answer-span

The groupant beside world/src/span.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

Implementation material. A Mova code model describes this code at one source revision: partial models, finite checks and their stated scope. It is not a reviewed general exposition, and its labels are shown as written, never raised.

Components

component	label	where
approximant	partial source model	forward
ast	partial	`syntax.json`
composition	partial source model	composition
cost	partial source model	cost
descriptant	partial source model	epoch-memory
invariant	partial source model	invariant
tests	partial source model	checks

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.

id	type	label	where
`answer-span-behavior-complement`	behavior_complement	partial source model	behavior-complement
`answer-span-reverse-reconstruction`	reverse_reconstruction	partial source model	reverse-reconstruction
`answer-span-routing-fiber`	reverse_reconstruction	partial source model	routing-fiber
`answer-span-predicate-priority`	priority_relation	partial source model	predicate-priority
`answer-span-cleaning-fiber`	reverse_reconstruction	partial source model	cleaning-fiber
`answer-span-early-return-shadow`	behavior_complement	partial source model	early-return-shadow
`answer-span-numeric-bypass`	information_loss	partial source model	numeric-bypass
`answer-span-location-time`	constraint_relation	partial source model	location-time
`answer-span-reason-horizon`	resource_boundary	partial source model	reason-horizon
`answer-span-feeling-envelope`	information_loss	partial source model	feeling-envelope
`answer-span-fallback-objective`	counterfactual	partial source model	fallback-objective
`answer-span-rule-fiber`	reverse_reconstruction	partial source model	rule-fiber
`answer-span-fallback-tie-intervention`	counterfactual	bounded native correspondence	fallback-tie-intervention
`answer-span-candidate-expansion-and-retry`	composition_relation	bounded native correspondence	candidate-expansion-and-retry

Parts

Model text

model

answer-span

Subject: world/src/span.rs, whole-file SHA-256 in manifest, baseline 4279da541. Domain: finite acyclic valid trees with consistent child lists, valid indices and terminating dependencies. Observation: a single call return, not semantic truth or reachability. Hold Tree, ans::lemmas and types::FEELING dependencies fixed; parser/model loading and external caller wiring are gaps. All entries are source-derived partial models. Exact-target definitions are not executable inverses; finite tests cannot establish universal correspondence.

forward

want_of routes metadata; extract finds predicate then dispatches ordered rules, usually cleaning and joining selected indices. Num bypasses clean. minus_question is separately called by external code.

invariant

Non-Num success contains ordered unique indices, positive content and echo<=half. Inputs unchanged. Acyclic consistent Tree is a termination assumption, not checked here.

cost

find_pred O(pnL) with morphology cost L. Conservative O(n^3+n^2 log n+pnL) for repeated subtree/sort/clean operations under constant-cost memberships; clean's front removals cost O(m^2). String bytes, hash worst cases and dependency internals remain gaps.

epoch-memory

2026-10-29: initial source-bound epoch. Preserve Num bypass, global Feel selection, early-return cleanup suppression, and verb_kids comment mismatch: code expands Xcomp only and checks no shared subject. No source changes or task-accuracy claim. Partial syntax has no semantic analysis or macro expansion.

checks

Structural inventory checks hashes, references, types and node bounds only. Proposed oracle: enumerate small annotated acyclic trees and compare literal expected returns, including None and labels; mutate priority or echo inequality to obtain a red control. Any executed bounded evidence is recorded separately in root evidence.md; catalog examples remain proposed until recorded there.

composition

Children share immutable Tree and question state. Routing yields Want, predicate selection yields p, dispatch yields candidate indices, normalization yields text; Num bypasses normalization. Guarded branches compose as a union, sequential stages via existential intermediates. Reverse keeps all compatible intermediates; complement is this node's U minus composed R, not union of child complements. Why/clauses overlap extraction. minus_question is separate API and caller wiring is outside file. Do/Clause delegation falls through after After None but most direct branch returns stop on clean None.

behavior-complement

U=terminating extract calls on valid acyclic trees paired with arbitrary optional (text,rule); R=actual return graph. Inner U minus R excludes Some on empty trees, a Num rule other than nummod, and non-Num success with no content or majority echoed content. Num bypasses cleaning; extending the latter exclusion to Num is false.

