📐world-evaluation-boundaries

The groupant beside models/world-evaluation-boundaries.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

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.

Model text

descriptant

world-evaluation-boundaries: wave29

tests03's normalized phrasing relation must preserve above/below distinction while identifying agent renamings; load_txt reset is literal '1', not numeric equivalence. StepGame score ties map to left, so replacing missing model weights with abstention changes graph answers. bAbI relation adjacency uses subtraction, unlike StepGame's directly signed reciprocal graph; an adapter that copies edge orientation would invert displacement. Command parser errors happen before apply in fixtures; bAbI read increments t before parser failure and is not transactional. CLI configuration checks cannot establish legacy school failure/publication semantics.

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-evaluation-boundaries

forward

Three distinct evaluation boundaries compose: bAbI text → UD-backed Story updates/questions; StepGame text/JSON → Item → lexical-or-perceptron edge vectors → first-discovered BFS displacement → sign label; command-language fixtures → lang::parse_all → World::apply/run_v3 → snapshots, gates and qa verdicts. tests03 directly exercises StepGame lexical/load normalization and induction Change coverage. parser_failures/school_bench construct a separate world CLI Command with isolated paths; this wave tests that construction, not child execution. No direct adapter equates bAbI Story with command World.

invariant

Bounded selected observations retain once-only speaker identity, holder/carry coherence, first-path vector orientation/tie order, strict grammar and rejected-command version observations, canonical-versus-claim/dream separation and partial temporal order. Command helpers explicitly remove MOVA_STATE/PRAG_OFF and scope hot/bench roots to Scratch. These are finite observations, not all-input guarantees or persistent-school measurements.

cost

Text/token normalization linear in traversed bytes plus emitted/cloned stories; StepGame lexical cue scan worst quadratic in token count and BFS O((V+E) log V) with ordered maps and adjacency clones. bAbI reach O(VE), histories O(H log H) when sorted, parser cost delegated. Synthetic world tests delegate gate/state/time closure work. CLI helper cost fixed path/env construction plus temporary directory cleanup. No timing, throughput, numerical error or end-to-end task score claim.

composition

tests03's normalized phrasing relation must preserve above/below distinction while identifying agent renamings; load_txt reset is literal '1', not numeric equivalence. StepGame score ties map to left, so replacing missing model weights with abstention changes graph answers. bAbI relation adjacency uses subtraction, unlike StepGame's directly signed reciprocal graph; an adapter that copies edge orientation would invert displacement. Command parser errors happen before apply in fixtures; bAbI read increments t before parser failure and is not transactional. CLI configuration checks cannot establish legacy school failure/publication semantics.

domain

Finite UTF-8 literals, bounded synthetic Word/Item/World fixtures, coherent internal Story maps and available host temp directory; default CPU build. Integer operations within range and successful allocation. Selected library tests do not load parser models, datasets, saved VMM test fixtures or teacher. Integration helpers never spawn world binary.

limitations

Global compiled relation metadata supplies north/east and inverse/step knowledge for bAbI graph fixtures; it is read-only compiled source data. General bAbI read/ask and live parsing remain unexecuted. Tests.rs saved Wen and natural-language model paths, tests3 saved merchant, legacy parser_failures/school_bench school commands, StepGame train/eval/contrast are excluded. No percentage executable surrogate metric for these categorical/effect relations.

checks

• 17 exact named tests, twice debug and twice release; sorted named outcome equality across all four
• Four compiled runtime negative controls: identity, tie ordering, parser env removal, bench CPU mode; exit101, byte restoration and exit0 rerun
• Seven paired identical syntax projections; 14 scope-specific fragments with UTF-8 ranges and hashes
• native-commands.json and source-test-map.json retain exact selections
• External verify.rs checks 21 manifests, hashes, scopes, relative projections and exact 17-name outcome set

catalog

behavior-complement

Exclude a composed StepGame answer when query agent count is not two or target unreachable; exclude lexical certainty for clock/conflicting-axis phrases. Exclude replacing known Story speaker and dropping another holder. In selected command fixtures exclude grammar acceptance of listed malformed commands, observed rollback growth, claim/dream changes to canonical observations and invented total temporal order. Exclude explicitly configured state/prag flags in child helpers. Do not exclude bAbI pre-error t increments, contradictory first-path graphs, or unexecuted legacy CLI side effects.

reverse-reconstruction

StepGame label preimage consists of all Reader/Item graphs with first-path displacement in its sign orthant; phrasing and load discard agent names/support/numbering. bAbI exact holder-map output admits all histories with identical final holder domains and equal last surviving ownership on every target key; an empty target requires empty final ownership, and state_lines omits speaker/attributes. Passing command assertions determines only observed projections and finite fixtures. Equal canonical world cannot reconstruct differing claims/dreams. Configured Command recovers explicit model/mode/path inputs but not ambient env or eventual subprocess effects.

coordinate-transport

bAbI offset from reference subtracts relation adjacency vectors; StepGame edge reference→subject adds Reader vector. Both mean subject minus reference but edge storage is opposite. Model each adapter explicitly; tests use c→a displacement(1,1) and C→A displacement(-2,0).

transaction-boundary

Command fixture world parses all lines before any apply; tests record unchanged observed state on rejects. bAbI read increments t before annotator and may update subjects before unknown-class failure. CLI Command setup has no binary startup; successful setup does not certify failed model handling.

observation-fibers

World test snapshots omit K for claim invariance and inspect only selected version/event observations for rollback. Partial time order preserves Unknown for incomparable events; no numeric duration inferred. Story state_lines is a different lossy projection and cannot be substituted.

execution-scope

17 selected bounded tests and four compiled controls exercise helpers/state/grammar. The two integration sources' legacy process-running tests remain excluded, as do model-backed readers and stored corpus fixtures. Scratch cleanup is observed under normal OS success only.

This page in Markdown → https://md.groupants.elamur.ai/free/en/mova/models/world-evaluation-boundaries.groupant/