📐prag-model

The groupant beside prag/src/model.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

• prag-model-gain
• prag-model-perceptron
• prag-model-tree
• prag-model-knn

Model text

model

prag-model

domain

Aligned finite supplied x/y/keys, SLOTS-width valid dictionary coordinates (UNSEEN allowed at prediction), nonempty names of at most 65536 classes, valid class indices, representable offsets/counts and finite weights. Lower-level gain and distance helpers accept their explicitly stated dimensions. Fixed target arithmetic and successful allocation are premises. Macro bodies and semantic AST analysis are absent.

forward

fit dispatches a shared Train to one of three learners: perceptron with 12 epochs, IGTree, or kNN. A common encoded row then produces Pred{class,conf,why}. The child gain function supplies both tree slot ordering and kNN overlap weights. Class indices mean exactly the supplied names order; Dict slot/value coordinates and sentence key order must remain aligned across fitting, prediction and explanation.

invariant

All class indices stay within nonempty names for valid models. Repeated same supplied input/seed on a fixed target must preserve prediction observations. conf is a margin for perceptron, local support for IGTree and vote share for kNN, so confidence is not a shared calibrated probability. bytes serializes full perceptron but selected core tuples for IGTree/kNN; it is not full resident-memory usage.

cost

Dispatch cost equals selected child training plus allocation, not sum of all learners. Prediction chooses one child. Shared gain work is charged once within that learner. Serialization allocates bytes; info formats metadata. No training-throughput or accuracy measurement is claimed.

behavior-complement

For any two-class zero-weight perceptron and equal-scoring row, class 1 is absent; the tie selects 0. For the supplied three-row tree, absent first-slot value cannot produce the minority class 1. For positive k and fixed four-row kNN memory with two exact matches, a prediction counting only one exact match as its vote denominator is absent. These algorithm-tagged transition families define a concrete source-derived inner behavior complement in the universe (Kind,valid state,row,Pred); malformed state and untested arithmetic extremes remain unknown.

reverse-reconstruction

An observed class 0 may come from zero perceptron ties, tree majority fallback or kNN tie resolution. Even Kind plus conf/why leaves inactive dimensions and training histories ambiguous. Candidate reconstruction must retain shared names order, dictionary coordinates, row/key alignment, seed and child constraints. Without these correlations an invented tree default or nearest key is not a compatible antecedent.

specialization

Confidence-observation projection intentionally erases kind and explanation. Equal numeric confidence then has incompatible meanings across learners (margin versus support versus vote fraction); it cannot support direct probability calibration. The public tiny fit test retains kind and compares class/conf/why and byte counts across two independent fits; actual eval::fit_set and predict_seq use common dictionary coordinates and per-field names.

composition

Shared gain ratios determine tree order and kNN weights. Perceptron reads the same Dict coordinates independently. Caller eval::fit_set binds one label/name vector per Field to common rows/keys; predict_seq overwrites prev-slot from its prior predicted act. That caller path is inspected here; this leaf executes Dict::build/encode -> fit -> predict on three supplied rows only. Child fragments partition production mechanics; test module remains at the file root. No parser/teacher/corpus/held-out/training-job execution.

checks

Five native tests in model::wave13_models check canonical histogram arithmetic, literal one-update averaging, tree pruning/fallback, 80 kNN shell-label cases and repeated public fit observations. Exact names and evidence hashes are in root evidence.md. Two debug and two release runs must have equal normalized named outcomes. Negative production control changes shell admission and must fail a named assertion. No automatic inverse solver, general proof, calibrated probability claim or percentage-error approximant.

epoch-memory

2026-10-30 initial scoped model. Prospective prediction identified randomized floating reduction in gain_ratio. Old implementation produced nine f64 values on 64 identical histogram calls. BTreeMap now fixes ascending category order; the expected grouping cost increases to O(N log V), with no timing claim. Other reviewed semantics were retained: perceptron averages initial state, tree pruning can lower confidence, and kNN explanations can extend beyond voters. Malformed state, huge counters, long-run floating accumulation, serialization durability and semantic accuracy remain unverified. All typed records bind the exact source hash in the manifest; they are partial source models supported by finite native checks, not universal proofs.

evidence

Native evidence, 2026-10-30

Five tests in model::wave13_models pass twice in debug and twice in release with equal normalized named outcomes (5/5 each, no ignored cases). Raw logs live under $MOVA_DATA/runs/octopus-wave13-model-2026-10-30/. These are tiny supplied-input checks, not the legacy 200-row training test, a corpus evaluation or live training job.

