📐knowledge-memory

The groupant beside models/knowledge-memory.groupant 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

Knowledge-memory descriptant

2026-10-29 epoch: twenty viewpoints expand the one-child registry group. Boolean metadata success cannot prove durable acknowledgment, residency, provenance lineage or refresh. The architectural contract in ../../docs/knowledge.md motivates prospective obligations; source bindings cover registered descendants only.

Useful observation: global retention, activation eligibility and skeleton proposal coexist. Future wiring must preserve this across identities, failures and revisions. No performance gain or full-engine implementation is claimed. AST extraction, integration traces and prospective oracle execution remain open.

2026-10-30: registered the existing local hot-state mechanism as a third tagged branch. Cursor saturation and checked ancestry repair witnessed arithmetic overflow and source-inspected cycle nontermination. Strict materialisation and permissive story/branch missing-link policies remain distinct. Local line reconstruction does not establish the global graph or retained semantic learning. Child fragments now carry separate codec versus ancestry domains and costs.

catalog

Typed model catalog

Source epoch: parent manifest source_bindings. The first two entries cover the tagged metadata, catalog and bounded local-state relations in core-model.md. Earlier prospective entries retain their metadata-specific assumptions. Local file reconstruction does not execute the proposed global storage/residency protocols.

nonbehavior

Type: behavior_complement. U is the disjoint union of operation-tagged child input/output universes; R is the corresponding tagged relation. An inner approximation B=sum_i tag_i(B_i) is valid only where B_i excludes actual child behavior under its assumptions. This rule applies to independent tagged alternatives, not to a sequential pipeline.

Metadata still excludes wrong recommendation flags or changes to opaque external state from its pure call. Local-state exclusions include wrong generated-delta reconstruction and successful checked traversal of a reachable cycle. Child finite witnesses and source inspection support conditional partial models, not a complete complement or universal proof.

antecedents

Type: reverse_reconstruction. Pre_R(tag_i,y)=tag_i(Pre_Ri(y)), retaining every compatible child input and latent state. A decision does not identify a durable record; equal decoded lines do not identify unique ancestry, edit scripts or semantic knowledge. Erasing the tag widens the compatible family without creating a call edge.

Native complete fibers belong only to the enumerated child domains. General inversion and runtime file reconstruction from outputs remain unimplemented.

durability-handoff

Type: interaction_obligation. Domain/observation: prospective admitted record q with identity.

Operator and approximation: Desired accept(q) -> q in G after durable acknowledgment. g=1 is a recommendation, not acknowledgment; no such write transition is implemented here.

Assumptions: fixed revision and shared domain boundary above; external state only in prospective entries. Evidence: source/design inspection; tests.md records the executed subset. Oracle: Future integration: crash after acknowledgment, restore q and provenance.

eligibility-residency

Type: projection_boundary. Domain/observation: eligible records and prospective RAM set H.

Operator and approximation: h=1 licenses consideration; desired H subset of eligible durable records under a separate budget. h does not imply q in H.

Assumptions: fixed revision and shared domain boundary above; external state only in prospective entries. Evidence: source/design inspection; tests.md records the executed subset. Oracle: Future oracle: reject automatic residency claims when budget is zero.

eviction-conservation

Type: interaction_obligation. Domain/observation: prospective eviction of q.

Operator and approximation: Desired evict(q):(G,H)->(G,H minus {q}); global content and provenance unchanged.

Assumptions: fixed revision and shared domain boundary above; external state only in prospective entries. Evidence: source/design inspection; tests.md records the executed subset. Oracle: Future storage oracle: compare durable snapshots across eviction.

compile-conservation

Type: interaction_obligation. Domain/observation: prospective skeleton promotion.

Operator and approximation: Desired promote(q):(G,H,K)->(G,H,K') retaining q and evidence; k=1 is only a proposal.