Assumptions/evidence/status: common domain above; source-derived partial relation at bound revision. Approximation and held-fixed inputs are stated in the relation. Oracle: finite literal fixtures exercising both membership and nonmembership against the actual function, preserving ambiguous preimages. No linguistic correctness or universal proof inferred.

reverse-reconstruction

Observed (text,rule) retains every compatible Want, tree, anchor and candidate index list reaching that label and output. None retains empty trees, absent anchors, rejected candidates and exhausted scans. This necessary-constraint outer preimage preserves repeated words, deleted edges and distinct syntax; there is no unique semantic inverse.

routing-fiber

Fixed other frame fields: Say/Think map to Say; Want/Obj/WhoObj to Obj; WhoSubj/WhatSubj/WhoIs to Subj; Happen/HappenTo to After iff sub is Some, else Clause; Generic to Clause. Fibers retain all matching kinds and all payloads of the required sub presence.

predicate-priority

For each preds entry in caller order, rank matching tokens VERB=0, ADJ/AUX=1, NOUN=2, other=3, preserving earliest rank tie. Accept rank<3 then stop. Earlier predicate noun beats later predicate verb. PROPN alone is rejected, even though downstream branches mention PROPN.

cleaning-fiber

Fixed tree/qlem and observed indices J: retain every input list whose sort/dedup and asymmetric edge trim yields J and whose content passes 2*echo<=content, content>0. Duplicates, permutations and removable edges are lost; text-only observations additionally merge repeated forms.

early-return-shadow

Inner U minus R on a fixed dispatch path excludes later candidate outputs once an earlier matched branch executes return out, including when clean returns None. Obj final oblique loop retries after failure, and Do/Clause After if-let falls through after None; these exceptions prevent a global no-retry claim.

numeric-bypass

Want::Num returns raw form of first global base=Nummod OR UPOS=NUM token, label nummod. Predicate and qlem do not constrain it. All changes preserving this first qualifying token preserve observation, regardless of relevance to the question.

location-time

Where local scan accepts selected locative Obl excluding TIME_N or LOC_ADV Advmod; global fallback scans Obl only. When searches verb_kids if p exists, all indices otherwise; first time-marked Advcl precedes temporal Obl. These lexical heuristics do not prove temporal/spatial semantics.

reason-horizon

Why scans p and at most two permitted head steps before lexical fallbacks. An ancestor beyond three chain nodes is outside the structural scan, but lexical because/for/since can still recover words there. Complement exclusion requires holding fallback triggers absent.

feeling-envelope

Feel scans all tokens irrespective of p for first eligible emotional predicate, then emits contiguous envelope of selected adjective/intensifier/conjunction indices. Intermediate unselected words may enter; rejected first candidate ends branch. Emotion noun fallback occurs only when no adjective candidate exists.

fallback-objective

minus_question chooses maximal CONTENT-count run split by echoed content tokens, earliest tie, count>=2, then cleans once. Maximizing token length or trying a runner-up after cleanup failure defines a different system; equal-content unequal-length runs distinguish them.

rule-fiber

Rule labels retain branch ambiguity: conj-after has child and sibling paths; obj/ccomp/xcomp occur in Obj and Say. A rule is an operational diagnostic, not a proof of causal or answer truth; different Wants and trees remain compatible.

native-epoch-2026-10-30

Prediction before native probes: exact cleanup threshold permits half echoed content; ties preserve earlier candidates; direct cleanup rejection shadows later alternatives while final oblique scanning retries. All six synthetic native tests passed the first debug observation. The complete bounded preimage checks are documented in normalization.groupant/model.md#wave6 and predicate.groupant/model.md#wave6; root wave6.md retains intervention and wiring observations. Current source corrects the helper comment recorded as a mismatch in the earlier epoch; behavior remains unchanged. The earlier standalone-link limitation is replaced for these native checks by a fresh Cargo build of this checkout. General parser correctness, other Wants, malformed-tree termination and all caller wiring remain unmeasured. See evidence.md for final debug/release repeat and negative-control records.

evidence

Answer-span evidence

2026-10-29, source baseline 4279da541. Seven nodes: root 12 typed models, six fragments with four each, total 36. Extraction owns Why and clauses as nested refinements; other root children cover routing, predicate selection and normalization. All six fragments have syntax node IDs and exact enclosing byte ranges. Root syntax projection has 3708 nodes, source 30078 bytes, SHA-256 a7db75bd3d52ee90eee6f562d4cef3fccbf7407bbf2531ff5e6377ad41c51b0c.

