagnosis regresses in the field.

Testing

Limatus's test suite is written in Gherkin and run with Behave:

behave

Pytest unit tests:

python -m pytest tests -q

Judge calibration canary (fixture judge, no API key):

limatus canary --manifest features/fixtures/editorial-canary/manifest.yml

License

MIT — see LICENSE.