Elamur Groupantsworld-bookans-retrieval
The groupant beside world/src/bookans.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-bookans-retrieval-behavior-complement | behavior_complement | partial source model | behavior-complement |
world-bookans-retrieval-reverse-reconstruction | reverse_reconstruction | partial source model | reverse-reconstruction |
world-bookans-retrieval-selection-fiber | reverse_reconstruction | partial source model | selection-fiber |
world-bookans-retrieval-stale-index | error_partition | partial source model | stale-index |
Model text
Partial source model, 2026-10-30. Each typed view shares this source revision, scope, assumptions and checks unless narrowed below. Observation is one finite helper call, not full conversation. Source-derived relations remain hypotheses supported by bounded correspondence; no general semantic proof.
forward
Valid k score=2*index term occurrence count + |(lemmas union ws) intersect sentence words|. Skip comments/header/unparsable/out-of-range rows. Keep positive scores, order(-score,k), top30 then union valid neighbours and take90 ascending. Index occurrences repeat; lexical terms deduplicate.
invariant
Supplied worlds require valid one-based event sentence anchors and consistent entity/state references. Book retrieval returns unique increasing in-range zero-based indices. Answer helpers do not mutate worlds. Copy gate acceptance establishes neither factual correctness nor a general copyright guarantee. Partial syntax is not semantic analysis.
cost
Index bytes I, BTreeMap updates logK, sentence word sets WlogW, ranking KlogK, <=90 neighbour insertions.
composition
Normalization feeds hashed membership and independent quote scanning. Retrieval combines index/lexical evidence before ranking and expansion. run feeds notes to prompt and at most two teacher calls, gates responses, then logs after printing. Query annotation, sentence splitting, teacher transport, environment mutation and log errors remain unmodeled semantic/effect dependencies. Local evidence does not certify that path.
epoch-memory
Initial wave8 source review and native checks. Preserve ordering, ambiguity and external dependency gaps. No useful approximation metric is supplied for categorical dispatch, so executable Rust approximants remain absent. Tests are not fast surrogates. See root evidence.md for repeated profiles and compiled controls.
checks
cargo test -p world --lib wave8_models -- --test-threads=1; repeat with --release before -p. Only named native helpers exercised; no live parser/teacher. Fragments use whole-file partial syntax projections plus source byte envelopes, not inferred types, expanded macros or semantic AST.
behavior-complement
Universe U is admissible supplied inputs paired with categorical observations of scoped relation R below. The source-derived inner exclusion hypothesis rejects observations violating local priorities, thresholds or constructor outputs. Unmodeled branches and dependencies remain unknown; U minus a partial model is not a certified complement of the entire Rust file.
reverse-reconstruction
B(y)={x in the declared domain:R(x,y)} retains all compatible latent tokens, scores, ties and states. Equal text or score does not identify a unique input/history. This is a relational specification, not an inverse solver. Native finite fibers give scoped evidence, not completeness outside their domain.
selection-fiber
Same neighbour union admits distinct scores/hit sets. Enumerate64 fox/crow masks over six sentences; independent distance<=1 predicate checks union.
stale-index
Invalid IDs cannot consume top30 slots. Before guard usize::MAX overflowed k+1 in debug and index=count could include last sentence. Regression includes both, valid row and empty sentences, plus forty higher-scoring invalid rows competing with one valid hit. The valid hit and neighbours survive, distinguishing early filtering from late output filtering.
This page in Markdown → https://md.groupants.elamur.ai/free/en/mova/world/src/bookans.rs.groupant/retrieval.groupant/