The rules

Vocabulary — The Core Lexicon

The word list several rules quietly depend on. G3 (drop a verb’s preposition), S2 (phrasal → plain), S3 (preferred sense), and S6 (regular collocations) each quote a per-word list to do their job. Without it those rules are promissory notes — statable in prose, but not verifiable or reversible.

Source of truth

The data lives in ../tools/data/lexicon.json, not here. This file is its schema documentation, a set of representative highlights, and a coverage statement — kept honest by tools/test/vocabulary.test.ts, which parses the highlight tables below and the coverage counts and asserts both against lexicon.json itself. If this file and the data disagree, the data wins and the test fails.

../tools/src/core-lexicon.ts loads lexicon.json and builds everything the tooling consumes from it:

• toAbolishedEntries() feeds the linter/translator dataset (dataset.ts) — dropped-prep and phrasal-verb forms the same way every other abolished form is tracked.
• buildPhraseTransforms() feeds the forward translator (translate.ts): every inflected surface form of a drop-ruling verb or phrasal verb, matched longest-first.
• buildPrepRestorations() feeds the reverse translator (reverse.ts): the canonical preposition each drop verb restores, inserted and always flagged as a guess.

Schema

Six arrays. droppedPreps and phrasalVerbs are machine-actionable (the translators read them); sensePreferences, collocations, falseFriends, and registerDefaults are doc-only — recorded for the human record, never applied or even flagged by tooling, because applying them needs word-sense disambiguation or judgment the tools don’t have.

droppedPreps.ruling is one of three buckets (the G3 keep/drop/replace test): drop (verb-selected, meaning-neutral — dropped, and reverse-restored by lookup), keep (marks a real relation — left alone), or replace (changes the verb’s meaning — routed to phrasalVerbs via replacedBy, naming that entry’s phrasal). At most one drop ruling exists per verb — the invariant that makes the drop reversible — enforced by tools/test/core-lexicon.test.ts.

droppedPreps.forward defaults to "apply" (the forward translator drops the preposition). "flag" withholds the forward transform for a verb whose canonical preposition has a high-frequency competing reading the tools can’t disambiguate without syntax. No entry currently uses this field: the one candidate, wait for (duration “wait for three minutes” vs. object “wait for the bus”), is handled instead by the not-duration guard in core-lexicon.ts — the forward translator drops for only when the following span is not a length of time, so wait for the bus auto-translates and wait for three minutes is kept. The reverse translator restores wait for regardless — a stoplist (prepositions, conjunctions, common adverbs, -ly words) is what keeps a duration phrase like “wait for three minutes” from getting a second for inserted; see tools/README.

phrasalVerbs.plain is the one single-word machine replacement; alternates are additional human-readable options folded into the linter’s report text (woe = [plain, ...alternates].join(" / ")) but never produced by the translator. separable is recorded (give it up) but v1 of the translator only matches contiguous phrasal spans — a separated phrasal is left alone and flagged by the scanner as ordinary prose, since it isn’t a bigram the linter tracks.

Coverage

Frequency spine: ../tools/data/ngsl.json — NGSL 1.2 (Browne, Culligan & Phillips, 2013; CC BY 4.0), 2,809 ranked headwords. The sweep walks it rank 1→2,809 in six ~500-word bands; a word earns a row only when at least one array has something non-default to say about it (a fully regular word like table needs none).

Bands swept: 6 of 6 — the full NGSL 1.2 spine (ranks 1–2,809). Row counts (test-checked against lexicon.json by vocabulary.test.ts):

This is a first full pass, not an exhaustive one: only words with a clear, low-collision-risk case earned a row (see tools/README.md’s “Sweep methodology” for the criteria and the cases deliberately left out). Growth continues opportunistically as new collisions or gaps surface.

Table A — G3 canonical prepositions (droppedPreps)

Verbs whose selected preposition is meaning-neutral: G3 drops it and the verb goes transitive. The canonical preposition is what the reverse translator restores by lookup, so each drop-ruling verb has exactly one.

wait is still a drop verb — its canonical World English is wait the bus, same as the others. The for-vs-duration test is handled by the not-duration guard in core-lexicon.ts: the forward translator drops for only when the following span is not a length of time, so “wait for the bus” → wait the bus and “wait for three minutes” is kept (the for marks a duration, per S5). The reverse translator restores wait’s for regardless, using a stoplist to skip duration phrases instead (see the schema section above).

Table B — S2 phrasal → plain (phrasalVerbs)

Non-compositional phrasal verbs and their plain replacement. Transparent phrasals (sit down, stand up) are not listed — S2 leaves them alone; only the opaque ones are replaced.

look for and believe in are routed here from G3’s drop/replace test (they change the verb’s meaning, so they are phrasal, not droppable).

Table C — S3 preferred sense (sensePreferences, doc-only)

Where a common word has a rare or risky sense, use the plain word instead. run, get, and take are reserved for their most concrete meaning.

Table D — S6 collocation → regular pairing (collocations, doc-only)

Where standard English forces an arbitrary word-partnership, a regular literal pairing is allowed. Includes the adjective- and noun-selected prepositions that G3 explicitly hands to S6 (a predicate adjective can’t take a direct object, so these are not dropped).

Table E — False friends (falseFriends, doc-only)

Words that resemble an L1 word but mean something else — a source of confident error (PAIN-POINTS §4). L1-specific, grows per language pair; shown here in full (small enough not to need sampling).

Table F — Near-synonym register (registerDefaults, doc-only, highlights)

Germanic vs. Latinate layers give English two or three words for one concept, differing only in register. World English picks the register-neutral default and drops the rest for everyday use.

Acceptance criteria

An entry is “done” only when:

1. It is consistent with lexicon.json — tools/test/vocabulary.test.ts parses every table above and every coverage count and asserts both against the data.
2. It satisfies the structural invariants tools/test/core-lexicon.test.ts checks: at most one drop ruling per verb, every replace ruling’s replacedBy names a real phrasalVerbs entry, phrasalVerbs.plain is a single word and isn’t itself a dropped form, and G3-drop / S2-phrasal bigrams don’t collide.
3. It stays consistent with samples — e.g. listen music / wait the bus agree with Passage 9.
4. The forward/reverse round-trip property holds: every drop + forward: "apply" verb translates SE→WoE by dropping its preposition, and WoE→SE restores it, flagged (tools/test/reverse.test.ts).

These tables are highlights; lexicon.json is authoritative for anything not shown here.