Elamur Groupantsworld-qud
The groupant beside world/src/qud.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
Implementation material. A Mova code model describes this code at one source revision: partial models, finite checks and their stated scope. It is not a reviewed general exposition, and its labels are shown as written, never raised.
Components
| component | label | where |
|---|---|---|
| approximant | partial source model | forward |
| ast | partial | syntax.json |
| composition | partial source model | composition |
| cost | partial source model | cost |
| descriptant | partial source model | epoch-memory |
| invariant | partial source model | invariant |
| tests | partial source model | checks |
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.
| id | type | label | where |
|---|---|---|---|
world-qud-behavior-complement | behavior_complement | partial source model | behavior-complement |
world-qud-reverse-reconstruction | reverse_reconstruction | partial source model | reverse-reconstruction |
world-qud-recency-priority | priority_relation | partial source model | recency-priority |
world-qud-slot-eligibility | constraint_relation | partial source model | slot-eligibility |
world-qud-uncertainty-partition | constraint_relation | partial source model | uncertainty-partition |
world-qud-polarity-fiber | reverse_reconstruction | partial source model | polarity-fiber |
world-qud-givenness-order | encoding_relation | partial source model | givenness-order |
world-qud-recency-resolution | information_loss | partial source model | recency-resolution |
world-qud-annotation-boundary | constraint_relation | partial source model | annotation-boundary |
world-qud-lexical-dependency | dependency_relation | partial source model | lexical-dependency |
Parts
Model text
model
Preannotated QUD stack selection and accommodation
forward
State=(open stack,given lemmas,last subject). Prepare tokens, push questions or select latest congruent, accommodate if none, then update given/last. One Step per input. No probabilistic ranking.
invariant
Stack IDs are an ordered subsequence of pushed sentence IDs. Explicit answer targets an open question. Complete removes only winner; partial preserves stack. Implicit never pushes.
cost
For S sentences, T tokens and Q open questions, conservative O(S Q T^2 + S T log(S T)) comparisons plus string/global animacy costs; output reasons can grow with rejected candidates.
composition
Preparation feeds slot eligibility, then tracking applies completeness only after ranking. track adds external annotator/splitting. Tests use track_toks only; malformed heads remain trusted.
epoch-memory
Predicted latest eligible and partial persistence; 46,656 bounded histories confirm. Generic-hedge prose overstates implementation: Who+maybe is complete. No production repair or continuous approximation metric proposed.
checks
Five native tests: 46,656 six-step histories (279,936 transitions), uncertainty interaction, empty-content slot gate, four polarity preimages, ordered givenness.
behavior-complement
Universe: admissible inputs and candidate projected outputs, fixed dependencies. Source-derived inner exclusion: outcomes contradicting the stated invariant/priority/eligibility constraints are outside R. These constraints are partial models of implementation, not a complement of linguistic truth. Finite tests support only their declared domain.
reverse-reconstruction
B(y)={x in declared domain | F(x)=y}, including latent state and dependency snapshot. Retain all compatible antecedents; rejected rows/tokens and discarded projections create ambiguity. Full text/reasons/values narrow fibers compared with key/status projections. No complete solver, unique recovery or causal inference claimed.
recency-priority
Winner=max eligible stack index regardless of completeness. Empty eligible set accommodates. Bounded oracle uses maximum eligible ID, not production search. Source-derived partial relation; checks above and explicit unexecuted gaps delimit evidence.
slot-eligibility
Nonempty question content needs overlap plus slot; empty content waives overlap only. Proper names avoid global lookup; Unknown noun animacy is ineligible. Source-derived partial relation; checks above and explicit unexecuted gaps delimit evidence.
uncertainty-partition
Complete is a branch property, not truth certainty. YesNo hedge is partial; definite Who slot ignores generic hedge. Source-derived partial relation; checks above and explicit unexecuted gaps delimit evidence.
polarity-fiber
Project away text/reasons: sleep asserted/not/never/no yield same complete attachment and empty stack. Full Step has narrower fibers. Source-derived partial relation; checks above and explicit unexecuted gaps delimit evidence.
givenness-order
Missing requested lemmas produce leakage if in answer else hallucination, retaining order and duplicates with exact string equality. Source-derived partial relation; checks above and explicit unexecuted gaps delimit evidence.
recency-resolution
Listed pronouns use last subject without gender/number agreement. Questions update last too. Subjectless input leaves prior last intact. Source-derived partial relation; checks above and explicit unexecuted gaps delimit evidence.
annotation-boundary
Tok::new lowercases, public fields and toks_of need not. First head-zero root and head=i+1 children are trusted; no parser accuracy claimed. Source-derived partial relation; checks above and explicit unexecuted gaps delimit evidence.
lexical-dependency
Who nouns depend on sense::animacy → global categories/selection; conflicting animate/inanimate categories yield Unknown. Freeze dependency snapshot. Source-derived partial relation; checks above and explicit unexecuted gaps delimit evidence.
evidence
Evidence — 2026-10-30
CPU native checks: cargo test -p world --lib wave8_models -- --test-threads=1, twice debug and twice release. Each run passes 10 tests (five per owned file), 218 filtered out. The four ordered test-name/result projections match exactly: SHA-256 df12ad1b6b325d1334fcda986aa2718c5fba2b44d0aaeca22cc6ccf3be83d7a2. Existing unrelated build warnings remain. No runtime throughput claim follows from test duration.
Negative control: Reverse newest-first selection to oldest-first; bounded_stack_matches_independent_latest_eligible_oracle fails a compiled assertion. Exit 101 with named FAILED and assertion diagnostic, not compiler failure. Saved source restored byte-for-byte; restored combined ten-test run passed. Raw mutant-qud.log SHA-256 038eb01757387f5f1d7f990f07e8fa32c71d09c453475a1cdc94161d10a0bb34.
Run artifacts are relative to $MOVA_DATA/runs/octopus-wave8-questions-2026-10-30/: debug-1.log, debug-2.log, release-1.log, release-2.log, *.normalized, mutant-qud.log, restored.log, changed.json. Large raw output is outside the repository.
Each syntax projection generated twice identically, then copied to syntax.json. A two-source isolated snapshot plus external temporary registry validates both file roots and six fragments; repeated structural reports match, with zero diagnostics and one intentionally missing temporary parent sidecar. Full-worktree check with only the two-source registry diagnoses unrelated ownership; no full-inventory validity is claimed here. changed.json reports both changed Rust files with missing=[]; ancestor integration belongs to root task.
Coverage: 10 typed root views plus three four-view fragments per file, 44 views total. Models are partial source relations and finite correspondences, not semantic ASTs, proofs, approximate speedups, or linguistic accuracy. Helper tests do not execute parser/teacher, full mining output, global hot state or I/O failures. Dependency hashes are explicit extension metadata; optional absent/runtime dependencies remain gaps.
This page in Markdown → https://md.groupants.elamur.ai/free/en/mova/world/src/qud.rs.groupant/