Old production failed gain_ratio_canonical_histogram_oracle: 64 identical calls produced nine different f64 bit patterns. Ascending category accumulation repairs this target-local nondeterminism. The earlier canonical-value assertion also failed by one low bit; both logs are retained. No semantic prediction-change or speedup claim follows from these bit differences.

The compiled production mutation distinct > k to distinct >= k fails knn_distance_shells_and_frequency_ties at its mask=0, k=1 assertion. Source bytes were restored exactly; all five positive tests pass again.

The independent histogram test has 17 category histograms and analytic empty/zero/one endpoints. Perceptron uses one example and one update with literal expected margins; tree uses three supplied rows and literal fallback/pruning behavior. kNN enumerates 16 labelings x five shell counts = 80 vote checks, plus skip/UNSEEN and prior ties. The public fit join trains only three supplied rows twice per kind. Two syntax projections are byte-identical. Four fragments plus root declare 15 typed views; this count is not proof coverage.

Actual caller join: eval::fit_set shares x/dict/keys across fields, then predict_seq overwrites prev-slot by prior predicted act. The leaf executes Dict::build/encode -> fit -> predict; broader per-field and previous-sentence joins belong to integration. No filesystem durability, parser, teacher, corpus, GPU, held-out or calibrated probability evidence. Existing unrelated compiler warnings remain.

Bindings

Final source SHA-256: e18c587fd1f87b18c7e0118e0009b16fb469e26ba9cd0edf26b097226130ee25

• predictions.md: b06357400771b84108bc1b5b572433bd117118c78c05b11f322ec25b7a7cec99
• old-code.log: 95fef9db8366108136d5a2f42c36968817e83235c77fbd715173f09e6fbc1928
• old-nondeterminism.log: eca4cb464efa87b1c82794507d18b6ccceca1f4027fa4f97a4db58175b3fe9e1
• debug1.log: 0f6a8a64366821793e69e63e16a5f9906b138acd368c97a9070e5341c4891ff1
• debug2.log: 5bea463837ce5aaa92fef411d824f1e5e1d5dd94719a0b75e8434e0e5cec1d30
• release1.log: 3a10dab388cbbfa889e682893c13edfd1404c1165cc8f67feece939609281052
• release2.log: 2dbd7f8d7caf930b3077ab5395a12645ce159ad6eeabb32027003441a56a0e05
• mutant-shell.log: 8a1a91e2b812f342ce657d3ac26f01ab9849bbecf235eff0f6cfbed5182cb34c
• restored.log: 783d61c3ee7704aa44b5e1173c485d1bd6762256952535ad83f7769b728857c0
• normalized.txt: 70bd9773789e1f09a75441c4e5f1a7c554994418c32c4df5af4cbe6505a44d88
• syntax1.json: d1e176089771c44de040b35da944811ab1c38e54290d99e0adcfe3978d86662c
• syntax2.json: d1e176089771c44de040b35da944811ab1c38e54290d99e0adcfe3978d86662c

predictions

Prospective model review, 2026-10-30

Domain: supplied in-memory aligned x/y/keys; rows have SLOTS.len() entries; positive class count <=65536, valid labels and Dict coordinates; representable counters/features; finite weights. No parser, teacher, corpus, GPU or training job.

Forward: gain ratio aggregates label histograms per slot value then computes entropy reduction divided by split entropy. Perceptron uses active known indicators plus bias, seed-hashed example order and timestamp averaging. IGTree sorts slots by gain ratio and index, retains plurality defaults (prior then smaller class breaks ties), prunes same-default leaves. kNN sorts weighted-overlap distances then memory index, includes all examples in first k distinct distance shells, and chooses k by leave-one-out accuracy with smaller candidate on ties.

Suspected witnessed bug: gain_ratio sums floating contributions over randomized HashMap iteration. Equal data can produce different low bits, potentially changing tied slot order or weighted distance boundaries. Predict deterministic ascending slot-value accumulation fixes repeated bits; no signature change. Cost changes expected grouping O(N) to O(N log V); traversal still O(VC), memory O(VC). No speedup claim.

Invariants: unknown values are inactive in perceptron; UNSEEN mismatches even itself in kNN; deterministic smaller-class tie; IGTree missing edges use current node default; k counts distances, not rows. Prediction confidence differs by model and is not comparable probability calibration.

Tests will use analytic entropy endpoints, an independent sorted-value histogram oracle, literal single-step averaged weights, literal tree defaults/pruning, exhaustive small kNN shell histograms and tiny public fit/dictionary joins. Mutation must break a named assertion. Empty classes, malformed lengths/serialized state and overflow-sized training remain outside valid-input model. Actual eval::fit_set passes common x/dict/keys and one label vector per Field; predict_seq supplies encoded rows and previous predicted act. Only tiny in-memory API joins will execute.

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