📐question-generation-composition-wave20

The groupant beside qagen/tests/groupant_wave20.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

Wave20 native joins

forward

Three synthetic joins compose real Fixes validation, prompt numbering, repair gating, anchor-ID mapping, span checking and injected lemma classification. They preserve admitted behavior and exclude invalid preflight side effects without teacher/model execution.

invariant

One-based local anchors map to the same paragraph's sentence IDs. Explicit spans and implicit preserved answers retain distinct meanings. Stub callback counts are exact work observations, not real model throughput.

cost

Three constant-size fixtures, one paragraph/two sentences/one item; lemma calls bounded by one answer plus listed anchors after validation; allocation and normalization costs remain. No live service timing.

composition

Each named test is a separate tagged relation. Composition requires matching coefficient conventions, finite domains, normalization metrics, probability mass or radius, observation coordinates and explicit tolerance assumptions. Error caps apply to their declared metric only; no uniform whole-pipeline percentage follows. Distinct tests do not constitute one end-to-end pipeline. Independent finite or literal oracles remain distinct from cross-implementation equality.

epoch-memory

2026-10-30 prospective design retained below; no speed or accuracy gain is claimed.

Wave20 independent integration prospective — 2026-10-30

Eight bounded algorithm joins cover twenty production modules: SPD factor/triangular/iterative solve; QR least-squares transported to quadratic optimization; scalar Brent/Armijo; Lasso threshold solution; interpolation/Bezier/quadrature; monotone cubic with Romberg; subset sum transported to knapsack; totally-monotone/partition/BST DP oracles. Independent exact fixtures supplement cross-implementation agreement. Residual observations are not solution-error certificates on ill-conditioned systems; quadrature differences are not rigorous error bounds. Matrix optimizer domains and DP assumptions remain explicit.

Three synthetic qagen joins exercise real validation/gating/repair paths under deterministic injected lemma stubs. No teacher, model, corpus or database execution. Equal output rows do not recover original order or metadata.

Before measurement: compare bitset subset sum with existing boolean DP on identical deterministic bounded item/target fixtures, exact decision equality, two executions, seven alternating-order timing samples, black-box outputs and bounded repeats. Prediction: packed words reduce O(nT) scalar work to O(nceil(T/64)); tiny cases may regress. Qagen stub call-count profiles compare valid/invalid anchor routes with same deterministic stub, predicting invalid inputs perform zero analysis and lawful behavior remains unchanged. Stub wall time is not model/service throughput. Root runs timings after harness review. No global speedup claim.

Validation: actual production joins and independent oracles; twice debug/release normalized outcomes; compiled semantic negative controls; exact source hashes retained by parent integration. External harness pending integration; worker sources read-only to reviewer.

checks

3 named joins pass twice per profile; compiled adapter rejects, restoration passes and paired partial syntax matches. See evidence.md. Macro expansion and semantic AST remain absent.

catalog

Wave20 supplied-input relations

Source-derived specifications; bounded native evidence and repeated partial syntax are recorded in evidence.md. No general proof or complete inverse engine is claimed.

behavior-complement

Type: behavior_complement. Named check: all named joins.

Admissible repair outputs retain item and paragraph identity, valid one-based anchors, gate-approved spans or preserved implicit answers. Its complement includes dropped/misaligned items, out-of-range anchors, invalid UTF-8 slicing, callback effects before invalid-input rejection and explicit outputs failing the span relation. Synthetic checks cover only selected branches.

reverse-reconstruction

Type: reverse_reconstruction. Named check: all named joins.

A repaired explicit span identifies compatible sentences/anchors but not a unique old answer; implicit output preserves the chosen old answer and anchors but loses generating prompt provenance. Lemma coverage is many-to-one in forms, order and content labels. Validation errors retain many malformed antecedents; no inverse dataset reconstruction.

implicit-repair

Type: validated_prompt_repair_lemma_relation. Named check: validated_fix_prompt_gate_and_lemma_join.

One synthetic paragraph with cat/dog sentences and item cats sleeps anchored1. Validated Fixes builds prompt Item1; implicit gate output preserves old answer and anchor. Counting lemma stub maps cats->cat, treats the as function word; try_doubt returns (1,1.0), exactly two calls (answer then anchor). No real annotator/model.

explicit-anchor

Type: anchor_coordinate_span_transport. Named check: explicit_fix_anchor_ids_join_verbatim_and_lemma.

Same paragraph, response Item1 anchor2 the dog runs becomes explicit. Numeric anchor2 maps to sent_id s2; independent span_reason accepts output. Same actual repair output fed to try_doubt returns(2,1). Word normalization may erase case/punctuation information; sentence identity is retained.

invalid-preflight

Type: validation_effect_exclusion_relation. Named check: malformed_contexts_are_rejected_before_lemma_callbacks.

Fixes rejects anchor0; try_doubt rejects [1,0] before any callback; legacy doubt with anchor2 on one sentence returns None with zero calls. Multibyte first label is classified malformed by real fix gate without panic. Errors do not authorize partial effects; no file/model/database operations.

evidence

Wave20 harness evidence

3 named joins pass twice in debug and twice in release within the 69-check package suite. Corresponding compiled adapter mutation rejects by assertion, exact restoration passes and paired partial syntax matches. Source SHA-256 f50e7759c6a4e14d228678f323ac8d9449888c9241f610009231335c1063d5d4.

Forward: Three synthetic joins compose real Fixes validation, prompt numbering, repair gating, anchor-ID mapping, span checking and injected lemma classification. They preserve admitted behavior and exclude invalid preflight side effects without teacher/model execution.

Invariant: One-based local anchors map to the same paragraph's sentence IDs. Explicit spans and implicit preserved answers retain distinct meanings. Stub callback counts are exact work observations, not real model throughput.

Native receipts bind actual logs; finite observations do not prove arbitrary input behavior or inverse uniqueness.

This page in Markdown → https://md.groupants.elamur.ai/free/en/mova/qagen/tests/groupant_wave20.rs.groupant/