📐coref-doc-adapters

The groupant beside coref/src/doc.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

Components

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.

Model text

descriptant

coref-doc-adapters

retree validates aligned raw forms; from_words is a permissive adapter with a strictly weaker validity contract.

Epoch 2026-10-31: scoped view of parent state; no independent completeness claim.

approximant

Scoped forward relation

retree validates aligned raw forms; from_words is a permissive adapter with a strictly weaker validity contract.

Parent transition model (restricted to this byte scope): For each nonempty sentence, validate all heads<=n then walk each parent path to reject cycles after depth>n. kids[p] records source order. Descending-depth propagation computes sub[p]=hull(sub[p],sub[child]). Split at any newdoc prefix and mark quotes per document, resetting open (not qid) at paragraph boundaries. retree clones gold, validates nonempty-sentence counts then word counts/forms, replaces six syntax columns, clears DEPS and refreshes. from_words clamps head>n to0 and bounds parent walks without rejecting cycles.

invariant

Scoped invariant/cost

For acyclic in-range input, depth[root]=0 and each edge increases depth by1; sub is an interval hull, not necessarily a contiguous descendant set. Retree preserves gold raw comments/MISC/forms and returns no partial clone on errors. documents ignores empty sentences; from_words retains them.

For sentence lengths n: O(sum n² + sum n log n + raw bytes), O(N+raw bytes) storage. retree clones all gold bytes and checks aligned words; from_words also traverses depth quadratically. u32 depth/qid and usize lengths have no explicit overflow policy.

Scope: retree validates aligned raw forms; from_words is a permissive adapter with a strictly weaker validity contract.. Parent bounds are conservative; no separate timing measured.

models

Scope: retree validates aligned raw forms; from_words is a permissive adapter with a strictly weaker validity contract.. Relations below are restricted to this fragment's transitions and retained parent state.

Typed models

behavior_complement

documents success excludes out-of-range heads and directed cycles in nonempty parsed sentences; does not require exactly one root or projectivity. Entity on nonword rows is rejected after tree validation. from_words success excludes no cycles: do not transport the stronger documents guarantee.

reverse_reconstruction

For a built view D, compatible raw inputs preserve its token projection, retained metadata and Entity values while allowing discarded comments/DEPS and skipped empty sentences. retree output admits compatible gold/tree pairs agreeing on aligned forms, with the six replaced syntax columns from trees and remaining raw columns from gold except word DEPS, which is forced to _. Original overwritten syntax and original word DEPS are unrecoverable.

tree_validation

All head bounds precede cycle checks; cycle checks precede nonword Entity rejection. Empty parsed sentences bypass build validation.

quote_automaton

State(open,qid): opening while closed increments qid; closing while open labels then closes; paragraph resets open only. Quotes can span sentences within one document.

retree_alignment

Align only nonempty sentences, then exact case-sensitive forms by word offset, preserve nonword gold rows, clear word DEPS and refresh after each complete sentence.

adapter_domain

from_words sets every space_after=true, no gold/speaker and newpar=false; clamps bad heads but permits cycles. downstream DFS mention detection needs acyclicity despite bounded depth walk.

No executable percentage surrogate: categorical state/effect semantics currently lack a useful cheaper observation metric. Supported future caps are 1%, 2%, 5%, 10%; no profile is claimed here.

tests

Evidence — 2026-10-31

Safe synthetic command: env -u COREF_X -u COREF_MD -u PRAG_OFF cargo test -p coref tests:: -- --skip score::. Two debug and two release executions are retained externally with numeric exit status and sorted named outcomes. Selected checks:

• tests::wave27_document_failure_order_quotes_and_retree
• tests::resolver_ignores_gold_entities
• wave27_tests::synthetic_load_preserves_gold_and_rejects_mismatched_trees

See source-test-map.json and controls.json in the wave27 coref run receipts. The library declaration file has compilation evidence only. No corpus, teacher, GPU or speed measurement.

Partial scope: from_words cyclic-input downstream termination is not established; no fault injection, u32 exhaustion or nonprojective exhaustive checks.

Measured receipts

Final source state: 13 named tests (12 library, 1 binary) passed twice debug and twice release, 52 total executed test instances; all four command exits0 and sorted named outcomes identical. Five added tests are four library boundary tests plus one main load test. The pre-existing score tests and groupant_transfers were filtered out.

Compiled actual-source control: tests::wave27_document_failure_order_quotes_and_retree. Compile exit0, assertion-run exit101, exact bytes restored, positive rerun exit0. Retained logs: control-doc-compile.log, control-doc-red.log, control-doc-restored.log.

Nine source AST projections were generated twice identically. Shared ancestor declarations are deferred to root integration; local owned leaves have no structural diagnostics. No timings or general accuracy claim.

Review witness, 2026-10-31: an external Rust binary linked the actual coref/doc public APIs against bounded four-word synthetic input. It confirmed crossing same-eid spans [0,2],[1,3] reconstruct as [0,3],[1,2], existing Document Entity caches remain stale after raw write, and retree forces output word DEPS to _ while preserving the original input. Compile and execution exited0. This single debug witness is additional to the paired suite, not a new paired/release test. Source and numeric receipts: annotation wave27 external run, coref-review-witness.rs and coref-review-witness*.exit/log.

This page in Markdown → https://md.groupants.elamur.ai/free/en/mova/coref/src/doc.rs.groupant/adapters.groupant/