📐prag-frames-questions

The groupant beside prag/src/frames.rs in the Mova repository: what the code computes, what it never does, how to go back from its outputs, its invariants, cost and checks.

Facts

Components

Obernants

The typed relations around the forward model: the behaviour complement (what the code never does), the reverse reconstruction (every input compatible with an output) and specialised views.

Model text

Subject: prag/src/frames.rs SHA-256 bfbc2f7905d878080a7e7dd30ae7f1bfe8f07313c3a4cf8f0254be29e72a4b88. Partial source model, finite native evidence. Fixed configuration and admitted finite resource domains; source-derived inner exclusions and ambiguous preimages, no inverse solver.

forward

ask_of maps why to Cause, where to Place, when to Role(Attr) (the time the reader keeps as when: attributes; PRAG_OFF=wh-time restores Frame), who subject to Agent and other WH cases by tree role; Asked frame references predicate rank. units splits plain text and applies statement_order only for questions. plain gives a speaker label its verb of saying back before parsing (speaker_label, PRAG_OFF=speaker-label): a sentence-initial capitalized label of at most six words before its first comma and twelve in all, with no digit, quotation mark, LABEL_CLAUSE word, verb-only word or (after the first word) possible VBZ/VBD form, followed outside quotes by a colon and an opening quotation mark, gets " says" inserted at its first comma or end (Dad: "pass me the salt" -> Dad says: "pass me the salt"). repair functions can alter supplied word dependencies before frame extraction.

behavior-complement

Universe: supported who-as-Nsubj question trees and candidate Asked observations. Exclude an Agent gap naming a different frame or changing it to Place. For empty text exclude any parser call in read_with. For speaker_label exclude any change to a text with no colon followed by an opening quote, any change inside a quotation, any byte change other than one inserted " says" per accepted label, and a second insertion into an already rewritten label (idempotence).

reverse-reconstruction

An Agent gap does not recover the question's wording, subject agreement or normalized do-support. Different questions share Ask::Frame; when- and how-questions share Role(Attr). A bare time phrase keeps only its whole value (March 2022 -> when:march 2022), so the year alone is not recoverable from the option's attribute set as a separate value. speaker_label is injective on accepted labels: deleting each inserted " says" before its colon recovers the input, but a text that already read Dad says: "..." has the same image as Dad: "...", so the preimage of a rewritten text is the set of those spellings. Anchor role projections lose unstated/shared arguments.

invariant

Question-specific rewrites are conditional; parser callback sees normalized unit text. Bare predicate sentinel is usize::MAX, not a valid token.

cost

Sentence normalization linear-to-quadratic in text/token copying; anchor lookup scans predicates/roles. General repair correctness is not proven.

question-coordinate

For two supplied units with Who as Nsubj in the second, asked.frame=1 and what=Role(Agent), while pred[1]=(1,1). Native fixture rejects conflating token position and global frame position.

composition

Only supplied fixtures execute; natural language parser quality, all repairs, bare fragments and ambiguity are explicit gaps.

epoch-memory

Time and name roles, 2026-11-04. Prediction before code: the Wikipedia school's when-questions are all guesses because the reader asks Frame and keeps only when:<lemma> (the year lost), and which-questions tie because the answer is a name inside the remembered entity; a whole time value as a when: attribute, when asking Role(Attr), and a name role (Ent::names) should lift when to ~60-75% and which to ~55-70% with the articles-shuffled control flat. Measured (CPU, peta build, train and cv courses, two identical runs): when 259 -> 881 of 1106 (shuffled 263), which 65 -> 200 of 214 (shuffled 64) on wikipedia; cv-a 143 -> 457, 27 -> 90; cv-b 155 -> 444, 24 -> 106; subjects -12 net on wikipedia. Ten school courses 3234 -> 3248 of 7283. Tests wh_role_tests: page 29 and their march have no time value, a cup of tea names no cup; the mutant that accepts any month word fails the time test (exact source restored).

Speaker labels, 2026-11-03. Prediction before code: labels are read with the speaker as subject of the quoted clause (Dad: "Leo, pass me the salt" -> dad pass salt to=i); restoring the verb of saying lets the existing say/quote reading assign me to the speaker and the vocative/you to the addressee; score effect expected small, reading effect large. Measured on 180 labelled test questions (commands, world, adult-life): speaker read as the one who says 14 -> 102, speaker read as doer of the quoted clause 20 -> 14. School courses (CPU, two identical runs on commands): commands 43 -> 40 of 115, nine primary courses 3053 -> 3065 (odd lessons +9, even 0). A guard-removal mutant (no possible-finite-verb check) fails what_is_no_speaker_label_stays_byte_for_byte; exact source restored.

Wave15, 2026-10-30. Prospective repair: reject unresolved right heads in identity with a constant six-item guard, preserving coordination membership and its nontransitivity. Old production failed unresolved_identity_is_symmetric_even_in_coordination; repaired code passes. Compiled guard-removal control fails the same named assertion and exact bytes are restored. Eight focused tests pass twice debug and twice release. No speed, full semantic coverage or universal proof claim.

limitations

Partial syntax is not expanded semantic AST. No live parser, teacher, whole corpus, GPU or environment mutation. Acyclic input trees and fixed configuration/resource premises matter; malformed recursive ancestry, counter overflow, complete TAM/lexicon/quotation/repair semantics and all active-rule effects remain open. No executable approximant/error budget is claimed.

This page in Markdown → https://md.groupants.elamur.ai/free/en/mova/prag/src/frames.rs.groupant/questions.groupant/