Elamur Groupantscorpus-register-ranking
The groupant beside world/src/register.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 | Forward |
| ast | partial | ../syntax.json |
| cost | partial | Cost |
| descriptant | present | model.md |
| invariant | partial | Invariant |
| tests | bounded native checks | ../model.md#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 |
|---|---|---|---|
nonbehavior | behavior_complement | partial source derived model | nonbehavior |
antecedents | reverse_reconstruction | partial source derived model | antecedents |
finite-domain | domain_partition | partial source derived model | finite-domain |
sign | monotonicity_relation | partial source derived model | sign |
rates | format_projection | partial source derived model | rates |
Model text
descriptant
corpus-register-ranking
ranked_rows and run score formatting/output.
Forward
Evaluate pooled score, filter false comparisons including NaN, rank and render rounded values. run orchestrates counts and output.
Invariant
Retained keys belong to the union maps and satisfy both source comparisons. NaN z or NaN threshold excludes a row. Ordering is descending total_cmp(z), then lexical key when scores compare equal; count ties need not be score ties. Original u64 counts are retained in rows and compared exactly to min. Score/rate conversion to f64 can lose precision above 2^53. Finite corpus-swap antisymmetry is checked on bounded fixtures.
Cost
O(V log V), O(V) keys and rows.
Composition
This fragment has no children. ranked_rows receives maps/totals, computes pooled priors and scores, filters and sorts. run formats two-decimal z and one-decimal per-million rates. Reverse reconstruction retains exact compatible scores before formatting; write completion is a separate conditional trace.
Epoch memory
2026-10-30. Native cases cover 64 independently expanded score pairs, swaps, thresholds, ties and degeneracies. The large-singleton fixture demonstrates rounding can leave positive complementary denominators and admit z=0. The numerical algorithm is preserved; review corrected event-specific invariant prose. No executable approximation/error-cap claim.
Gaps
Syntax is source-bound and partial: macro expansion, names/types, cfg evaluation and external modules absent. Native fixtures exercise pure helpers and body slicing. Filesystem errors, invalid UTF-8 reads, formatted nonfinite rates and overflow remain unexecuted; large-count behavior has one targeted witness, not exhaustive coverage. No parser/annotator, sentence-splitting, live corpus, teacher or GPU execution is involved.
Exact-count correction
2026-10-30: independent large-count native witnesses fail on the former f64-to-u64 count round trip. ranked_rows now retains raw ca/cb, checks ca >= min and emits ca/cb exactly. The f64 z formula and NaN policy are unchanged. At ca=2^53+1 the old gate wrongly excluded equality; at ca=2^53+3 it could admit min=ca+1. The corrected native fixtures cover both directions.
catalog
corpus-register-ranking typed models
Subject: source hash in manifest, epoch 2026-10-30. Scope: ranked_rows and run score formatting/output.. Domain: terminating Rust calls without integer overflow, narrowed below. Models are partial source-derived relations, not proofs. Complement is an inner bound in the stated input/output universe. Reconstruction retains all compatible antecedents, not a unique inverse. Oracle status is explicit below; tests.md records executed coverage.
nonbehavior
Type: behavior_complement.
U=(maps,options,rows). Excludes admitted NaN, missing-union key or ascending finite-score pair. Oracle threshold/degeneracy tests.
antecedents
Type: reverse_reconstruction.
Exact tuple word,z,ya,yb admits totals and other counts satisfying sum and score equations; other words not uniquely identified. Two-decimal output loses more information. Oracle proposed finite fibers.
finite-domain
Type: domain_partition.
If 0<ya+alpha<na+1000 and same for b, log arguments positive finite and variance positive. Zero complementary mass can create infinities/NaN. Oracle 64 interior pairs and singleton.
sign
Type: monotonicity_relation.
For equal totals and finite domain, shared prior and monotone log odds imply sign(z)=sign(ya-yb); swap negates. Oracle native pair loops.
rates
Type: format_projection.
Rates ya1e6/na and yb1e6/nb print one decimal; empty reference may print nonfinite rate for finite admitted z. Header attempted books differs from successful reads. Oracle source-only formatting.
This page in Markdown → https://md.groupants.elamur.ai/free/en/mova/world/src/register.rs.groupant/ranking.groupant/