Assumptions: fixed revision and shared domain boundary above; external state only in prospective entries. Evidence: source/design inspection; tests.md records the executed subset. Oracle: Future promotion receipt must retain source record identity.

conflict-independent-activation

Type: wiring_constraint. Domain/observation: leaf flags consumed by two hypothetical downstream handlers.

Operator and approximation: Conflict handler consumes a, activation handler consumes h. For equal i and differing t, residency eligibility is equal even when alternatives differ.

Assumptions: fixed revision and shared domain boundary above; external state only in prospective entries. Evidence: source/design inspection; tests.md records the executed subset. Oracle: Leaf toggle check executed; downstream consumer wiring not implemented.

review-not-truth

Type: information_loss. Domain/observation: review metadata joined to evidence.

Operator and approximation: r is reviewed applicability/evidence status; neither r nor k implies truth. Equal flags can describe mutually inconsistent assertions.

Assumptions: fixed revision and shared domain boundary above; external state only in prospective entries. Evidence: source/design inspection; tests.md records the executed subset. Oracle: Future oracle: preserve two reviewed rival claims without selecting truth from flags.

premise-lineage

Type: provenance_obligation. Domain/observation: prospective derived claim q.

Operator and approximation: Desired durable record carries premise IDs, derivation rule and model version; Boolean Decision has no encoding for these.

Assumptions: fixed revision and shared domain boundary above; external state only in prospective entries. Evidence: source/design inspection; tests.md records the executed subset. Oracle: Future oracle: follow every derived record to named premises and version.

revision-refresh

Type: version_coherence. Domain/observation: prospective G revision with resident q.

Operator and approximation: Desired cached(q.version)=G(q).version or mark stale before use; metadata F does not inspect versions.

Assumptions: fixed revision and shared domain boundary above; external state only in prospective entries. Evidence: source/design inspection; tests.md records the executed subset. Oracle: Future oracle: mutate global version and reject stale materialization.

derivation-invalidation

Type: dependency_obligation. Domain/observation: prospective changed premise p.

Operator and approximation: Desired affected derived records reachable from p become invalid or revalidated with history retained; metadata does not discover dependency edges.

Assumptions: fixed revision and shared domain boundary above; external state only in prospective entries. Evidence: source/design inspection; tests.md records the executed subset. Oracle: Future oracle: revise p in a two-edge derivation chain.

scope-join

Type: identity_coherence. Domain/observation: record scope and recommendation scope.

Operator and approximation: Intended consumer must bind Candidate.scope to the same record/version being considered; otherwise Instance can borrow a general mechanism decision.

Assumptions: fixed revision and shared domain boundary above; external state only in prospective entries. Evidence: source/design inspection; tests.md records the executed subset. Oracle: Future oracle: replay a general decision against an instance record and reject identity mismatch.

budget-feasibility

Type: resource_constraint. Domain/observation: prospective resident selection x.

Operator and approximation: For m_j>=0 and M>=0 require sum x_j*m_j<=M and x_j<=eligible_j; policy supplies eligibility only.

Assumptions: fixed revision and shared domain boundary above; external state only in prospective entries. Evidence: source/design inspection; tests.md records the executed subset. Oracle: Future oracle: oversized useful record remains eligible but unselected.

utility-choice

Type: counterfactual_objective. Domain/observation: prospective eligible records with fixed costs.

Operator and approximation: Proposed maximize sum x_j*(p_j*c_j-r_j) under budget; alternatives can tie, so no unique selected set is claimed.

Assumptions: fixed revision and shared domain boundary above; external state only in prospective entries. Evidence: source/design inspection; tests.md records the executed subset. Oracle: Future finite brute-force knapsack oracle; no scheduler exists here.

admission-order

Type: stage_order_obligation. Domain/observation: prospective ingest -> admitted -> recommend.

Operator and approximation: Candidate contract begins after admission; recording unverified claims must retain their unverified status. F cannot implement admission or downgrade evidence.

