Skip to content

CHANGELOG.md

Latest commit

 

History

History
33 lines (29 loc) · 1.83 KB

CHANGELOG.md

File metadata and controls

33 lines (29 loc) · 1.83 KB
Raw

Changelog

All notable changes to this project are documented here. Format loosely follows Keep a Changelog; this project uses semver.

[0.1.0] — 2026-06-23

Initial scaffold sprint: the reusable foundation for screening sequence-mixing primitives against the Konjo Architecture Charter. No novel primitives.

Added

  • docs/CHARTER.md — the Konjo Architecture Charter (source of truth).
  • SequenceMixer plug interface (gen.primitives.base) encoding the four-gate filter (Expressivity / Trainability / Hardware / Scaling) as a GateCard, with forward (causal), optional step, required flops, and boundary shape assertions.
  • Reference baselines: AttentionRef (causal multi-head softmax attention, the known-capable control) and SSMRef (gated linear-attention recurrence, the known recall-limited control with a true step path).
  • Tiny harness model embed -> [mixer block × L] -> norm -> head.
  • Synthetic gauntlet: induction, MQAR (accuracy-vs-K curve), selective-copy, and formal languages (Dyck / parity / modular addition), with a deterministic train+eval harness and a report/verdict renderer.
  • gen.stats.wilcoxon — self-contained 30-run paired Wilcoxon signed-rank gate (exact for small n, normal approximation with tie + continuity correction).
  • gen.ledger — append-only JSONL four-gate decision log with a generated DECISION_LOG.md view.
  • Scripts: run_gauntlet.py, paper_screen.py, verify_deps.py.
  • Blocking CI (lint, format, type check, tests, fast gauntlet sanity) and pre-commit hooks (ruff, verify_deps). No continue-on-error.
  • Tests for the interface, gauntlet scorers, the Wilcoxon gate, and the ledger.
  • docs/gauntlet_baselines.md — V1 capability profiles for both baselines with the MQAR accuracy-vs-K curves and the environment/version header.