Elamur Groupantsmath-orchestration
The groupant beside models/math-orchestration.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 |
expectation-join | representation_transport | partial source derived model | expectation-join |
fallback-state | state_transition | partial source derived model | fallback-state |
effect-order | failure_precedence | partial source derived model | effect-order |
budget-composition | resource_boundary | partial source derived model | budget-composition |
Model text
descriptant
math-orchestration: wave28
Problem kind/source→Expect integer/options→gate; plan/curriculum normalizers and tolerances differ. Template substitution lossy then gate rechecks. CLI held policy outside run_c. Vmm cap precheck not reservation. Build output enters lex module; suite uses separate small-unit engine.
This source-bound partial group organizes existing Mova code. Native evidence and excluded effects are recorded in source sidecars and the root wave28 report. Registration is not a universal proof or whole-system runtime claim.
model
math-orchestration
forward
Problem→Expect→Plan→Trace; verified=fails.empty&&answer.Some. B.correct=verified&&grade. Verified B→Memo regardless correct; C tries<=3 compatible templates then gates else copies B fields. CLI selects data/effect routes; Vmm/Lean explicit external boundaries.
invariant
Verification distinct from gold correctness. C fallback keeps by_slm=false. Strict over-cap and forbidden-proof preflight exclude corresponding call effects only under actual early return guards.
cost
Sampling O(n log n), delegated arithmetic/enumeration, perceptron training and bounded template attempts; teacher retry multiplier max_retries+1, CLI training folds/epochs. No global wall-clock or atomic budget bound. enumerate polls time every4096 visited leaves; claim caps only condition-passing cases, has no leaf-level time poll, and delegates evaluator checks.
composition
Problem kind/source→Expect integer/options→gate; plan/curriculum normalizers and tolerances differ. Template substitution lossy then gate rechecks. CLI held policy outside run_c. Vmm cap precheck not reservation. Build output enters lex module; suite uses separate small-unit engine.
domain
Bounded synthetic helpers/plans/records. Public module compile-only, classic CLI usage-only. Live teacher/Lean/models/corpora/held/DB/GPU and large benchmark excluded.
limitations
Finite test floats, local compiler/filesystem; delegated arithmetic retains own limits. Lexical AST lacks semantic resolution/macros. No-panic perceptron comparisons require comparable non-NaN accumulated scores and weights; finite weights alone do not prevent score overflow/NaN.
checks
- paired syntax for14 legacy sources
- twice debug/twice release native selections and normalized outcomes
- compiled assertion-inversion controls, exact restoration and rerun
catalog
behavior-complement
Verified B excludes failures/missing answer but not wrong NL model. Over-cap/Lean-ban preflights exclude only guarded effects.
reverse-reconstruction
Records admit all antecedent plans/prompts producing those observations; template and summaries lose values/order/metadata so inverse is many-valued.
expectation-join
Expect integer for int or number from TheoremQA, options verbatim. Numeric gate and grading differ in tolerance/normalization.
fallback-state
Fallback copies answer/option/verified/correct/cost/tokens/secs/method, not B trace/history; keeps C config and by_slm=false.
effect-order
Vmm cap→mkdir→teacher→spent→append→is_error; Lean bans→write→process→diagnostics. Failed append not rollback.
budget-composition
Per-plan limits/reserve estimates do not bound parallel composition; concurrent prechecks can overspend.
This page in Markdown → https://md.groupants.elamur.ai/free/en/mova/models/math-orchestration.groupant/