📐world-big

The groupant beside world/src/big.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

Components

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.

Parts

• world-big-sampling

Model text

descriptant

world-big

ids scans store indices ascending, book eligibility from fourth TSV column and h(i)=high4(iK mod2^64), K=0x9E3779B97F4A7C15. If C<=n return all candidates, otherwise choose candidate[floor(kC/n)] with f64 arithmetic for0<=k<n. parse keeps5..40-token sentences, annotates/hand-repairs and restores index order. sample caches Arc by exact name, at most4 entries, evicting first nonguard else oldest. judge_changes alternates A/B then counts every accepted response line.

Modeling epoch 2026-10-31

Prospective source/branch/cost hypotheses were saved externally before source edits. This partial source-derived model retains observed boundaries rather than claiming complete verification. Only domain/group runtime test; sampling, cache and external judgment source-modeled, no executable surrogate.

approximant

Forward relation

ids scans store indices ascending, book eligibility from fourth TSV column and h(i)=high4(iK mod2^64), K=0x9E3779B97F4A7C15. If C<=n return all candidates, otherwise choose candidate[floor(kC/n)] with f64 arithmetic for0<=k<n. parse keeps5..40-token sentences, annotates/hand-repairs and restores index order. sample caches Arc by exact name, at most4 entries, evicting first nonguard else oldest. judge_changes alternates A/B then counts every accepted response line.

Domain: Valid store and metadata, representable lengths and heads; only pure name/group mapping executed. No store, annotator, cache, manual or teacher execution.

No executable percent surrogate: these categorical/effect observations lack a useful approximation metric. Future supported caps1/2/5/10% need a defined domain/metric and real reference evidence; no error or speed bound is claimed.

invariant

Output ordering follows sampled store index, not worker completion. Name prefix big-tale maps tale, everything else real; big-worlds uses offset bucket%16 over manuals. Cache names omit sample size and content version. Prompt word index zero-based; head0 root, positive heads1-based and unchecked.

Cost component: cost.md.

cost

O(S log B+B log B+C) sampling time, O(B+C) memory; n limits returned ids not full candidates. Selected parsing delegates tokenizer/annotator/hand costs, <=10workers, four variable-byte cache entries. Teacher/I/O no latency bound, usize/f64 large inputs uncertified.

composition

family fresh names @2..15 and guard names feed induce::explore_big; parsed sample has hand repairs but no induced rules. induce::changed_words coordinates feed judge_changes. Domain bucket1 differs from world-manual offset1.

models

behavior_complement

bucket>=16 yields no ids; n0 yields empty after store scan. Cache hit excludes re-reading new MOVA_BIG_N/store environment. Guards may be evicted if all entries are guards. Teacher totals may exceed item count because repeated numbered responses count repeatedly; malformed/out-of-range rows skipped, unrecognized labels count neither.

reverse_reconstruction

Observed ids admit any sentence contents at same eligible indices and arbitrary skipped candidates consistent with stride. Empty admits n0 or no eligible books or invalid bucket. Cache Arc admits many changed environments after insertion. Teacher tally admits many response orderings and duplicates, not unique verdict reconstruction.

world_big_invariants

Output ordering follows sampled store index, not worker completion. Name prefix big-tale maps tale, everything else real; big-worlds uses offset bucket%16 over manuals. Cache names omit sample size and content version. Prompt word index zero-based; head0 root, positive heads1-based and unchecked.

world_big_resource_cost

O(S log B+B log B+C) sampling time, O(B+C) memory; n limits returned ids not full candidates. Selected parsing delegates tokenizer/annotator/hand costs, <=10workers, four variable-byte cache entries. Teacher/I/O no latency bound, usize/f64 large inputs uncertified.

world_big_representation_join

family fresh names @2..15 and guard names feed induce::explore_big; parsed sample has hand repairs but no induced rules. induce::changed_words coordinates feed judge_changes. Domain bucket1 differs from world-manual offset1.

world_big_domain_partition

Valid store and metadata, representable lengths and heads; only pure name/group mapping executed. No store, annotator, cache, manual or teacher execution.

tests

Finite actual-code evidence — 2026-10-31

Names: ["big::wave29_worldlearn::sample_name_domain_and_groups"]

Selected CPU command: CARGO_BUILD_JOBS=2 CARGO_TARGET_DIR=ISOLATED_TARGET cargo test --offline --locked -p world --lib --example retention_probe [--release] wave29_worldlearn -- --test-threads=1

15 selected cases passed twice debug and twice release with identical sorted names/outcomes, exits0. control-big: compiled source mutation failed with exit101; source restored byte-for-byte and15positive tests passed. No live teacher, corpus, state activation, GPU, or full-school run. Temporary synthetic files use unique process-id paths and are removed after passing tests.

Only domain/group runtime test; sampling, cache and external judgment source-modeled, no executable surrogate.

Source SHA256: afe8d7cc25363ac760140ae959cb97a249db63c0e8f7422d02abb01ca3e9c8eb. Syntax generated independently twice and compared; lexical only, no semantic proof or CUDA compilation. Strict whole-crate Clippy was not repeated merely to reproduce existing dependency/source lints; Cargo warnings retained in receipts.

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