Assumptions: fixed revision and shared domain boundary above; external state only in prospective entries. Evidence: source/design inspection; tests.md records the executed subset. Oracle: Future oracle: trace status through admission and durable write.

failure-atomicity

Type: failure_prefix. Domain/observation: prospective write then activate sequence.

Operator and approximation: Desired failed durable write cannot produce a falsely durable success receipt; tentative working memory needs distinct status. No transaction implementation is modeled.

Assumptions: fixed revision and shared domain boundary above; external state only in prospective entries. Evidence: source/design inspection; tests.md records the executed subset. Oracle: Future fault injection between durable write and activation.

checkpoint-boundary

Type: scope_boundary. Domain/observation: knowledge-memory versus retained school checkpoint.

Operator and approximation: Registered closure also contains local hot-state reconstruction. Neither catalog tables, line snapshots nor a school checkpoint graph establish a durable global store, resident cache or automatic skeleton generator.

Assumptions: fixed revision and shared domain boundary above; external state only in prospective entries. Evidence: source/design inspection; tests.md records the executed subset. Oracle: Inspect registry closure; future integration must bind cross-subsystem artifacts explicitly.

replay-idempotence

Type: interaction_obligation. Domain/observation: prospective duplicate admitted record/version.

Operator and approximation: Desired replay preserves one canonical identity and provenance history; pure repeated F outputs do not establish idempotent persistence.

Assumptions: fixed revision and shared domain boundary above; external state only in prospective entries. Evidence: source/design inspection; tests.md records the executed subset. Oracle: Future oracle: replay same record twice and compare identities/derivations.

aggregate-cost

Type: cost_composition. Domain/observation: n independent recommendations plus hypothetical downstream work.

Operator and approximation: Metadata work O(n); total cost n*O(1)+write+fetch+refresh+training terms, all downstream terms unknown here. No RAM-speedup claim.

Assumptions: fixed revision and shared domain boundary above; external state only in prospective entries. Evidence: source/design inspection; tests.md records the executed subset. Oracle: Leaf source inspected; future instrument stages independently.

versioned-policy-recovery

Type: reverse_reconstruction. Domain/observation: decision history under a known policy version.

Operator and approximation: Compatible histories require each d_j=F_version(c_j); without version, union over possible policies enlarges fibers. Current manifest hash binds only current metadata policy.

Assumptions: fixed revision and shared domain boundary above; external state only in prospective entries. Evidence: source/design inspection; tests.md records the executed subset. Oracle: Future oracle: store policy version, distinguish present conflict gate from relaxed hypothetical gate.

tagged-state-fibers

Type: tagged_reconstruction_boundary. Equal observation bytes from different operation tags are not interchangeable. Decoded lines leave deleted source lines, edit-run grouping, ancestry and metadata ambiguous. Retain those possibilities in the inverse; do not infer learned facts or storage receipts from line equality. Evidence: local finite codec cases and source-level branch separation, not an integration run across all children.

line-codec-composition

Type: generated_delta_reconstruction. For bounded child line fixtures, apply(a,diff(a,b))=b. Diff selects one script among potentially many. Bind the same base a, not a merely equal-length replacement. File-level splitting/delimiters are a separate serialization boundary; arbitrary newline-bearing persisted payloads are outside this equation's evidence. Native generation/application and an independent script oracle check the bounded relation.

ancestry-before-decoding

Type: validated_history_path. A finite parent-map snapshot and selected ID determine a checked path under an explicit missing-link policy. Reachable cycles reject before materialise decodes deltas. Story and branch selection retain permissive missing-link behavior but terminate on cycles. Unvisited components do not invalidate a local successful path. Child path tests plus source-level caller inspection are the evidence; file reads are not made atomic and concurrent publication is untested.

knowledge-memory-wave12-principle-review-projection

Type: derived_knowledge_review_boundary.

