📐qagen-groupant-wave17

The groupant beside qagen/tests/groupant_wave17.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

Wave17 native joins

forward

Join selected Para identities to prompt/gate numbers; closed schemas and rule versions to explicit spans; normalization fibers to provenance; quotas to selected book counts.

invariant

Same ordered batch, local p/anchors, explicit skip identities, rule version and source sentence provenance.

cost

Reached selection/maps/sorts, prompt bytes and gate preprocessing/span checks; finite fixtures only.

composition

Each named test is a separate tagged relation. Composition retains the same ordered Item or Para batch, one-based row/anchor coordinates, closed labels/rule version, attempted cardinality, validated reply and isolated ledger. Independent branches do not share a live acquisition pipeline. 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.

Qagen synthetic integration harness — prospective model

Date: 2026-10-30. Source inspection base: 8910dafb6fcaebedc498b536dc62301e63826866. This document precedes harness construction. No execution results are claimed. Root owns installation, adjacent groupant, syntax/source binding, compilation, mutation runs and final evidence.

The finite forward model is an explicit composition of pure production functions: supplied Para/Pick/skip -> select indices -> ordered Para references -> prompt body and gate coordinates. A second observation joins schema and Pick rule to prompt constraints and accepted/rejected supplied TSV. The supplied TSV is a synthetic independent witness, not generated by the prompt. These four named tests do not constitute one live teacher pipeline; no corpus, DuckDB, teacher, parser or database runs are involved.

The domain consists of tiny valid in-memory paragraph vectors, unique doc#para identities, consistent synthetic word counts, positive sentence weights, ordinary finite sizes, and literal TSV. Sentence IDs and source paragraph numbers deliberately differ from local numbering. Selection consumes skip keys, not Pick.exclude file paths. Fixture metadata is supplied, not read_paras acquisition evidence.

First obernant — behavior_complement: selected outputs exclude ineligible and skipped identities; prompts exclude source sentence IDs and skip payloads; gate local p/anchor references exclude source paragraph/sentence coordinates. V2 rejects explicit spans from nonanchor sentences and context, and punctuation-changed spans. Closed schema excludes invented labels. Rule V1 intentionally has a larger explicit-answer acceptance domain. Advisory paragraph notes do not imply pair rejection. Prompt-only semantic constraints such as question quality and answer length remain outside the implemented gate guarantees.

Second obernant — reverse_reconstruction: a gate (p,anchor) and the ordered batch recover the source (doc,para,sent_id) for that observation; the gate result alone cannot recover those identities. Quotients introduced by case/quote/whitespace normalization admit multiple raw texts; the ambiguity witness uses distinct raw sentences and IDs with the same normalized span. Selection output is compatible with many rejected input paragraphs and cannot recover excluded content. Quota observations are compatible with many size/cap inputs; no unique inverse is claimed.

Predicted tests:

1. selected_identity_exclusions_join_prompt_and_gate_coordinates: selected input positions [1,3], keys tale#17 and tale#19, literal prompt P1/P2 sentence numbering, and accepted gate coordinates (1,1,[2]) and (2,1,[1]) recover literal source IDs ending 702 and 901. Excluded and heading payloads do not appear.
2. closed_schema_and_rule_join_prompt_and_span_gate: all seven literal schema names agree with prompt type lines; V1 prompt lacks VERBATIM SPAN, V2 includes it; one punctuation-changing explicit response is accepted under V1 and rejected under V2, while an exact span stays accepted. Invented type and explicitness are rejected under both versions.
3. sentence_provenance_survives_normalization_ambiguity: two raw quoted/cased/space-varied sentences normalize to king's bell rang.; same accepted local coordinates and preserved answer text map to different source IDs. Wrong-anchor and context-only answers fail in both paragraphs. This witnesses compatible preimages rather than reverse uniqueness.
4. quota_caps_join_book_selection_and_prompt_cardinality: total sentence weights (10,6), eligible non-skipped caps (1,2), request 4, produce quotas (1,2) and indices [0,5,6]; prompt contains three paragraphs. A large ineligible pool influences weight but cannot exceed caps; one skipped B paragraph contributes weight but no selection capacity.

Invariants: selection indices are increasing and map to the same ordered prompt/gate batch; gate p and anchors are 1-based local coordinates; accepted pair k is compact after rejected rows; explicit source provenance requires an anchor span under V2; schema/rule version is shared rather than guessed. Literal expected strings, indices, accepted tuples and rejection reasons are independent observations, not computed by the operation under test.

Cost model: at these fixed fixture sizes, bounded memory/time. General composition adds selection bookkeeping and seeded sorting, prompt output bytes, gate preprocessing and span/token comparisons; no whole-pipeline asymptotic claim or performance gain is measured. Reconstructing source identity from known batch and pair is O(1) indexing per anchor; compatible raw-text preimages are unbounded. No approximation/error-budget claim.

