[jonwillits]

The corpus names the pair fluents and the places of phase (c) (docs/specs/WORLD_PHASE_C.md):

• The lexicon: a preposition for every pair fluent and every SAME literal that some precondition or condition reads, nouns for the categories of places, adjectives and part nouns for the place features, two pole adjectives per place scalar, transitive verbs for the place event types with dynamics, and a quarter of the place patient capacities; the concepts come after the generic noun, so every earlier content lexeme keeps its label. The new part of speech preposition takes no affix, and the word-form pipeline needed no change.
• The grammar: a pair state is the copula (or become) and a prepositional phrase, the preposition before or after its noun phrase by word_order.adposition, placed where an object stands; a place event type is a transitive verb with a place as its object; interpret, the readings, the JSON form, and the propositional rendering cover both (HOLDS and BECOME over PAIRFL.k(REF.a, REF.b), EVENT over PLACEEVENTTYPE.n(REF.a, REF.q)).
• Places as referents: introduced with "a", mentioned later with "the", named by mention.place_level_weights, told apart by place properties and parts, never a pronoun, never with a relative clause.
• The predicate kinds pair_state and place_event_type, with their truth from the replayed pairs and places, in the corpus and in the oracle.
• Moves and acts narrated in main clauses and in relative clauses; result sentences state a pair change, a move's new place, or the loss of the old place (WM.72).
• Filters that keep this stage's sentences out of the test sets (stage c4c's), the new counts of stats.yaml (ruling 3 of WM.72), and the figures.

Jon's decision WM.72 is recorded in docs/DECISIONS.md with the engineering choices WM.E335 to WM.E343; the proposal file docs/proposals/2026-10-29-world-stage-c4b-decisions.md has the choices, the filters, the lexicon's entries, the moved figures, the run times, and four open questions.

The tiny and default worlds keep their identities (2f80537bb72b, baa625dfd0fb; byte-identical to the reference run). The corpora were regenerated and the word forms remade from the new requests with all (ruling 1 of WM.72); the snapshot moved by 131 figures, all of the corpora (the tiny corpus 195 sentences and 589 pairs, the default 100,472 sentences with 2,560 moves, 604 acts, and 1,430 pair changes stated); the guides' figures moved by 34. Run times: define default 8.5 s (8.4 s), generate default 164 s (160 s), the word forms with all 5.6 minutes with the audio cache of the b3b run (about 30 minutes for a first run), render default 2.6 s; the fast tier 2 min 24 s; the full check list 9 min 8 s. Nothing skipped.

Co-Authored-By: Claude Fable 5.1 noreply@anthropic.com