Executed checks

The source-bound syn projection repeated byte-identically. Two complete groupant inventories with the unchanged registry repeated byte-identically and reported valid=true, diagnostics=[], 19 sidecars and six registry group sidecars. This new leaf remains unregistered until the integration step; registry-node.json is its descriptor for inference-core. Inventory validation establishes structure, not model semantics.

The standalone Rust harness probe.rs.txt includes the current authoritative span.rs directly; it links pre-existing world/en dependency rlibs for Tree, question types, morphology and emotion tags. Those dependencies were not rebuilt or certified by this probe. It never loads a parser/model or contacts a teacher.

Thirty-two literal fixtures passed: three predicate-priority/PROPN cases, one numeric echo-bypass case, sixteen question-mask fallback cases, ten kind/sub-presence routing cases, and two object selection/rejection cases. An intentionally false retry-after-echo-rejection prediction was rejected. Two subsequent harness runs produced identical summaries:

literal fixtures=32; wrong retry model rejected=true

This is a negative formula control, not a production-source mutation test. The result does not cover all Wants, malformed trees, end-to-end answering, parser accuracy, or general inverse completeness. Catalog experiments beyond these fixtures remain prospective.

Reproduction

From repository root, generate syntax with mova-modelcheck syntax into a new output and compare with syntax.json. The retained build_models.rs.txt is a Rust authoring helper using serde_json; compile it with --crate-name span_models and an available serde_json rlib, then run from root to regenerate model.md and manifest.json files. Re-generation overwrites those files; review model changes before running it.

Compile probe.rs.txt directly with rustc --edition=2024 --crate-name span_probe, supplying --extern world, --extern en and the dependency directory using local build artifacts. Keep the output binary outside the repository. The relative include resolves the unchanged span.rs beside this sidecar. Dependency artifact versions are an explicit reproducibility limit.

Source review observations

Num bypasses clean, predicates and echo filtering. Feel is global, and Say's quote pair is global once a predicate exists. Candidate cleanup often ends a branch with None instead of trying another candidate; Obj final obliques and Do/Clause's After delegation have distinct retry behavior. find_pred prefers earlier caller predicates and rejects rank-3 PROPN-only hits. verb_kids expands only Xcomp children and checks no shared subject, despite its broader comment. minus_question maximizes content count with earliest ties.

These are operational facts or comment/implementation mismatches, not established linguistic bugs. No runtime source, shared registry, active data, or learned state was changed. The partial AST has no macro expansion, name resolution or semantic verification.

Native epoch, 2026-10-30

Baseline 48f537bb250b47e571aaf2758a39c82ab6bddadd. The historical standalone probe above is retained; the new evidence compiles the actual world library and its path dependencies with Cargo. Six native tests in span::span_models construct annotated Tree values directly, without parser assets or teacher calls. Production behavior is unchanged. The verb_kids comment now accurately states its one-level Xcomp expansion, lack of Conj expansion, and lack of subject checking.

Five specialized models are added in wave6.md at root, normalization and predicate scopes. The original behavior_complement and reverse_reconstruction entries remain first. Total: seven groupant nodes, six scoped fragments, 41 typed models (previously 36). Two complete bounded domains are measured: 341 index lists for each of four echo masks (1364 clean calls), and all 64 rank triples over VERB/ADJ/NOUN/PROPN. Tests compare entire actual and independent oracle preimage maps, including None. Marker asymmetry, unequal token-length ties, content-count intervention, branch shadowing versus retry, and one-level child expansion use separate literal fixtures.

Regeneration is two-stage: regenerate source syntax, run build_models.rs.txt, then run wave6_models.rs.txt. The latter adds the new declarations and precise models; do not run it repeatedly on an already extended manifest. Both helpers are authoring artifacts, not production modules. The syntax projection and all six fragment source bindings are refreshed to the current source. Partial syntax still omits macro expansion, including assertion bodies, and does not establish semantics.

The initial groupants structural check reported exactly two stale ancestor bindings (inference-core and mova) because this task owns only the source and its adjacent sidecar. Ancestor refresh belongs to integration. No leaf diagnostic was reported. A groupant inventory is not a correctness proof. Focused Cargo checks emit existing world/dependency warnings; this work makes no whole-world lint or full-suite claim.

