Elamur Groupantsonto-query
The groupant beside onto/src/query.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 | approximant |
| ast | partial | syntax.json |
| composition | partial | composition |
| cost | partial | cost |
| descriptant | present | descriptant |
| invariant | partial | invariant |
| tests | executed bounded checks | tests |
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 |
|---|---|---|---|
behavior_complement | behavior_complement | partial source derived model | behavior_complement |
reverse_reconstruction | reverse_reconstruction | partial source derived model | reverse_reconstruction |
lookup_order | lookup_order | partial source derived model | lookup_order |
relation_completeness | relation_completeness | partial source derived model | relation_completeness |
proof_paths | proof_paths | partial source derived model | proof_paths |
parser_precedence | parser_precedence | partial source derived model | parser_precedence |
Parts
Model text
descriptant
onto-query
lookup explicit ID/POS→merged exact/lower/capitalized→first successful noun plural fallback. Context/core/prior rank retains within0.3. ask resolves first --layer before command. Property verdict preserves Kb distinction. parts sorted ancestor(depth,id), PartOf<=3, MemberOf/MadeOf depth0.
Modeling epoch 2026-10-31
Prospective model saved before edits; nine existing sources read fully. Added bounded inline tests; store reader overflow repaired only after compiled old-code failing witness. No live corpus, teacher, model or database effects executed. Source syntax generated twice identically; partial lexical projection only.
approximant
Forward relation
lookup explicit ID/POS→merged exact/lower/capitalized→first successful noun plural fallback. Context/core/prior rank retains within0.3. ask resolves first --layer before command. Property verdict preserves Kb distinction. parts sorted ancestor(depth,id), PartOf<=3, MemberOf/MadeOf depth0.
Equations
lookup=ordered_unique_senses(exact,lower,capitalized) ⊕ first successful noun singular candidate. rank retains s(c)≥max(s)-0.3. Parts=first unique(part,rel) across sorted ancestor owners and≤3 incoming part edges.
Domain
Valid lookup ordering/nonempty senses assumed. IsA miss No, parts miss Unknown. Dedup(part,rel) retains first owner/path. Ranking may load global prior; one-sense fixture avoids.
Categorical behavior and effect boundaries have no validated percentage surrogate. Executable Rust approximant unsupported; 1/2/5/10% future caps require a useful metric first.
Ranking score domain
Multi-candidate ranking claims additionally require finite per-candidate scores (finite prior shares and finite weighted sums). The prior parser can accept NaN: total_cmp may place NaN first, then every score >= best - 0.3 comparison fails, producing an empty ranked vector; isa/has_part can subsequently panic at cands[0]. Nonfinite scores are an explicit unverified failure domain, not a successful unknown-answer policy. Executed query fixtures have one sense and bypass rank_candidates scoring/global prior loading.
invariant
Invariant and domain
Valid lookup ordering/nonempty senses assumed. IsA miss No, parts miss Unknown. Dedup(part,rel) retains first owner/path. Ranking may load global prior; one-sense fixture avoids.
Limits
Finite synthetic fixtures are not semantic proof. Public invalid indices, allocation failures, huge counters, environment races and live effects outside explicit checks remain unverified.
Ranking score domain
Multi-candidate ranking claims additionally require finite per-candidate scores (finite prior shares and finite weighted sums). The prior parser can accept NaN: total_cmp may place NaN first, then every score >= best - 0.3 comparison fails, producing an empty ranked vector; isa/has_part can subsequently panic at cands[0]. Nonfinite scores are an explicit unverified failure domain, not a successful unknown-answer policy. Executed query fixtures have one sense and bypass rank_candidates scoring/global prior loading.
cost
Cost approximation
Lookup sense dedup O(S²); rank O(S log S); ancestry O(Va+Ea), repeated target BFS and broad3-level parts product costs; output-sized proof memory.
Symbolic costs only; no timing or throughput claim.
composition
Composition
lookup=ordered_unique_senses(exact,lower,capitalized) ⊕ first successful noun singular candidate. rank retains s(c)≥max(s)-0.3. Parts=first unique(part,rel) across sorted ancestor owners and≤3 incoming part edges.
Graph builder→store shape check→Kb; Kb→query→bridge/describe/pron. hs Parsed→hsmap preference/scoring/gates. Core SynMap last-duplicate vs Graph first-duplicate requires unique synset keys for interchangeable lookup. Syntax children below share parent projection; exact byte ranges bind original source semantics.
-
lookup: Found=merge(exact,lower,capitalized) else first successful plural candidate for nouns
-
meronymy: owners=sort ancestors(depth,Cid); walk incoming PartOf<=3, MemberOf/MadeOf at depth0
models
behavior_complement
Unknown word excludes guessed Boolean; no part witness excludes No; direct form excludes plural; MemberOf cannot recursively traverse.
reverse_reconstruction
Found allows case/underscore aliases and plural forms with earlier fallback misses. Yes admits compatible paths; No constrains selected senses only; answers lose unseen paths/seeds/extra args.
lookup_order
Explicit ID POS mismatch returns None immediately; adjective satellite same class.
relation_completeness
IsA/verb/role No vs meronym Unknown on no witness; relation-specific closed world.
proof_paths
BFS shortest, owners depth/id sorted, first(part,rel), depth<=3.
parser_precedence
First --layer parsed before unknown/missing command; only first pair drained, extra args ignored.
Ranking score domain
Multi-candidate ranking claims additionally require finite per-candidate scores (finite prior shares and finite weighted sums). The prior parser can accept NaN: total_cmp may place NaN first, then every score >= best - 0.3 comparison fails, producing an empty ranked vector; isa/has_part can subsequently panic at cands[0]. Nonfinite scores are an explicit unverified failure domain, not a successful unknown-answer policy. Executed query fixtures have one sense and bypass rank_candidates scoring/global prior loading.
tests
Actual-code checks
query::wave29_tests::lookup_query_and_open_world_parts
Evidence 2026-10-31
Each selected test passed twice debug and twice release after exact mutation restoration (15 named suite checks; all four sorted outcome sets equal, exits0). Compiled source control: negative-query.log (compiled source mutation assertion failure101). Controls compile successfully and fail an assertion with exit101. Store control is the original unsafe length addition before repair.
Finite graph/token/temp fixtures; public invalid states and unselected branches not exhausted
Whole CLI, live teacher/corpus/model/prior loading, filesystem failure/durability and huge inputs were not executed. No exact wrapper is advertised as a percentage surrogate. Added test blocks rustfmt-formatted independently; unrelated source formatting preserved. Strict broad Clippy not repeated: existing dependency/source warnings are outside this work.
This page in Markdown → https://md.groupants.elamur.ai/free/en/mova/onto/src/query.rs.groupant/