📐prag-model-perceptron

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.

Model text

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

A row activates one indicator per known in-range value plus bias. Scores are class-wise sums of weights over active indices. Training orders examples by h64(index,seed xor epoch), chooses largest score then smaller class, and adds +1 to gold and -1 to guessed class indicators on errors. Timestamp accumulator U records step*delta; final weights are W-U/(T+1). Prediction ranks scores, emits winning margin and up to five strictly positive winner-minus-runner contributions sorted descending with stable original feature order.

invariant

UNSEEN and out-of-vocabulary values do not activate lexical indicators; bias remains. Zero epochs yields zero scores, class 0, margin 0 and empty why. Positive explanations are a selected projection: their sum need not equal margin, since negatives and truncation are omitted. Single-class prediction has zero margin.

cost

F=sum dictionary sizes+1, C classes, N rows, S slots, E epochs. Dense state O(CF+NS); training O(ECNS+EN log N+CF), prediction O(CS+C log C+S log S). Hashing string indices adds bounded index-digit work. Representable feature offsets and finite weights are premises.

behavior-complement

For a fixed finite model and all-UNSEEN row, observation candidates assigning a lexical feature to why are absent: active set is bias alone. For a two-class single-example gold-1 update starting at zero, margin other than S+1 on that row or other than 1 on an all-UNSEEN row is excluded by averaging over the initial and updated states. At zero epochs, class 1 or positive margin is absent. These explicit parameterized families are an inner complement on valid constructed models.

reverse-reconstruction

Pred(class,margin,why) cannot recover inactive weights, omitted negative contributions, truncated explanations, original dictionary row or training order. On an all-UNSEEN row it constrains only bias differences; infinitely many finite inactive weights remain compatible. Zero-epoch output admits all dictionaries/training labels with the same nonempty class count. A reconstructed compatible weight matrix is not the actual learning history.

specialization

Unknown-projection intervention holds trained weights fixed and replaces all coordinates by UNSEEN; the observation becomes bias-only. A single supplied training row with gold 1 gives independent literal weights +/-0.5, hence literal margins S+1 and 1. Zero epochs is a separate abstention/tie boundary.

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.

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