📐lexical-semantics

The groupant beside models/lexical-semantics.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

model

Lexical and command observation group

Forward approximant

For fixed lexical tables let N be the number API, V the current value-judgment relation, P the typed-command parser and G the trigger observation API. Model their tagged union, retaining actual caller edges separately. Child domains and failures apply; quantity and normalization differ. Value rules implement command-history predicates, not universal ethical or factual truth.

Composition

The input languages are separate; no local call translates normalized prose into commands. A joint path needs an encoder and lexical identity preserving quotes, IDs and provenance. Global catalogs and English dictionary dependencies cross the registry tree. These lexical observations alone do not certify persisted learned graph state. Parsed commands can feed the separately registered World runtime, whose validation and mutation are additional gates.

Value fields and same-goal peers traverse string-key order. Help and mercy read harm history before a set command. Another field cannot create their prior justification in that same command, but can support a later one. Lifetime suppression and duplicate-key last-write behavior remain. Determinism does not imply unique preimages or retained learning.

Invariant

Preserve operation and observation tags. Source bindings omit generated dictionary contents. Numeric bounds remain explicit; partial AST integrity is not semantic proof. Distinct-key permutations preserve current value output; changing final duplicate-key values need not.

Epoch memory

Baseline1e63aaf0a showed two sorted judgment projections from one mixed command under fresh HashMaps. This epoch orders fields/goals and freezes prior harm/harmer eligibility. The historical witness describes the older source. Both canonical explanation order and historical eligibility matter: sorting alone was insufficient. Existing single-state-field examples retain behavior.

The command/trigger epoch adds native parser fibers, role/registry cells and quoted/duplicate/error-precedence fixtures. A real parse_all-to-world::run test crosses the ownership tree, preserving one shared program and world history. Successful parsing does not imply runtime acceptance. Fixed-time trigger arithmetic has pure fixtures; live clock activation and publication effects remain source-inspected. No general natural-language or course-quality conclusion follows.

Dictionary interpretation expansion

Add the meanings observation family M to the tagged union N, V, P and G. Keep normalized records, operation projections, fact projections and cached snapshots as distinct tags. The earlier views retain their stated domains. Reverse reconstruction of M requires an equivalence class of records and lexical snapshots, rather than a unique source dictionary. Native supplied-record fixtures cover extraction and interpretation; external layer loading and cache invalidation remain separate boundaries.

catalog

Typed lexical interaction models

Subject: child revisions in manifest.json. Finite calls, fixed dependencies and child domains. Current value rules use canonical maps and prior-command history. Historical hash-order evidence remains in the child epoch record. See tests.md for executed oracle scope.

excluded-tagged-observations

Type: behavior_complement.

U contains tagged number/text, value/command, command-parser and trigger observations with candidate outputs under fixed dependencies. R is the disjoint union of the child relations on their stated domains. Actual joins additionally preserve shared program, state and dependency identity. Candidate inner U minus R excludes changed operation tags, a cardinal from quantity on standalone seventh, a judgment from an absent rule, and help justified only by harm first recorded in the same set command.

Oracle: Check ordinal/absent-rule boundaries and prior-history fixtures.

compatible-texts-and-traces

Type: reverse_reconstruction.

For (b,y), retain all inputs and dependency configurations satisfying R_b(x,y). Spelling loss, overwritten fields, hidden histories and absent rules prevent unique reconstruction. Value preimages use prior-command harm; mercy requires the current agent in that prior harm history.

Oracle: Enumerate bounded histories; retain ambiguity and absent-rule effects.

prior-command-coupling

Type: command_snapshot.

For external agent a and target x freeze h=(x in H0) and s=(a in H0[x]) before scanning a set. Relief can add help only with h and no previous helped pair; mercy also requires s. New harm can support later commands. Ordered keys fix explanation representatives; duplicates stay last-write.

Oracle: Permute fields across absent/other/same prior harm; compare full vectors and reject live-history mutants.

seed-and-gap-boundary

Type: dependency_obligation.

Values consult global catalogs; numbers consult source tables and English correction dictionaries. File hashes do not bind every dependency. gaps and Found are observations/proposals, not persisted learning receipts.

Oracle: Vary a controlled dependency and inspect persistence boundaries.

separate-input-languages

Type: composition_boundary.

Normalization takes prose; judgment takes commands. Neither locally calls the other. A serial composition requires an encoder preserving quoting, IDs, names, lines and final assignments. This parent selects a branch.

Oracle: Test an encoder before asserting a commutation law.

observation-quotients

Type: information_loss_partition.

Quantity, text plus Found, full ordered judgments and principle membership are distinct projections. Canonical traversal fixes current value order but not injectivity: projecting details loses hidden history, overwritten assignments and explanation fields.