Prospective compiled mutation: in prompt::prompt change only rendered sentence index k + 1 to k + 2, retaining valid Rust and all other behavior. The named selected_identity_exclusions_join_prompt_and_gate_coordinates test must fail assertion marker W17-J1-PROMPT-COORDINATES after successful compilation. Restore exact source bytes before positive reruns. This is a wiring observation: gate still interprets anchors from 1 while prompt would instruct from 2. An alternate mutation in the harness batch adapter reverses selected references and must fail the same literal prompt-body assertion; production mutation is preferred.

Evidence to collect by root: owner API confirmations; exact final source/harness hashes; installation sidecar and partial syntax projection; focused named debug twice and release twice with normalized outcomes; compiled mutation failure at the intended assertion, restoration hash and final positives. Current evidence is source inspection only; this worker neither compiles nor mutates repository sources.

Construction receipt: external harness written and rustfmt --edition 2024 completed successfully; compilation and test execution remain pending. Harness SHA-256: 7f9447e887eae69a2a95f443555140cc8dbfa95d83a29b8062383fca5f9d38a4. Inspected source SHA-256: select.rs f8ed8d3d50fdae707957bdce4872ed66e9a9591b963f9fe6c875f29104d1c151; prompt.rs 6fe8471363a925217a3b9b80a97ce4a8e322318d2b4274394dbf4b461f6b04bb; schema.rs fb31471cabbbd651d312bcd06ffffea17f2b99694242170dfdf05bd82ef4d249; gate.rs 8a01354b712e6c60e6a5216f640f12f1965ebfa6ec10a12f56c1a0871d1b2721. Root must rebind post-integration identities.

API coordination: wave17_select confirmed public select/quotas/Pick/Para APIs unchanged, including sentence weights before exclusions. wave17_prompts confirmed schema/rules/prompt signatures and prose unchanged; checked_add adjacency repair does not affect fixtures. wave17_gate confirmed gate API unchanged, batch/local coordinate semantics, and span normalization limits. Its planned strict malformed-anchor/prefix parsing does not affect the well-formed supplied TSV.

checks

4 named tests pass twice in debug and twice in release. The compiled adapter fails its named assertion, with exact restoration. Partial syntax repeats identically; evidence.md retains native and structural receipts. Macro bodies remain opaque; semantic AST and whole-system proof are absent.

catalog

Wave17 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 four qagen joins.

Universe: supplied Para/Sent/Pick/skip, versioned schema/prompt and independent TSV observations. Exclude skipped/ineligible selections, source-coordinate substitution for local p/anchors, invented labels, V2 explicit nonanchor/context spans and punctuation-changing spans. Prompt semantic constraints not enforced by the gate remain outside this complement; no acquisition/teacher behavior is inferred.

reverse-reconstruction

Type: reverse_reconstruction. Named check: all four qagen joins.

Retain every input vector, metadata, raw sentence text, Pick/skip and response sharing selected identities, prompt/gate observations and quotas. The same retained batch plus p/anchor recovers source doc/para/sent_id; gate output alone cannot. Case/quote/whitespace normalization has many raw preimages, selection loses rejected content and quotas lose weight/cap details.

selected-prompt-gate-coordinates

Type: paragraph_sentence_coordinate_transport. Named check: selected_identity_exclusions_join_prompt_and_gate_coordinates.

Input positions [1,3] yield source paragraphs17/19, prompt P1/P2 and [1]/[2], accepted (p,k,anchors)=(1,1,[2]),(2,1,[1]), recovering source702/901 in the same ordered batch. Exclusions never enter the prompt; source paragraph numbers and sentence IDs are not local coordinates. Costs include selection, rendering, gate and O(1) source indexing per anchor.

rule-schema-span-constraint

Type: versioned_span_constraint_transport. Named check: closed_schema_and_rule_join_prompt_and_span_gate.

Seven literal schema labels agree with prompt type instructions. Pick rule1/2 selects the corresponding prompt constraint and gate; exact Red, blue flowers passes both, punctuation-changed red blue flowers passes V1 and fails V2. Invented causal/inferred labels fail. Literal response supplied independently; no language-model obedience claim.

normalization-source-ambiguity

Type: normalized_span_provenance_relation. Named check: sentence_provenance_survives_normalization_ambiguity.

Distinct quoted/cased/spaced raw sentences normalize to king's bell rang.; accepted local anchors preserve identical answer text but recover distinct source IDs through their own batch positions. Wrong-anchor and context-only explicit spans fail. The raw sentence fibers are nonunique; context is visible in prompt but unavailable as an answer anchor.

quota-selection-cardinality

Type: capacity_selection_transport. Named check: quota_caps_join_book_selection_and_prompt_cardinality.

