Elamur Groupantsworld-stepgame
The groupant beside world/src/stepgame.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 |
| composition | partial | composition |
| cost | partial | Cost |
| descriptant | present | descriptant |
| invariant | partial | model.md |
| tests | bounded checks partial scope | 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 |
tie_selection | tie_selection | partial source derived model | tie_selection |
lexical_ground | lexical_ground | partial source derived model | lexical_ground |
training_schedule | training_schedule | partial source derived model | training_schedule |
empty_evaluation | empty_evaluation | partial source derived model | empty_evaluation |
Parts
Model text
descriptant
world-stepgame
load_txt resets at literal 1 and clones story on tabbed questions. JSON defaults absent fields, maps invalid numeric ids to zero, stably sorts numeric ids. read prioritizes perceptron for seen phrasing, otherwise lexical unless environment disables it. Scores tie by lowest class index. answer requires two distinct query agents, builds reciprocal edges, BFS from reference and projects displacement to sign class. train uses deterministic stride and integer accumulated weights; eval/contrast perform dataset/report effects. Equations: score_k=Σ_{f in feats}w[f,k], pred=min argmax_k score_k; p[q1]=0; vacant p[y]=p[x]+v_xy; label=LABELS[(sign p[q0].x, sign p[q0].y)]. lexical axes return None if both signs occur, otherwise first nonzero sign or zero; ground=a1 negates both axes.
Modeling epoch — 2026-10-31
Read full assigned source before editing; prospective predictions retained in external wave29 receipts. Stored/live/parser-dependent test paths are excluded from execution. Evidence is limited to named bounded fixtures; registration is partial modeling. Production behavior unchanged; only inline synthetic tests added to four assigned sources.
approximant
Forward relation
load_txt resets at literal 1 and clones story on tabbed questions. JSON defaults absent fields, maps invalid numeric ids to zero, stably sorts numeric ids. read prioritizes perceptron for seen phrasing, otherwise lexical unless environment disables it. Scores tie by lowest class index. answer requires two distinct query agents, builds reciprocal edges, BFS from reference and projects displacement to sign class. train uses deterministic stride and integer accumulated weights; eval/contrast perform dataset/report effects. Equations: score_k=Σ_{f in feats}w[f,k], pred=min argmax_k score_k; p[q1]=0; vacant p[y]=p[x]+v_xy; label=LABELS[(sign p[q0].x, sign p[q0].y)]. lexical axes return None if both signs occur, otherwise first nonzero sign or zero; ground=a1 negates both axes.
Domain and resource limits
Finite valid structures and allocation-success executions under the branch conditions in models.md; integer overflow is outside measured fixtures. Exact categorical relations have no supplied percent surrogate. Macro expansion, arbitrary model quality and external effects are not inferred from syntax.
invariant
Invariant
LABELS order is tie order. Coordinates mean first agent relative second, not meters. Deduplicated single ASCII capital agents only. finish retains scaled sums, not divided average; avg is not a read switch.
Cost
Features O(B) allocated ngrams; score O(9F) expected hash operations; lexical worst O(T²) backward cue scanning; BFS O((V+E)logV) plus clones. Training O(epochsNF*9); stride not a permutation if gcd(N,7919)>1.
composition
Composition
Scoped children: formats, lexical, graph. load_txt resets at literal 1 and clones story on tabbed questions. JSON defaults absent fields, maps invalid numeric ids to zero, stably sorts numeric ids. read prioritizes perceptron for seen phrasing, otherwise lexical unless environment disables it. Scores tie by lowest class index. answer requires two distinct query agents, builds reciprocal edges, BFS from reference and projects displacement to sign class. train uses deterministic stride and integer accumulated weights; eval/contrast perform dataset/report effects. Equations: score_k=Σ_{f in feats}w[f,k], pred=min argmax_k score_k; p[q1]=0; vacant p[y]=p[x]+v_xy; label=LABELS[(sign p[q0].x, sign p[q0].y)]. lexical axes return None if both signs occur, otherwise first nonzero sign or zero; ground=a1 negates both axes. Other functions and test additions are modeled only at this parent. bAbI Story and command World have different representations: command-language fixtures do not certify the UD parser adapter. StepGame lexical/load helpers connect directly to tests03; CLI helpers construct a separate world process but this wave does not spawn it. No composition-wide percent accuracy follows.
models
behavior_complement
Exclude lexical vectors for clock tokens, conflicting signs on an axis or directionless text without sameness; exclude answer for invalid agent count or unreachable target. Contradictory cycles still return first path. Zero perceptron intentionally yields left, not abstention. C_answer includes (q,G,Some(label)) whenever |agents(q)|≠2 or q0∉Reach_G(q1). C_lex includes (s,Some(v)) whenever clock(s) or axis_conflict(s).
reverse_reconstruction
Left label admits every graph with first BFS displacement x<0,y=0; explanation narrows to exact d but never unique story. phrasing collapses letter renamings/punctuation/case. load_txt loses support fields and original numbers. Predicted class k admits weight states where k is smallest maximizing score. Pre(left)={ (R,I) | BFS_edges(R,I)(q1)[q0]=(x,0), x<0 }.
tie_selection
Empty/default scores choose index0 left; repeated feature tokens count repeatedly.
lexical_ground
Unique cue wins then possessive subject then a2; right before under is intensifier.
training_schedule
t increments even for correct prediction; finish sums (t-last)*weight. Integer overflow is not certified.
empty_evaluation
Printed raw percentages can NaN at zero; contrast modulo-zero possible with no eligible templates and samples>0. Live data paths excluded.
tests
Bounded checks
Named checks associated with this source (not every fragment branch):
- stepgame::wave29_tests::lexical_ground_and_tie_boundaries
- stepgame::wave29_tests::graph_first_path_and_text_reset
- tests03::stepgame_lexical_reader_reads_new_phrasings
- tests03::stepgame_phrasing_and_corrected_format
17 unique named tests in total passed twice debug and twice release; sorted named outcomes equal across all four runs. Four compiled controls cover bAbI identity, StepGame tie ordering, parser-command state removal and bench-command CPU selection; each exits101 at a runtime assertion, exact source restored then positive exit0. Other sources have finite adversarial fixtures but no separately compiled mutant.
CLI integration coverage is command construction and temporary cleanup only; legacy school/model/cadence paths are not executed. No speed benchmark, GPU, live teacher, datasets or held-out replay.
Limits:
- Partial syntax projection only; no macro expansion or proof
- No useful percentage surrogate for exact categorical/state/effect semantics; executable approximant unsupported
- No dataset training, eval/contrast or GPU
- No overflow/allocation fault injection
External receipt root: $MOVA_DATA/runs/octopus-wave29-worldeval-2026-10-31.
This page in Markdown → https://md.groupants.elamur.ai/free/en/mova/world/src/stepgame.rs.groupant/