Elamur Groupantsmath-solution-memory-perceptron
The groupant beside math/src/memory.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 |
| descriptant | present | descriptant |
| invariant | partial | invariant |
| tests | partial bounded evidence | 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 |
failure_precedence | failure_precedence | partial source derived model | failure_precedence |
representation_transport | representation_transport | partial source derived model | representation_transport |
resource_boundary | resource_boundary | partial source derived model | resource_boundary |
domain_boundary | domain_boundary | partial source derived model | domain_boundary |
Model text
descriptant
Source-bound description
Sort classes and memo ids; update wrong predicted class per feature, accumulate timestamped weights then average by final t. Guess sorts descending scores then lexical class, margin top1-second or0, retains five highest signed feature contributions.
Epoch 2026-10-31: prospective model before bounded synthetic tests; production behavior unchanged. Source scope bytes[5291,8283). No external memory load/generalization. load drops malformed lines and read errors; learned accuracy unmeasured. Not every branch in this fragment executed.
approximant
Forward relation
Sort classes and memo ids; update wrong predicted class per feature, accumulate timestamped weights then average by final t. Guess sorts descending scores then lexical class, margin top1-second or0, retains five highest signed feature contributions.
Domain: No-panic comparison domain requires comparable non-NaN accumulated scores and feature weights. Finite weights alone do not prevent accumulation overflow or NaN; partial_cmp unwrap panics on NaN. Finite meaningful margins additionally require finite score differences. Duplicate features can update twice.
Executable Rust approximant unsupported: categorical/effect observations have no useful percentage metric yet. Future caps1/2/5/10% only; no automatic substitution.
invariant
Invariant and cost
No classes excludes Guess; zero epochs can still leave classes and lexical tied prediction. why has<=5 entries, not necessarily positive weights.
Ordering: Class/id sorting before epochs; accumulated old weight added before each update; final residual time before threshold1e-9.
Cost: O(epochsmemosclasses*features) scoring, plus key clone/hash and sort; weights O(class-feature pairs). usize ticks unchecked.
composition
Composition and epoch memory
features from Problem feed both train and guess; solve compares margin threshold but it is not calibrated probability.
Source scope bytes[5291,8283). No external memory load/generalization. load drops malformed lines and read errors; learned accuracy unmeasured. Not every branch in this fragment executed.
Partial syntax and finite tests do not establish all caller assumptions.
models
behavior_complement
No classes excludes Guess; zero epochs can still leave classes and lexical tied prediction. why has<=5 entries, not necessarily positive weights.
reverse_reconstruction
Winning class/margin admits many weights and feature vectors; averaged weights lose update history. Empty model may come from empty memos or zero updates.
failure_precedence
Class/id sorting before epochs; accumulated old weight added before each update; final residual time before threshold1e-9.
representation_transport
features from Problem feed both train and guess; solve compares margin threshold but it is not calibrated probability.
resource_boundary
O(epochsmemosclasses*features) scoring, plus key clone/hash and sort; weights O(class-feature pairs). usize ticks unchecked.
domain_boundary
No-panic comparison domain requires comparable non-NaN accumulated scores and feature weights. Finite weights alone do not prevent accumulation overflow or NaN; partial_cmp unwrap panics on NaN. Finite meaningful margins additionally require finite score differences. Duplicate features can update twice.
tests
Evidence, 2026-10-31
Selected actual-source checks: cargo test -p math --lib memory::. Exact commands, numeric statuses, normalized outcomes, compiled assertion controls and restoration receipts are in the external wave28 mathcontrol run.
Source scope bytes[5291,8283). No external memory load/generalization. load drops malformed lines and read errors; learned accuracy unmeasured. Not every branch in this fragment executed.
No live teacher/API, trained model, corpus/held-out, GPU, DB or expensive benchmark run. Parent syntax generated twice identically. Scope remains partial.
This page in Markdown → https://md.groupants.elamur.ai/free/en/mova/math/src/memory.rs.groupant/perceptron.groupant/