Supplied knowledge facts pass through statement rendering and derived/not-yet-checked key selection before optional review and publication. Rendering retains the selected link and negation but discards frequency, source and derivation metadata; it is not a replacement for the durable fact. Selection and review cannot by themselves establish truth or promote every derived claim to the skeleton. Leaf fixtures execute the pure rendering/selection relations; know storage, teacher/cache I/O and multi-file publication remain external dependencies.

Wave12 source-derived view; the supplied domain and observation are above. Executed scoped integration evidence is recorded in world/tests/groupant_wave12.rs.groupant/model.md#evidence. Fixed dependencies and shared identities are premises; finite native joins do not establish full-pipeline behavior or recover unique input history.

wave29-graph-language-encoding-grammar-transport

Morph Gram::from_vesum keys meet uk analysis grammar and textgen synth grammar after NEG/IMPLICIT masking. Implicit subject copies predicate person/number/gender; main stores fact gram in every gs and member gram in gd. This preserves packed bits, not original VESUM tags, discarded xp tags in Gram, or semantic correctness.

wave29-graph-language-encoding-identity-boundary

Registry::Id=(heap<<32)|local is split exactly; relation edges use local only and require separate rel heap context. Global concept IDs are lexical vector positions with no heap/local encoding. Equating these numbers would fail even for a one-concept compiler and first global heap2^27.

wave29-graph-language-encoding-completeness-vs-realization

render_check sorts normalized words; generate chooses conditional/global-frequency variants and DP order. Therefore render_check(text,Fact)=Ok does not imply generate(Fact)=text, nor does generation establish role correctness. Synthetic tests separately witness each actual function.

wave29-graph-language-encoding-effect-frontiers

sub bad mask precedes reads; parse_lines loads input/names/Morph before output creation; ingestion creates names before Registry open and writes columns before registry save. Build expressions writes before global.rs. Success/failure does not compose into a transaction across these files.

wave29-ontology-core-selection-categorical-state-transport

TSV known/unknown/n/a transports to Field Known/Unknown/NotApplicable; risk none to Known(empty), never Unknown. coregate validates complete categorical cells but not field-value grammar.

wave29-ontology-core-selection-sense-identity

coresel filters exact synset; q_synsets accepts noun accepted/also mapping; Core SynMap collect last duplicate vs Graph find first means unique key/POS is necessary for agreement.

wave29-ontology-core-selection-csr-join

Builder dedups(s,r,t) and builds both orientations; model borrowing assumes valid offsets. store validation belongs between loaded bytes and Kb queries; builder does not itself exclude cycles.

wave29-ontology-core-selection-teacher-effect

corevmm/judge prompts and pure parsers can be tested without wdvmm::ask. Output append errors and in-flight worker effects remain outside demonstrated contracts; probability estimates are not measured frequencies.

wave29-ontology-query-state-layer-proof-transport

Known(value,win,chain,beaten,l)→query Yes/No plus seed/chain text; pron maps value to Tri and persona; describe counts only noncharacter applicable-known props. Undetermined and Unknown both map Tri Unknown but why differs.

wave29-ontology-query-state-unit-gate-join

Rat endpoint values already in base units before hsmap compares shared attr/base intervals. Numeric overlap formula is dimensionless; basis compatibility is not checked by current hsmap shared predicate.

wave29-ontology-query-state-snapshot-cursor

p0=0≤B; guarded n≤B-p yields p'=p+n≤B, establishing safe addition/slicing. Typed enum and shape checks then feed Kb but do not establish graph DAG/CSR inverse semantics.

wave29-ontology-query-state-hs-selection

hs main=first deepest Item else Head; hsmap matches original kind/level/NP then compounds candidate phrases. State key live/food separates confirmations; first accepted Q vs last confirmation rules cannot be interchanged.

wave29-ontology-external-alignment-identity-transport

Q→idx_g(Q)→nodes_g(index) is identity for present Q; idx_before cannot be reused in rebuilt g2. Rank vector must be recomputed in g2.nodes order.

