A theory-first, low-compute research program to rearchitect reasoning systems "from the ground up": separating computation from addressable memory, moving beyond pure next-token prediction, and targeting the real open problem the 2026 literature has converged on — the consistent ~2-3x compositional-generalization collapse from ARC-AGI-1 to ARC-AGI-2 that every paradigm exhibits, plus the fact that reasoning is still "knowledge-bound" and expensive.
The full staged program lives in docs/plan.md. The concrete build-and-test plan for the synthesized architecture (codename VNR — a fixed interpreter + growing abstraction library) lives in docs/vnr-build-test-plan.md.
- Context limits memory -> separate computation from addressable memory (von Neumann analogy). SSMs (Mamba-3), power-law memory (Sessa), external/episodic memory.
- "Just next-word prediction" -> predict in latent space and optimize for correctness/compression. JEPA / energy-based models; MDL (CompressARC); program synthesis ("Parrots to Von Neumanns").
- ARC is unsolved -> the gap is compositional, sample-efficient, fluid reasoning, not scale. The strongest ideas live in tiny zero/low-pretraining models (TRM ~7M params, CompressARC).
- Cost is the threat -> treat skill-acquisition efficiency (Chollet) and $/task as first-class targets.
| Directory | Purpose |
|---|---|
docs/ |
The approved research plan and high-level documents. |
lit/ |
Annotated bibliography and per-paper notes (one file per paper). |
ledger/ |
Living research ledgers: limitations.md, hypotheses.md, results.md. |
spec/ |
Architecture spec / manifesto drafts (architecture-v0.md). |
experiments/ |
Tiny, laptop-runnable proof-of-concept experiments + lab-notebook entries. |
data/ |
ARC-AGI task data (fetched via data/fetch_arc.py; not committed). |
references/ |
Vendored reference code (TRM, CompressARC, NVARC; cloned, not committed). |
This is a lab notebook, not a product. Every experiment gets a dated entry in experiments/notebook/ using experiments/notebook/TEMPLATE.md: hypothesis -> setup -> metric -> result -> decision. Hypotheses are pre-registered in ledger/hypotheses.md before experiments run; outcomes land in ledger/results.md.
Phase 0 (workspace scaffolding) complete. Next: Phase 1 — literature map + limitations ledger.
- Python 3.11+
- See
requirements.txtfor the (intentionally minimal) starting dependencies.