Oracle: Retain multiple antecedents for each projection.

bounded-domain-meet

Type: precondition_composition.

Numeric overflow/renderer bounds and catalog assumptions remain local. A joint model also needs an encoder. The value fix neither validates narrative facts nor resolves lifetime-harm policy or cross-field contradictions.

Oracle: Keep numeric boundary fixtures and same-command/later-command distinctions.

branch-local-cost

Type: resource_decomposition.

Selected cost is dispatch plus reached work. Ordered field parsing costs O(B+K log K) comparisons with string costs; goal updates O(log G); prior-harmer snapshot scans one history list. Other scans remain potentially superlinear. Nested views overlap.

Oracle: Use bounded same-output CPU batches; no course-wide gain follows.

command-language-gates

Type: parser_runtime_boundary.

Syntax errors and runtime reference, role, register, anchor and state conditions are separate gates. A syntactically valid move can fail because its destination is absent. The native five-command World fixture uses the actual parse_all output, accepts four commands, and checks the rejection and snapshots. Parser finite fibers preserve quoting, aliases, duplicate fields, zero anchors and existing quantity policies; these policies do not assert natural-language truth. Reverse queries retain both the typed command and runtime context, since parsing success alone cannot reconstruct a committed state change.

trigger-observation-tags

Type: clock_numeric_observation_boundary.

Fixed-time unit/calendar equations are distinct from live clock, activation and filesystem observations. Equal formatted text does not uniquely reconstruct times or logs, and a correct numerical projection does not ensure publication. Native bounded tests use fixed fire_at inputs and pure helpers. Live activation and time-dependent I/O are unexecuted here; source inspection supplies their current conditional boundary.

dictionary-record-preimages

Type: record_override_preimage.

The meanings projection accepts at least six TSV fields, normalizes keys to lowercase and retains selected fields. The last normalized duplicate wins. Preimages include ignored and overwritten rows; the result is not a union of every sense. Operation and fact projections lose more information, so they cannot recover original spelling or provenance uniquely.

dictionary-cache-snapshot

Type: first_observation_state.

Compiled records, optional layers and first-use caches have separate identities. A Rust source hash does not bind included data or runtime overlays. Updates to external files need not be observed after a cache captures its value. Pure extraction fixtures do not establish publication, invalidation or retained learning.

expression-token-identity

Type: dependency_token_identity.

MWE token IDs map to the same dense coordinates used by arcs and expression sets. A credited match needs distinct shared tokens and a shared VERB. Greedy order remains; these constraints do not establish maximum matching or linguistic truth. Dictionary snapshots and tag conventions remain domain inputs.

wordnet-bounded-hypernyms

Type: lexical_graph_truncation.

WordNet traverses synset identities breadth-first and returns a deterministic prefix of at most 40 distinct member labels. Labels can collapse different synsets; missing records and first-use snapshots remain explicit. Bounded graph fixtures do not validate external dictionary accuracy or cache refresh.

partial-domain-configuration

Type: finite_override_completion.

Absent or nonfinite domain overrides use shipped defaults. Explicit finite zero and off-grid values retain their meanings. Empty annotations under the empty override map remain consistent with shipped defaults. Tuning callers supply complete maps and explicit zero ablations. Lexical snapshots and thresholds remain inputs; these checks establish neither learned optimality nor held-out accuracy.

lexical-semantics-wave12-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.

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.

tests

Evidence and limits

Current value policy: seven source tests pass, including the fresh world Cargo test run. A direct-source oracle adds 1,280 complete judgment comparisons over reachable histories, state fields, cause gates and both field orders. Old live-history mutations and reversed peer order fail. See current-evidence.md in the value sidecar for exact dependency assumptions and local scope. Number checks remain those previously recorded for unchanged source.

The baseline1e63aaf0a mixed-field witness observed two sorted projections. It describes the old source; its historical harness requires that checkout. Current tests require canonical full output and pre-command help/mercy eligibility. Later commands can use newly recorded harm; lifetime suppression remains.

Structural validation checks identities and AST/model references, not mathematical completeness or generated dependency contents. No full-course, GPU, teacher or whole-language accuracy claim follows.

Command and trigger epoch, 2026-10-30

Six native lang tests and five native trigger tests pass twice in each profile within the final 68-test focused world suite. Parser evidence includes 784 bounded relation-fiber inputs, 45 role/registry cells, 25 command variants and selected quoting/duplicate/error cases. The separately owned world fixture executes parse_all -> world::run and observes snapshots and one runtime rejection. Compiled source controls fail relevant assertions. Fixed trigger inputs do not exercise live activation, clocks or filesystem publication.

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