📐world-groupant-wave12

The groupant beside world/tests/groupant_wave12.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

model

Supplied-input wave12 joins

forward

A tagged relation over ten native joins: common arithmetic, generated answers, state/grade provenance, narrative order, frame/role identity, lexical surfaces, agreement rows, document context, retained weight codecs and grouped folds. This is not one invented learning pipeline. The twelve catalog views comprise two baseline relations and ten specialized observations.

invariant

Preserve exact shared IDs, index conventions, temporal anchors, retained table values and declared denominators. Scope arithmetic equivalence to the shared admissible domain; keep ordinary abstention and honest controls separate. Input grouping does not establish semantic held-out independence.

cost

Compose only reached public-API costs and fixture enumeration sizes. Arithmetic grids are bounded; lexical initialization, annotation acquisition and persistence are separate dependencies. No optimization, asymptotic improvement, full-course speed claim or composed numerical error guarantee.

composition

world lessons/school/primers, QA/QA3 and world state, prag frames/roles/invariants, lens/domains, agree tables/tree forms and document-context APIs share supplied intermediates. Frame extraction, global lexical resources and expression/domain dependencies cross ownership boundaries. Codecs exercise supplied tables in memory, not filesystem durability or fresh-process retention.

epoch-memory

2026-10-30 prospective design, followed by retained-state/fold extensions. Earlier planning described eight joins; the current catalog adds two distinct state/partition observations.

Prospective integration model

2026-10-30, before executing root fixtures. Eight supplied-input joins preserve: scalar arithmetic on common valid domains; generated lesson answer and phrasing split; world-answer provenance/control denominators; narrative order and anchor grading; frame/role token identity and invariant hits; expression meaning and original forms; agreement row identity and voice policy; chronological document context with original output order.

Costs compose only along reached calls. Finite independent arithmetic expectations and explicit state constructions constrain inputs. Existing frame/global lexical models remain dependencies, not semantic truth proofs. No full learning or parser-acquisition execution is implied. Scalar equivalence excludes invalid school operations; QA and QA3 intentionally differ for expected none. Context summaries and roles lose information; inverse models must retain compatible words, source order, world states and policy configuration. Root fixtures must be nonvacuous. A compiled adapter fault will test rejection before restored repeated runs.

checks

cargo test -p world --test groupant_wave12. The ten joins pass twice per profile; the compiled wrong-question adapter rejects and exact restoration passes. evidence.md binds the executed receipts. Partial syntax is expected at syntax.json; macro bodies remain opaque.

catalog

world-groupant-wave12 scoped views

These partial relations bind the manifest source revision or transitive bindings. Each view's supplied-input domain, observation and operator are stated below. Stable dependencies and successful bounded execution are premises. Executed evidence is recorded in evidence.md; source-derived models and finite fixtures are not general proofs. Reverse relations retain compatible antecedents and do not recover unique histories.

behavior-complement

Type: behavior_complement.

Within tagged supplied-input domains exclude inconsistent accepted arithmetic, changed draw ranges or partition tags, lost answer/control provenance, reversed event/anchor order, mismatched token coordinates, changed expression surfaces, crossed row/document identity and changed retained weight/fold observations. This partial relation complement is not the complement of sampled successes.

reverse-reconstruction

Type: reverse_reconstruction.

Retain compatible expressions, seeds, failed draws, worlds, query alternatives, annotations, labels, lexical snapshots and serialized states. Equal sums, grades, projections and summaries have many preimages. Recovering a weight table does not recover its training history; preserve shared intermediates for each tagged inverse.

arithmetic-common-domain

Type: partial_arithmetic_transport.

For a,b in 0..12, independent scalar expectations compare lesson eval and school Problem::answer for addition, multiplication, nonnegative subtraction and nonzero exact division. Rejected school operations are excluded; lesson negative subtraction is a separate case. Problem::answer fallback is not typed arithmetic success.

generated-answer-identity

Type: sampled_numeric_projection.

One addition fixture supplies 32 exercises, tests and exams each. Parsed operands remain in 2..8 and 1..4; their sum equals the answer and primers reachable accepts the same observation. Repeated exercises agree. Supplied prefix tags establish fixture selection, not general semantic assessment separation; reachability is heuristic.

state-grade-provenance

Type: answer_provenance_denominator.

Construct Ada in Hall through typed commands. Parsed questions yield one hit, one miss, one ordinary abstention and one separately counted honest control. Invalid-first-query fallback agrees with the direct answer and preserves observed location history. Natural-language question understanding is outside this check.

narrative-query-order

Type: temporal_anchor_transport.

A three-event chain queried in permuted order retains E1,E2,E3 and anchors 1,2,3 through QA3 sort and grading; reversed expectation fails. Literal grading agrees with QA, while expected none intentionally differs. Missing events and empty sort abstain. Arbitrary narrative metadata remains outside the fixture.

frame-role-hit-identity

Type: annotation_frame_projection.

