Elamur Groupantsmath-qworld-learned
The groupant beside math/src/qworld.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 | approximant |
| ast | partial | ../syntax.json |
| descriptant | present | descriptant |
| invariant | partial | invariant |
| tests | executed bounded checks | tests |
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 |
zero_weight_boundary | zero_weight_boundary | partial source derived model | zero_weight_boundary |
greedy_cluster_history | greedy_cluster_history | partial source derived model | greedy_cluster_history |
Model text
descriptant
math-qworld-learned
script_labels visits numeric script occurrences and binds each to the first unused equal exact nonconstant quantity. WorldModel predicts argmax over ['=','+','-','r']; learned cells join the highest-scoring prior pair only if score>0.
Epoch 2026-10-31: replaced duplicate catalog sentences with concrete exclusions, compatible-antecedent examples and two distinct scoped behavioral views. Source and prior execution evidence are unchanged; this revision ran no tests.
approximant
Scoped forward relation
script_labels visits numeric script occurrences and binds each to the first unused equal exact nonconstant quantity. WorldModel predicts argmax over ['=','+','-','r']; learned cells join the highest-scoring prior pair only if score>0.
This source-derived relation is partial; no executable percentage surrogate or replacement is supplied.
invariant
Scoped domain and invariants
Labels must align with qs; floating weights/scores must permit partial_cmp. Training uses10 passes and does not guarantee separation of conflicting labels.
Cost
Training scans item quantities and all i<j pairs for10 epochs; learned write compares each item with all earlier ones. Feature string/hash and exact cell work add to O(NĀ²) pair enumeration.
models
behavior_complement
No source quantity occurrence can be labelled twice by script_labels. A pair with score0 cannot join a prior cell. At all-zero operation weights, a non-rate prediction is excluded because max_by keeps the last tied r.
reverse_reconstruction
Two script lines assigning the same number to different slots can match duplicate-valued quantities in order; swapping otherwise equivalent lines changes occurrence labels. Many scripts differing in ignored lines produce the same labels. Distinct feature-weight maps can share all selected op/pair scores.
zero_weight_boundary
Default WorldModel chooses r for every operation; write_learned then emits commands but leaves state empty. Pair score0 creates a new cluster instead of joining.
greedy_cluster_history
For item i, choose the last maximal positive prior-pair score and reuse that prior item's assigned slot; cluster names come from the first item that created a slot, not the newest item's description.
tests
Scoped evidence
Greedy script duplicate binding and default learned rate/no-cell dispatch; no training.
Named actual source fixtures: qworld::wave28::{exact_world_overwrites_and_labels_duplicates}.
Parent tests.md records commands and paired profiles. No additional coverage inferred from parent fixtures outside this byte range; source-scoped numeric/parse properties remain partial.
This page in Markdown ā https://md.groupants.elamur.ai/free/en/mova/math/src/qworld.rs.groupant/learned.groupant/