Skip to content

64_spike-go-native-linear-classifier.md

Latest commit

 

History

History
58 lines (48 loc) · 2.39 KB

64_spike-go-native-linear-classifier.md

File metadata and controls

58 lines (48 loc) · 2.39 KB
Raw
id 64
title Spike Pure-Go Embedded Weasel Classifier
status ✅

Spike Pure-Go Embedded Weasel Classifier

Goal

Evaluate a fully embedded, pure-Go classifier path for weasel-language (or verbose-actionable) detection with no runtime dynamic libraries.

Tasks

  1. Define a pure-Go model family to evaluate first (for example sparse linear classifier over cue and n-gram features).
  2. Build a minimal prototype inference package that runs with stdlib-only runtime dependencies and deterministic scoring.
  3. Define a weight packaging path that is fully embedded in the mdsmith binary (for example go:embed plus checksum verification).
  4. Measure CPU latency and memory on the same benchmark corpus used in previous weasel spikes.
  5. Measure binary-size impact versus current mdsmith and compare with the yzma spike artifact footprint.
  6. Define integration boundaries and fallback behavior for MDS029: backend mode switch, timeout policy, and diagnostic stability.
  7. Document maintenance workflow: training export format, versioning, and safe model update procedure.

Results

See eval/conciseness/spikes/go-native-linear-classifier/README.md.

Highlights from the spike:

  • Prototype is fully pure-Go with stdlib-only runtime dependencies.
  • Weights are embedded with go:embed and verified by pinned SHA256.
  • Spike report includes classifier implementation details and raw captured benchmark output in markdown.
  • Cue lexicons were moved into the embedded artifact with validation gates to keep list expansion maintainable and correctness-checked.
  • Determinism was confirmed across in-process and process-restart runs (unique_hashes=1).
  • Latency and memory metrics were captured on the same six-sample corpus used in the yzma spike.
  • mdsmith binary delta for embedded artifact was measured at +480 bytes, versus yzma's +0.5 MB binary delta plus external model/library artifacts.
  • Recommendation: adopt this path as the CPU fallback candidate for plan 58, pending full dataset quality validation.

Acceptance Criteria

  • Prototype runs with no YZMA_LIB or external dynamic libraries.
  • Embedded weights load from binary-only assets.
  • Deterministic outputs are confirmed across repeat runs.
  • CPU latency and memory metrics are captured.
  • Binary-size delta is measured and documented.
  • Recommendation is made: adopt, defer, or reject this path.