wave29-ontology-external-alignment-source-count-join

n_sources(mask)=1[label]+1[lex]+1[ili∨wn]=n_sources_str(src_names(mask)); all16 masks checked. This agreement does not validate candidate factual truth.

wave29-ontology-external-alignment-numeric-join

wdx quantities preserve amount/unit/bounds strings; wdwalk converts value and bounds using property-specific factor, computes mean/sample sd on values only, min/max on union. Unsupported units skipped; finite values not universally gated.

wave29-ontology-external-alignment-effect-precedence

wdvmm cap guard precedes mkdir/teacher; append log precedes is_error. Extractors may partly write before refusal; Zip slices precede compression dispatch. Compiled controls test selected boundaries, not live transactional behavior.

wave29-snapshot-forgetting

Let pi(K)=K.g and S(G)=pack(G). On the tested valid three-concept graph, unpack(S(G))=G preserves concept ids, words and CSR edges. The composite Kb::new(unpack(S(pi(K)))) discards K's assertions, additional layers and extra edges. Its preimage includes all admitted K states sharing G; the snapshot alone cannot recover a learned or layer-specific answer. Property specificity and layer visibility were observed only after explicit synthetic assertions were supplied.

wave29-shape-versus-edge-duality

Let V(G)=store::check(G). V checks array lengths, offsets and reference/string bounds; unpack separately checks encoded enum discriminants. V does not impose in(v)={(r,u): (u,r,v) in out}. A three-concept graph with a mismatched but in-bounds incoming edge passes V and pack/unpack, while ancestors follows outgoing edges. Therefore V=Ok excludes tested structural corruption but admits relational inconsistencies; it is not a certificate of complete graph semantics.

wave29-identifier-namespace

kg Id(h,l)=(h<<32)|l is injective over two u32 components. With equal l and distinct h, the tested identifiers remain distinct through edge sort, flags filter and column persistence; a reopening of the temporary registry retains both keys. onto Cid is an index into a particular graph and has no implemented identity-preserving conversion from Id. Dropping h identifies distinct vertices; any future adapter must supply namespace and provenance explicitly.

wave29-allocation-rejection

The tested 46-byte checksum-valid snapshot requesting an array of length u64::MAX and registry containing global heap u32::MAX must return errors before successful state extension. Old code panicked on unchecked cursor/successor addition; repaired code rejects these finite boundary fixtures in debug and release. This constrains integer-boundary handling, not allocation exhaustion, malformed-file completeness, transactionality or crash durability.

wave31-annotation-store-adapters-representation-transport

Tsv::row encodes Null as backslash-N, rejects sentinel string, preserves empty string and writes header at creation. COPY-to read has no header skip. JSON serializes anchor arrays and comments before TSV so control characters travel as JSON escapes.

wave31-annotation-store-adapters-identity-coordinate-join

load matches (sent_id,text) against pre-state while qa maps doc_key then insists each anchor has unique sid with same doc and Some(para). Equal annotation bytes alone cannot identify a sentence; matching sent_id alone can admit an incorrect text.

wave31-annotation-store-adapters-effect-frontier

Parse/duplicate/state checks and TSV validation precede write exec. Post TSV parsing/counts follow exec, and batch file loops preserve earlier effects after later failure. Partial TSV row buffers may contain prefixes even when count remains zero.

wave31-annotation-store-adapters-tagged-independent-paths

Prag label schema, QA question schema, CoNLL load/export and tales segmentation are separate dispatched paths sharing transport utilities. One schema's parser success does not imply another's admission or DB table constraints.

wave31-annotation-store-adapters-fixture-observation-limit

Existing DB integration tests are modeled but not run; new safe tests in those files exercise actual en::conllu Doc parsing and fixture layout only. They do not evaluate query strings, views, migrations or transactional behavior.

wave31-annotation-store-adapters-cost-composition

Sum actual invoked helper costs plus process/SQL costs on the selected tagged path; do not add all CLI branches as if executed together. Full input/result materialization and sequential per-file DB calls remain, with no measured optimization claim.

tests

Evidence and checks

2026-10-29, source SHA bound in manifest.json.