Two supplied subject-carry-object annotations produce frames and role triples at original one-based token positions. Valid frames have no invariant hits; patient replaced by agent produces the named row-one identity hit. Fixed lexical knowledge and anchors remain dependencies, not truth proofs.

expression-report-identity

Type: lexical_surface_transport.

Three casing variants of take place yield nonempty lexical readings. Original surfaces survive lens means and domains_with report projection with the same snapshot. This checks a real cross-subtree call, not lexical truth or general domain-classification accuracy.

agreement-row-policy

Type: label_identity_policy.

Two imperative tree blocks join label rows by sentence ID. Ordinary voice changes to imperative/request; character voice retains declarative/assert. Output reparses with identities intact; CRLF and LF observations agree. External witnesses and optional policy settings remain inputs.

document-context-order

Type: chronological_state_projection.

Three sentences in two documents are deliberately out of input order. Roles and pre-sentence context restore row order after chronological reading; document starts have no recent events and the second sentence sees its own prior actor. Input permutation preserves the ID-keyed context map. Equal ordering keys remain a separate tie assumption.

retained-weight-codec

Type: learned_state_codec.

Pure student/actor TSV codecs preserve current and accumulated tables for three supplied representable states, including the zero-accumulator activation witness. Eighty-one duplicate-vote pairs in -4..4 decode to independent exact sums and roundtrip canonically. MAX plus one clips once to MAX; canonicalization loses formatting and extreme totals. No filesystem durability, fresh-process learning or prediction-quality proof follows.

assessment-fold-identity

Type: grouped_partition_relation.

For folds 1..8, repeats 0..7 and indices 0..31, public assignment matches a u128 arithmetic reference. For nine group IDs, sixteen indices share each assignment and stay in range. Zero folds reject; supported 32/64-bit extreme indices/repeats remain bounded. Changing repeats does not establish data independence, semantic holdout separation or statistical validity.

evidence

Executed supplied-input evidence, 2026-10-30

The integrated world slice passes 244 native tests twice in debug and twice in release with identical sorted names/results: 194 selected regressions, 38 new source tests, two existing lesson tests and ten new API joins. The separate pragmatic slice passes 27 tests twice per profile: fifteen new source tests and twelve existing regressions. These are focused suites. Existing unrelated compiler warnings remain. PRAG_AGREE_ALL and PRAG_OFF are cleared for root runs; no external parser/teacher, books, live course, held-out data or GPU run occurs.

Eleven compiled production mutants fail named source assertions. The root changes fallback question row zero to row two: it compiles and fails state_answers_preserve_fallback_and_control_denominators at the provenance assertion. Exact source is restored before positive repeats. Syntax and structural reports repeat identically; twelve changed sources are registered. Stale-binding/wrong-child controls reject and restore. Whole-repository require-all remains red for 454 genuinely missing sources.

Codecs preserve supplied tables and accumulator activation; this does not establish filesystem durability or fresh-process learning. Extreme duplicate actor votes clip once to i64 after i128 summation. Learning/world updates still require representable intermediates. No universal semantic proof, new approximation bound or school speed gain follows. Fixture scopes and reverse ambiguity are in model.md and catalog.md. Root receipt set: octopus-wave12-integration-2026-10-30. SHA-256 receipts:

• world-check.log: 12800654b9a5997f9bb9b65669a264ea41d7aa148fbab423ca11d7bd1fb43d73

• prag-check.log: 328a03d6bb4c16897628c9b141e82c176b91ef79958b628e8f5dc34cf047afee

• adapter-control.log: b8d80295e11528c0c79e245630dce4785d039af059686e6a4284f0de61acdcfd

• adapter-mutant.log: e75c9ea6351bfd55bcd91dd6a332a9a0ef0eb72f6b5877b6e14d8842df191d6b

• native-groupant_wave12-debug-1-normalized.txt: fd431f9d6b55826cef715de956e76a77177f8a733ed0ec91da24b4ca4062eb6b

• native-groupant_wave12-release-2-normalized.txt: fd431f9d6b55826cef715de956e76a77177f8a733ed0ec91da24b4ca4062eb6b

• join-syntax-1.json: 3bcb9332a30db62d4b5992c925c37b233ab7b88f8acb0f5c8abf69ad8dff7be6

• groupants-final-1.json: bdd5cd88b42ab9daf29cd2819fd51903f82d7ce58400f6177ae4c1e44985fe6c

• structural-controls.log: a50d4387189f34044579c0d068475201acf1f6e7ee8e3297c0db6909f3180975

This page in Markdown → https://md.groupants.elamur.ai/free/en/mova/world/tests/groupant_wave12.rs.groupant/