Literal total sentence weights(10,6), eligible nonskipped caps(1,2), request4 yield quotas(1,2), indices[0,5,6], book counts(1,2) and exactly three prompt paragraphs. Ineligible/skipped paragraphs affect weights but add no capacity. Positive weights and tiny fixtures exclude overflow/zero-weight generalizations; no performance gain is measured.

evidence

Wave17 bounded evidence, 2026-10-30

4 supplied integration joins. Source SHA-256 7f9447e887eae69a2a95f443555140cc8dbfa95d83a29b8062383fca5f9d38a4; twice-identical partial syntax SHA-256 4fa7e9cc5bc618f1e3847216d978e97601ede5eead0c7d0f62186557ebed195f.

212 selected Mova checks pass twice in debug and twice in release, with identical normalized named outcomes: 155 prior focused checks, 33 new source-unit checks, 17 additional existing synthetic QA checks and seven new joins (three pragmatic, four question-generation). Root qagen library runs explicitly skip lemma::tests::doubt_with_en_model. The 100-algorithm library and its prior evidence are unchanged and are not remeasured in this epoch.

Seven existing sources gain models, with 21 nested fragments and 81 typed views; the two harnesses add 11 views, the new question-generation group adds six, and existing parents add five. Final inventory: 329/762 Rust sources, 433 pending (350 public,83 closed), 659 nodes,16 groups,314 fragments,3676 views,1109 type labels,319 partial syntax projections. Filename coverage in prag is complete; semantic coverage is scoped. Six qagen sources remain in the backlog.

Witnessed repairs: exact u128 Hamilton quotient/remainders remove the >2^53 weight-order error and usize sum overflow; checked prompt adjacency prevents i64 wrap/panic; QA anchors reject empty list components and paragraph labels permit at most one P/p prefix; accounting refuses nonfinite fields before opening and propagates non-NotFound I/O errors; run reports identify Codex transport. No broad option-compatibility change or transport rename was made.

Owner evidence records eight old-code failure witnesses and 14 compiled production mutants, followed by exact restoration and positive runs. Two root compiled adapter controls additionally fail named assertions: reversed Item/label identity and a wrong sentence-source mapping. Compiler errors are not counted as mutation rejection. Stale group bindings and wrong child identities are structurally rejected and restored exactly.

Independent source review found no further production counterexample within declared domains and refined costs: quota initialization/cap scanning remains at zero rounds; string comparisons/cloning and general JSON-object lookup work are explicit. The small quota oracle covers 45,056 tuples; this finite evidence is not a Lean proof or complete input verification. Native code and syntactic projections remain separate from models.

A gate worker also invoked a broader 26-test legacy run containing an optional local-model test; captured evidence does not establish whether that model branch ran or returned early. It is documented separately and excluded from the root 212-check receipt. Root checks use supplied fixtures and isolated temporary files, with no live teacher, parser model, corpus/database acquisition, GPU, whole-school or adoption workflow. No speed gain or new approximation-error budget is claimed.

Raw receipts: $MOVA_DATA/runs/octopus-wave17-integration-2026-10-30. native-evidence-hashes.json SHA-256 2aee840e096dcaf3d558e55c95b86f6a6623506aa620351c1e9d009ba29e94e5; mova-normalized.txt e7d315a6d937998d03e2cdc2555ba4fd96137ef61ef873104329ca632b697274; structural-controls/summary.json a50d4387189f34044579c0d068475201acf1f6e7ee8e3297c0db6909f3180975. Focused formatting passes. Cargo Clippy completes with warnings; this is not a warning-free strict-lint claim. Macro bodies remain opaque in partial syntax. External child dependencies are not inferred to have complete models from these joins.

Supplied-path dependency snapshots

These inspected production files participate in the tested paths; listed hashes are execution snapshots, not complete semantic dependency closure.

• teacher/src/lib.rs: 718e03bf89dcf8e3c69f3af6fa137d88c70d9797b04f21f1752d998e2a5b6b90
• prag/src/data.rs: b3146b6af3ca37c5a64f82c60970bea0420e5f7f62cf38642c305e09b8ce4015
• prag/src/schema.rs: 7dade219cf710dd08df7a60853a6101058b90af22a564ac35c9b3c5d4b23eff8
• prag/src/silver.rs: adf915c8d5e9059d3f58b20eca8d9b94721ee7fd9f39ac69f93775a6b5c856c9
• prag/src/opus.rs: 0169f72bddab55f9e493b85a86918b5f0adbcf5153c7c12725bb249f2ed5c136
• qagen/src/gate.rs: 09328f711abf4b6453f95bed3d873deb205c8801d939841974416fe74e3e5806
• qagen/src/schema.rs: 03e113198711324f029d8e9e4d9fc899416e3144ea856240b9d962775c78c2fe
• qagen/src/prompt.rs: 2e30ac525e2bf8cf8432298ff35eaf2ec08933c17aff60e480017a2b7cca89f6
• qagen/src/select.rs: 0c35caf4eda91c0a5d36bce728f3846b360b483a66fa4ca3c884964839cd2554

This page in Markdown → https://md.groupants.elamur.ai/free/en/mova/qagen/tests/groupant_wave17.rs.groupant/