Elamur Groupantsworld-rule-learning
The groupant beside models/world-rule-learning.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
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 | forward |
| ast | not materialized | |
| composition | partial | composition |
| cost | symbolic source derived | cost |
| descriptant | present | descriptant |
| invariant | partial | invariant |
| tests | see source local evidence | 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 |
|---|---|---|---|
behavior-complement | behavior_complement | partial source derived model | behavior-complement |
reverse-reconstruction | reverse_reconstruction | partial source derived model | reverse-reconstruction |
rule-scope-join | representation_transport | partial source derived model | rule-scope-join |
count-kernel-join | coordinate_transport | partial source derived model | count-kernel-join |
brain-book-join | projection_loss | partial source derived model | brain-book-join |
effect-order | failure_precedence | partial source derived model | effect-order |
Model text
descriptant
world-rule-learning: wave29
family::change_line/change_of uses induce::Rule::machine/parse_rules, carrying scope separately. in_domain and apply_scoped transport changes without commutative-repair guarantee. big sample hand repairs precede induced rules; GPU counts only eligible correct words. worldbook::gather -> write -> brain_body feeds family brain; knowledge English serialization feeds state rules. retention_probe delegates Retained state, not legacy know cache.
This source-bound partial group organizes existing Mova code. Native evidence and excluded effects are recorded in source sidecars and the root wave29 report. Registration is not a universal proof or whole-system runtime claim.
model
world-rule-learning
forward
big sample -> induce ordered rule scores -> family scoped change -> state guard; family calls worldbook principles/write/brain_body and know English rule learning. Fix=cardinality(head-correct,label-wrong,subset-match), Break=cardinality(head-and-label-correct,subset-match); candidate order=(-Fix+Break,atomcount,Rule). D_tale(sc)=all; D_other(sc)=general. Worldbook merge key=(lower subject,predicate,lower text); retained brain projection truncates each part12 but counts all lines.
invariant
Label repair preserves heads/forms; scope replacement preserves scopes/positions. Gold/big/knob gates are distinct, no cross-domain guarantee from leaf counts. Cited count differs from emitted brain lines. CPU tests do not prove CUDA kernel equivalence.
cost
Subset counting W*sum C(A,k),k1..3 plus candidate sorting; family composes repeated parsing/guards; worldbook multiple corpus scans/triples/ranking and whole output; know fact indexing/derivation/ranking. Effects dominate unmeasured latency; no speed claim.
composition
family::change_line/change_of uses induce::Rule::machine/parse_rules, carrying scope separately. in_domain and apply_scoped transport changes without commutative-repair guarantee. big sample hand repairs precede induced rules; GPU counts only eligible correct words. worldbook::gather -> write -> brain_body feeds family brain; knowledge English serialization feeds state rules. retention_probe delegates Retained state, not legacy know cache.
domain
Finite synthetic tests and explicit temporary fixtures; rule atoms within internal21bit encoding for subset claims, aligned sentence/feature arrays, bounded counters and valid tree heads for traversal. No live corpus/teacher/school/GPU effects.
limitations
Runtime fixtures cover pure helpers plus fresh list/book and early CLI failures. Nonempty CUDA route source-modeled only; no kernel compile/execution. Actual training/state transitions/assessments and successful retained probe unexecuted.
checks
- 15named cases twice debug/twice release, all exit0 and equal sorted outcomes
- 6compiled source controls exit101 with assertions; exact restore and15positive rerun
- 7source syntax projections each generated twice identically;11scoped fragments share parent syntax
catalog
behavior-complement
Wrong-head words never count fixes/breaks, but helper explore_changes alone can return nonsignificant p; family must enforce source-specific thresholds. Scope adapter cannot promote an existing equal domain rule by Add. Brain13th line cannot be emitted though it increments count. pd=false excludes source quote; no successful graph/teacher effects are inferred.
reverse-reconstruction
Joint outputs admit input fibers: different word orders with equal subset counts; scoped rule lists differing in removed-rule multiplicity; source facts differing in omitted citations or beyond12 brain entries; different retained histories with same projected reports. None uniquely recovers corpus, rule history or full source graph.
rule-scope-join
Change carries Rule but not scope; family separately serializes scope and reattaches it. Replace retains each preexisting scope. Applying ordered labels need not commute, so independent candidate gains do not sum; combined child guard required.
count-kernel-join
CPU Feats[si][i] nonzero21bit codes -> GPU flat codes plusu32 offsets and enum labelu8->u32; group dense index determines count slot. Correctness assumes offset/count widths and unique table keys; no GPU execution witness.
brain-book-join
worldbook write emits<=3citation notes/fact, brain_body takes first note and<=12lines/part but returns total cited lines. family uses total as report and truncated body as stored knowledge; neither recovers all Fact attrs/support.
effect-order
Family hot write precedes state transition/guard; ignored rollback errors prevent transaction claim. Retention probe rejects argc before reads, parses train before exam, saves before equality. Teacher failure and filesystem failure are not numerical approximation errors.
This page in Markdown → https://md.groupants.elamur.ai/free/en/mova/models/world-rule-learning.groupant/