Elamur Groupantswordgraph-graph
The groupant beside wordgraph/src/graph.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.
| 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 graph file format (21 little-endian, 8-aligned sections: nodes, relation codes, forward and reverse CSR, OOV
strings, provenance) with its Writer, and the zero-copy reader Graph (an Arc over a memory map or a byte
buffer; Clone + Send + Sync) with the queries: node lookup, ranked neighbours, edge, plausibility, plausibility with
the isa back-off, and the writer's hint rank. Designed to be shared by threads of one process (the Owls, later).
Approximant
open: validate magic, section table bounds, presence of all 21 sections, length agreement, monotone CSR offsets with correct ends, edge targets < N and group relations < R. edges(n, r, d) = the group of n with code r, in stored order. edge(h, r, d) = forward lookup in h, else reverse lookup in d. plausibility = PMI of that edge or 0. plausibility_isa = direct PMI, else max over the first 8 isa parents p of d of 0.5·plausibility(h, r, p). rank sorts candidates by (score desc, freq desc, node asc).
Invariant
Every accessor reads in bounds after a successful open (offsets checked once). Quantization: PMI q = round(16·pmi) clamped to 1..255 (monotone, error ≤ 1/32 bit within range); count q = round(8·log2 c), back-transform within 5 %. Nodes sorted by (word, UPOS), so a word's nodes are a contiguous range. The file is never written in place.
Cost
open O(N + G + E) checks (50 k graph: 38 MB, a CLI query 7 ms wall, 2026-11-06); node lookup O(log N) + one dictionary probe; edges O(groups of n + bound); edge O(bound) each way; rank O(|cands|·8·bound).
Composition
Written by compile; read by eval, the CLI, and future in-process users. Uses rel for codes and extract::OOV_TAG.
Tests
graph_is_shareable (compile-time Send + Sync + Clone), quantization_is_monotone_and_bounded,
corrupt_files_are_rejected (bad magic, edge target out of range, truncation: red side).
Behavior complement
Does not mutate, append to, or lock the file; does not verify the provenance JSON or the dictionary fingerprint against the running binary; does not handle a file truncated after open (mmap: SIGBUS); does not disambiguate senses beyond UPOS; returns 0 plausibility for absent pairs (no negative evidence).
Reverse reconstruction
From query answers one recovers the stored edges, quantized PMI and approximate counts; from a 0 plausibility,
nothing (any of the absences listed in compile's model). The section layout is invertible: Writer::finish and
Graph::new are mutually consistent on all arrays (checked by the hand-built file test).
Provenance
This page in Markdown → https://md.groupants.elamur.ai/free/en/mova/wordgraph/src/graph.rs.groupant/