@rhetorlint/rules-en

The English tell-pack for RhetorLint: declarative rules for the families of rhetorical manipulation, aligned to the SemEval-2023 persuasion taxonomy. Data, not code — so any engine in any language can read it.

Rule shape

{
  "ruleId": "agency-hiding.deleted-subject",
  "family": "agency-hiding",
  "technique": "Obfuscation (structural — RhetorLint extension)",
  "type": "structural",
  "detector": "agentless-passive",
  "level": "warning",
  "confidence": 0.7,
  "note": "an agentless passive — who acted is deleted from the sentence",
  "expected": ["(name who did it: 'I/we + verb')"]
}

Check-types (kept small on purpose, the Vale lesson):

| type | how it matches | example families | |--------|----------------|------------------| | lexical | word-boundary match against a terms list | intensifiers, absolutes, softener hedges | | pattern | a regular expression | weasel attribution, whataboutism, rehearsed contrition | | structural | a named built-in detector (currently agentless-passive) | agency-hiding |

Simple regex and word-lists cover most families. The structural check-type is what RhetorLint's deleted-subject / agency-hiding tells need — and the part that wants real grammar awareness over time.

Two optional fields qualify a pattern rule:

| field | meaning | |-------|---------| | caseSensitive | true compiles the pattern with letter case honoured; absent means case-insensitive, which is the default for every other rule. Only @rhetorlint/core >= 0.1.2 reads it. | | minEngine | the lowest core version that honours every field the rule depends on. An engine below that floor may refuse the rule, warn, or drop it; it must not assume the rule still means what it says. |

The same pattern string is compiled by the JavaScript, Python and Go engines, so it may only use syntax all three share. Go's RE2 has no lookaround — (?!…) will not compile there.

The engine floor on shouting.caps

shouting.caps is the one rule about letter case, so it is the one rule that sets caseSensitive: true and declares minEngine: "0.1.2". An engine below that floor ignores the flag and compiles the pattern case-insensitively — where a rule for ALL-CAPS would otherwise match any run of two or more ordinary lowercase words, and bury the real tells.

So the pattern is built to match nothing under a case-insensitive compile. Each shouted word must open and close with [A-Z] around [^\x00-\x40\x5b-\x7f] — everything except the ASCII characters outside A–Z. Compiled case-insensitively that class rejects letters of either case, and the rule falls silent instead of firehosing. Under-marking is the doctrine: on an old engine the tell is lost, never invented. That class also admits accented capitals inside a word, so CAFÉS ARE OPEN is one run.

The seed families

11 rules across 5 of the taxonomy's 7 families:

| family | rules | |--------|-------| | manipulative-wording | intensifier.loaded · hedge.softener · hedge.deniable · weasel.attribution · contrition.rehearsed · lure.free-offer · shouting.caps | | simplification | absolute.universal | | distraction | deflection.whataboutism | | call | urgency.appeal-to-time | | agency-hiding | agency-hiding.deleted-subject |

attack-on-reputation and justification are named in the taxonomy and not yet seeded here.

The human-readable, teachable version — with definitions, worked examples, and SemEval lineage — lives in spec/taxonomy.yaml. This JSON is its compiled, runtime form.

Authoring note

Rules are authored for honesty over coverage: they err toward under-marking, and confidence reflects language-pattern likelihood, not certainty about intent. A rule that cannot be evaluated correctly marks nothing rather than over-marks. New tell families and locales are welcome — keep every rule pointing at a phrase a reader can see.

MIT.