Elamur Groupantsword-graph
The groupant beside models/word-graph.groupant 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 | Forward |
| Ast | not materialized | |
| Composition | partial | Composition |
| Cost | symbolic source derived | Cost |
| Descriptant | present | Descriptant |
| Invariant | partial | Invariant |
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.
| relation | type | label | where |
|---|---|---|---|
| Behavior complement | behavior complement | partial source derived model | Behavior complement |
| Reverse reconstruction | reverse reconstruction | partial source derived model | Reverse reconstruction |
Model text
Descriptant
The word graph (wordgraph/): passes over parsed corpora count typed content-word pairs; a build keeps the top
neighbours per (word, relation) both ways and writes one file that threads read in place. Epoch 2026-11-06: first
version; real runs over Wikipedia tier 1 (5 000 and 50 000 articles), measurements in wordgraph/README.md.
Forward
trees --extract--> events (rules v2) --pass--> parts (sorted counts, OOV tables, provenance) --compile--> file
(nodes, relations, forward/reverse CSR, OOV strings, provenance) --graph--> edges, plausibility, isa back-off, hint
ranking --eval--> agreement with the absurdity matrix. The CLI (main) wires them; rel supplies the codes.
Invariant
Determinism end to end: part bytes depend on the unit, rules version and seed only; the file depends on the set of parts, parameters and registry only (both measured twice on real data, 2026-11-06). Bounded size: edges ≤ Σ_groups bound(r), both directions. The reader validates every offset once and never writes.
Cost
Pass: O(T + E log E) per unit, parallel over units. Build: two k-way merges O(M log P) + candidate sorts. File ≈ 6 B per edge each way + 9 B per node + group tables. Measured 2026-11-06 (50 k articles): pass 35 s / 0.57 GB, build 58 s / 1.25 GB, 38 MB file, 2.07 M + 2.83 M edges, 135 k nodes.
Composition
Joins: extract::pack order = compile merge order (monotone OOV rank map keeps it); graph::Writer layout =
Graph::new validation; pass::Meta copied into the file provenance. Correct leaves do not prove these joins; the
end-to-end test (tests/build.rs) exercises all of them on a synthetic corpus, and the real runs exercise them at scale.
Behavior complement
The group does not: assert facts or truth (it counts what parsed text says), resolve word senses beyond UPOS, use coreference, deduplicate overlapping passes, weigh sources, smooth PMI, update a graph in place, or delete data. It does not answer for words outside its nodes (0 = no evidence, not «implausible»).
Reverse reconstruction
From a graph file: the kept pairs with quantized PMI and approximate counts, node frequencies, the passes' provenance (sources, parts' SHA-256, models, dates). Not recoverable: sentences, exact counts, pruned pairs, marginals. A pass part recovers the full event multiset of its unit (with OOV strings) but not sentence order or boundaries.
Provenance
This page in Markdown → https://md.groupants.elamur.ai/free/en/mova/models/word-graph.groupant/