Executed: compiled global/src/placement.rs directly with rustc --edition=2021 --test and ran all three existing unit tests: three passed. This isolates the unchanged metadata policy; no full crate or system suite was run.

A separate Rust probe imported the same source and ran twice with identical output. It enumerated 48 candidates, comparing all four output fields with F; checked all 768 candidate/decision pairs against the complement and converse equations (720 absent); counted six output fibers of sizes 2,2,10,10,12,12; checked 48 repeated calls and importance toggles; compared the relaxed-conflict counterfactual (four differing inputs). A contested general foundation rejects the deliberately relaxed skeleton rule, providing a negative control. The probe also checked all 40 manifest section IDs against exactly one matching catalog heading per ID.

The probe was external review tooling, not production code or an added registered Rust source. The finite relations are fully stated in core-model.md and catalog.md, allowing independent regeneration. No timing measurements were made.

Executed structural checks: installed mova-modelcheck groupants with source bindings and component paths; changed inventory against base ab4ba4321; git diff --check and repository secret guard. Structural inventory does not validate mathematical adequacy. Global repository sidecar backlog remains visible.

Only the explicit finite probe and three named existing tests are executed evidence. Other catalog oracles (including minimal repair enumeration, scope substitution probes, batch permutations and all prospective persistence/RAM/promotion integration checks) remain proposed. Missing AST and unimplemented surrounding engines are not covered by passing metadata tests.

Local hot-state epoch, 2026-10-30

The new child executes three pure native tests: 961 generated-delta roundtrips/minimality cases, 9317 iterator-decoder cases and 1250 ancestry-policy comparisons. Two compiled source-extracted mutants fail both profiles; the actual world module passes twice per profile. Parent models preserve operation tags and distinguish line equality from semantic retention. Caller wiring is source-inspected; no live hot-state files or environment variables were touched. The full parent storage/activation protocol remains prospective.

core-model

Knowledge-memory interaction model

Registered children: placement-metadata, global-catalog and hot-state. Their source closure is bound by the parent manifest. Model ownership is a tree; actual calls can cross that tree.

Forward

The modeled relation is a tagged union R = R_meta + R_catalog + R_local. Metadata recommendations, compiled catalog queries and local hot-state helpers have separate inputs and observations. There is no automatic join from one branch's successful result to another branch's storage, activation or learning operation.

Metadata: (c,e) -> (F(c),e), with opaque external state e untouched. F(c)=(1,i,(s!=Instance) AND m AND r AND NOT t,t). Durable recommendation is true; eligibility equals importance; alternatives equal conflict; proposal implies a reviewed non-instance foundation without conflict. Storage state is not an output of F.

Catalog: APIs observe compiled concepts, links, expressions and judgments. Generated tables and their seed provenance are not fully bound by the registered source hashes. Lookup success does not establish all-human knowledge coverage or persistent admission of a new fact.

Local state: the bounded model covers line-delta generation/application and checked parent traversal, with source-level caller wiring into materialisation, story and branch selection. A generated delta reconstructs its target line vector under the child domain. Ancestry validation checks only the selected path. File publication, concurrent writers, complete checkpoint integrity and other state-module operations remain outside the executable evidence.

Composition and invariant

For each operation, preserve its tag, line sequence, parent-map snapshot and missing-link policy. Reconstructing equal lines does not recover unique deltas or ancestry, and does not prove equal knowledge semantics. Child source identities bind an epoch; they are not proof that runtime files are authentic or unchanged.

Intended global wiring remains: admitted record -> durable provenance record; usefulness -> budget-constrained activation; reviewed mechanisms -> versioned promotion. The local hot-state module is an existing artifact mechanism, not this complete global-store protocol. No transaction protocol or automatic skeleton generator is established by these checks.

