Hard and soft rules

The two user-facing rule kinds, the five engines behind them, and the tier names the API still serializes.

lawlint has two user-facing rule kinds. The distinction is not a taxonomy for its own sake — it decides whether a check can run offline.

Hard rules

Deterministic checks over the text. No model, no network, no variance between runs: the same document always produces the same findings. Every rule enabled by default is a hard rule.

Soft rules

Inferential checks evaluated by the optional AI judge against a written rubric. They catch what a pattern cannot — whether a hedge actually says what is uncertain — and run only under --judge.

Engines

A rule declares one engine, which decides both how it is evaluated and which kind it belongs to.

Engine	Kind	Tier	What it does
`phrase`	Hard	`static`	Matches regex fragments against the text, optionally suppressed by surrounding context.
`leading`	Hard	`static`	Matches only at the start of a sentence — for openers and transitions.
`density`	Hard	`statistical`	Counts matches per 1,000 words and flags above a threshold.
`statistical`	Hard	`statistical`	Measures a document- or sentence-level metric and flags on the wrong side of a threshold.
`inferential`	Soft	`inferential`	Hands the text to the AI judge with a rubric and worked examples.

The engine has always serialized three tiers — static, statistical, and inferential — and still does. Every --format json payload, every lawlint rules --json entry, and every rule-package field keeps those exact values.

Hard and soft are the explanatory terms layered on top: static and statistical are hard; inferential is soft.

Intent: which rules move the score

Independently of kind, every rule declares an intent.

detection — corpus-validated to distinguish AI-generated from authentic human legal prose. These are the only rules that charge the human-likeness score. This is the default when intent is absent.

style — real writing advice that says nothing about provenance: the Oxford comma, semicolons, sentence length, the Orwell rules. Reports findings, participates in --fix, never moves the score.

Both kinds are reported, and both are fixable. Intent only governs scoring.

Severity

error, warning, or suggestion — set per rule, overridable per project in configuration. The legacy value info is accepted as an alias for suggestion.

Severity drives two things: the penalty weight a detection finding charges against the score, and the exit code. Soft-rule severity is capped at warning or suggestion — an inferential judgement never fails a build on its own.

Selecting rules

lawlint --rules no-em-dash,no-hedging draft.md   # only these
lawlint --disable no-legalese contract.docx      # all but these

Both accept full ids or bare aliases, and both have durable equivalents in .lawlint/config.json. The rule reference lists every built-in rule with its engine, tier, intent, and severity.