Skip to content

README.md

Latest commit

 

History

History
85 lines (66 loc) · 3.83 KB

README.md

File metadata and controls

85 lines (66 loc) · 3.83 KB
Raw

gen (源)

gen — 源, Japanese for source / origin. The package import name is gen; the PyPI distribution is konjo-gen (the name gen was already taken on PyPI).

A reusable scaffold for screening sequence-mixing primitives against the Konjo Architecture Charter (docs/CHARTER.md). It is the cheap, local gauntlet that every future "screen a new mixer" sprint drops a candidate into — so each candidate is a drop-in, not a rebuild.

This repo implements no novel primitive. It provides the foundation:

  • the SequenceMixer plug interface, which encodes the charter's four-gate filter as first-class structure (src/gen/primitives/base.py);
  • two reference baselines — plain causal softmax attention (the known-capable control) and a minimal selective linear recurrence (the known recall-limited control);
  • the synthetic gauntlet — induction, MQAR, selective-copy, and formal languages (Dyck / parity / modular addition) — that discriminates a capable mixer from an incapable one in minutes on CPU/MPS;
  • the 30-run paired Wilcoxon house gate (src/gen/stats/wilcoxon.py);
  • the append-only four-gate decision ledger plus a paper-screen scaffolder.

Why this exists (charter in one paragraph)

Anything nameable has been tried; what remains untried usually fails a hard constraint. So re-derive a sequence model from the ten invariant constraints (charter §1) and let novelty be the residue of rigor. Before any code, a candidate is paper-screened on four gates — Expressivity, Trainability, Hardware, Scaling — and most die at Gate 3 (hardware). A candidate that passes on paper earns a kill test: MQAR against these baselines. A primitive that cannot pass MQAR will not model language. Negative results are the deliverable.

Quickstart

# install (CPU torch)
pip install "torch==2.5.1" --index-url https://download.pytorch.org/whl/cpu
pip install -e ".[dev]"

# checks
python scripts/verify_deps.py        # dependency provenance gate
pytest -q                            # interface / gauntlet / wilcoxon / ledger

# run the gauntlet on a baseline (prints a capability profile + verdict)
python scripts/run_gauntlet.py attention
python scripts/run_gauntlet.py ssm

# reproduce the V1 baseline report + seed the decision ledger
python scripts/run_gauntlet.py --baselines       # writes docs/gauntlet_baselines.md

# scaffold a four-gate paper screen for a future candidate
python scripts/paper_screen.py tropical_max_plus  # -> docs/screens/tropical_max_plus.md

The reference baselines establish the separation the harness must reproduce: attention aces MQAR across K; the recurrence shows the known recall gap (degrades as the key count K grows) while still passing induction and selective-copy. See docs/gauntlet_baselines.md.

Adding a candidate mixer (future sprints)

  1. Implement SequenceMixer in src/gen/primitives/<name>.py: forward (causal, shape-asserted), optionally step, flops, and a filled-in gate_card (your paper-screen verdict on each of the four gates).
  2. Register it in src/gen/primitives/__init__.py:REGISTRY.
  3. python scripts/run_gauntlet.py <name> and read the MQAR curve against the baselines. Record the outcome in the ledger.

Layout

src/gen/primitives/   SequenceMixer interface + attention/ssm baselines
src/gen/gauntlet/     harness, report, and the four task families
src/gen/stats/        30-run paired Wilcoxon gate
src/gen/ledger/       append-only JSONL decision log
scripts/              run_gauntlet, paper_screen, verify_deps
docs/                 CHARTER, four-gate template, decision log, baseline report

Scope

Foundation only. No novel primitives, no training framework, no language data, no fast kernels. The first real candidate (e.g. a tropical max-plus mixer) is a separate sprint — see NEXT_SESSION_PROMPT.md.