Cost

Streaming n metadata recommendations costs O(n) time and O(1) auxiliary space. Catalog costs depend on the queried operation and fixed tables. Local diff uses O(nm) time and space for line counts n,m, subject to line-comparison cost; application consumes a bounded source cursor and emits copied/inserted lines. Checked ancestry follows reachable links with ordered-set/map lookup costs. Count these paths only when executed; no training, GPU or throughput gain is measured here.

Evidence boundary

Leaf tests establish finite correspondences. The new parent views organize their domains and caller obligations; they are not a new end-to-end persistence test. Missing semantic AST, unregistered dependencies and prospective storage/residency checks remain explicit.

wave29-graph-language-encoding

Actual module declarations in kg/src/main.rs connect cols/gram/id/morph/uk/textgen; Morph and uk share Gram; textgen consumes uk::Fact and Morph synthesis; main writes Registry IDs and Edges using relation.local and full vertex Id. Registry keys are append IDs, not lexical concept indexes. global/Cargo.toml has no runtime dependencies and kg/Cargo.toml has no global dependency. Compile seed path is a separate tagged branch, not a claimed graph pipeline. Tests exercise actual Morph→parse→TSV and Fact→generate joins on invented inputs.

wave29-ontology-core-selection

coresel uses wdwalk core resolution and corevmm key/verdict readers, build.rs consumes coregate then core.rs includes generated arrays. Graph model feeds store/Kb/query; core checks layer0 Kb properties and ancestry. Builder consumes Zip; judge/corevmm call wdvmm only in unexecuted effectful paths. Passing parser is not successful numeric code emission or teacher correctness.

wave29-ontology-query-state

Graph Builder.finish (ontocore)→store.check/unpack→Kb.new is real CSR boundary. Kb.prop→query.prop_of→bridge.answer_question/pron.info/describe.describe preserves Known/Unknown/Undetermined distinctions except pron Tri merges latter two with reason. hs::parse main feature/NP→hsmap candidates/pick/isa_verdict→Kb ancestors/query.parts_of. Core SynMap last duplicate vs Graph first duplicate is invalid adapter unless keys unique. Bound.overlaps requires callers align attr/base/basis; hsmap shared gate checks attr/base but not basis, so cross-basis comparison is an explicit limitation.

wave29-ontology-external-alignment

wdx::Tsv→wdgraph::for_each_row/read_isa_edges and wdlex rows; IsaGraph rebuild→rank_kinds/folk_ancestors; wdalign::src_names→wdwalk::n_sources_str; wdwalk::read_core→bridge::map_class or explicit model membership, candidates→coresel confirmation; wdvmm::ask→corevmm/judge live-only; wsd→query/kb/core; eval→query::Answer; build::Builder→Zip. P31 adapter differs between alignment and walker, and CRC/bounds remain ZIP gaps.

Wave29 composition feedback

The added graph and ontology branches are operation-tagged partial relations. Snapshot restoration reconstructs Graph arrays; it does not restore Kb layers/defaults. Structural array checks do not prove incoming/outgoing edge duality. Heap/local identifiers require an explicit namespace adapter before use as ontology concept indices. The four wave29 catalog views retain these measured boundaries and two repaired integer failures. Native observations use tiny fixtures; full durable/global integration, live alignment and crash behavior remain open.

wave31-annotation-store-adapters

Real joins: main->load/export/prag/qa/tales; load/export/prag/qa->Duck+Stage+Tsv; load/export/tales->en::conllu or tokenizer/annotator; util supplies SHA1/path/time. TSV write includes header while read expects COPY output without header: direct write/read is not identity unless header separated. Prag eight/nine-column input and QA eight-column input are distinct tagged schemas. Integration test fixtures exercise real CoNLL representation but SQL branches remain unexecuted. SQL include_str files are external dependencies, not selected Rust sources.

This page in Markdown → https://md.groupants.elamur.ai/free/en/mova/models/knowledge-memory.groupant/