Final focused validation: debug2/debug3 and release1/release2 each passed all six tests. A Rust validation helper removed build chatter and compared the six named outcome lines; all four normalized outcomes matched exactly. Native profiles are CPU-only, no optional GPU features. The threshold mutation changed the production comparison from echo*2 > content.len() to >=. Cargo rebuilt it; cleanup_complete_bounded_fibers failed its equality assertion for mask=1, list=[2,1], returning None instead of Some([1,2]). This is a compiled production mutation control, supplementing the earlier false-formula control. The source was restored and all six tests passed again. The restored syntax projection matched the retained projection byte-for-byte. The changed-file gate reports world/src/span.rs and missing=[].

Run artifact directory is named octopus-span-wave6-2026-10-30 under the operator's external runs root. It holds debug1 (initial), debug2/debug3, release1/release2, normalized results, mutation.log, validation.log, restored.log, repeated syntax projections and changed.json. No performance gain or broader language accuracy is inferred from these bounded observations.

Evidence SHA-256 bindings for this native epoch:

Artifact	SHA-256
span.rs	4d215ccd57884f41e77168839aac2710913efc4b233aa43ebdb2de36c7fcaa2c
syntax.json	0534f29ea3a624c338e297eda8b0c3d0f0e43ba86ef7bc69705540f96bcfbe49
debug2.log	fd05ac54b291712cf6221fe715165144a977b2667042e982c2d38a674dcac6e5
debug3.log	174bec41ad5869d7b5758d3983b6772dbe6536c4ac482b891b16846afabd5778
release1.log	7e05e81eb8116ff01b094b96be475bcc754baa9e8414a30885202e65c0445e51
release2.log	5edd7fa3e25f6624ca4d5caea08ff8e4f4a9f23fe031d94f04e6e72883a29902
mutation.log	3c15810aad554dffb7550bce1b33cc47a9506fdfe9cff536c9cc14e45885f18e
validation.log	c96ef0fd5fe258c834b6dc83c1020c0d2ff9a768f1e8c5aa26a5e97c69e5d5bd
restored.log	fd378ccad17e36ecdd69502523a1afd07ec73eede0c920445ba2321f9010a005

wave6

Native selection observations, 2026-10-30

Revision: exact source SHA in this manifest. Observations are single terminating calls. Tree fixtures are directly annotated, acyclic and index-valid; dependencies are compiled from this checkout. No parser or teacher is loaded. Metric is exact equality of optional indices/text and rule labels; zero tolerance. Exhaustive claims apply only to explicitly bounded domains, with no semantic accuracy or general Rust proof.

fallback-tie-intervention

Fix q={anchor} and the seven-token fixture ruby jade anchor the opal and pearl, with only ruby/jade/anchor/opal/pearl content. Runs have content counts (2,2) and token lengths (2,4). A chooses ruby jade; a token-length maximizer would choose the later run and is excluded. Append content amber to the second run: counts become (2,3), production yields the opal and pearl amber. A third question mask leaves every run below two content words and yields None. This is an exact three-call intervention relation with finite observed equality, not a general optimization or semantic superiority claim. Test fallback_content_objective_and_earliest_tie.

candidate-expansion-and-retry

Fix the eight-token try/take/ruby/give/jade/seek/pearl/they tree in verb_children_expand_one_xcomp_level_only. Helper output is [take,give,ruby,seek,they]: direct children precede exactly one level of direct Xcomp children; Conj grandchildren and deeper Xcomp grandchildren are excluded, while an independently supplied Xcomp subject is retained. Obj selection observes ruby with obj label. Separately two three-token fixtures differ only in Obj/Iobj versus Obl/Obl relations; with ruby echoed, direct selection returns None while the oblique retry loop returns jade/obl. These observations constrain parent wiring, order and cleanup failure propagation; they do not infer general branch reachability. Oracle arrays and optional text/rule outputs are literal and independent of production selection loops. Tests verb_children_expand_one_xcomp_level_only and object_shadowing_and_oblique_retry_are_distinct.

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

📐answer-span

Facts

Components

Obernants

Parts

Model text

model

answer-span

forward

invariant

cost

epoch-memory

checks

composition

behavior-complement

reverse-reconstruction

routing-fiber

predicate-priority

cleaning-fiber

early-return-shadow

numeric-bypass

location-time

reason-horizon

feeling-envelope

fallback-objective

rule-fiber

native-epoch-2026-10-30

evidence

Answer-span evidence

Executed checks

Reproduction

Source review observations

Native epoch, 2026-10-30

wave6

Native selection observations, 2026-10-30

fallback-tie-intervention

candidate-expansion-and-retry

answer-span