Elamur Groupantsstudy-proposal
The groupant beside tools/enrich/src/study.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 | forward |
| ast | not materialized | |
| composition | partial | composition |
| descriptant | present | epoch-memory |
| invariant | partial | invariant |
| tests | partial bounded checks | 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 |
|---|---|---|---|
study-proposal-forward | forward_relation | partial source model | forward |
study-proposal-behavior-complement | behavior_complement | partial source model | behavior-complement |
study-proposal-reverse-reconstruction | reverse_reconstruction | partial source model | reverse-reconstruction |
study-proposal-coupling | interaction_relation | partial source model | coupling |
Parts
Model text
Partial source model, 2026-10-29. The raw Rust file remains authoritative. AST is not materialized; byte ranges are snapshot anchors, not semantic proofs.
forward
Let L(P,B) be the ordered raw rows from unblocked sentences satisfying ten-column token shape, nsubj/obj noun/proper-noun, verb head, unique literal spelling and alphanumeric boundaries. Forward output is take16(Pool(L(P,B))). This models token clozes, not semantic entailment.
behavior-complement
In the universe of packet/blocklist/output triples, any output with a blocked source, invented literal answer or more than sixteen items is outside the generator relation. This inner complement does not classify every malformed parse.
reverse-reconstruction
Given output Q, compatible inputs include exactly packets and blocklists whose ordered accepted raw rows yield Q after pooling/capping. Unemitted rows and blocked sentences remain ambiguous; empty output does not identify an empty packet.
coupling
Blocking precedes the global pool: Pool(L(P,B)) = Pool(L(P without B,empty)). Removing a source can remove distractors for other sources, so filtering final outputs is not equivalent. The 64-subset oracle checks that coupling.
composition
The pool child consumes only accepted raw rows in original sentence/token order. Parent adds literal eligibility, source exclusion and the sixteen-result stop. Shared canonicalization is assumed; the child cost already includes its raw scans, so do not count them twice. Parsing plus head searches costs O(sum tokens^2); pooling is O(R^2 log R) plus string work.
invariant
Forward and reverse observations refer to the same fixed source snapshot and dependency behavior. Ordering is deterministic for fixed inputs; filesystem/environment stability is assumed where relevant. The complement is an inner bound relative to this fragment's input/output universe.
checks
See the file-level tests.md for executed bounded checks and limits. The first three model relations are specifications; tests sample correspondence rather than proving all inputs. No expanded semantic AST or executable general inverse solver exists.
epoch-memory
2026-10-29: retained the file-level 23 models and added this finer scope. An initial incidence oracle incorrectly treated three-byte fox as a token; the existing tokenizer rejects it. Corrected the independent membership matrix rather than changing production behavior. This is a model correction, not an algorithm defect.
This page in Markdown → https://md.groupants.elamur.ai/free/en/mova/tools/enrich/src/study.rs.groupant/proposal.groupant/