H4 Polytopic Attention + Project Olympus
Geometric attention using the 600-cell polytope for O(log t) token lookup, with ternary quantization and multi-specialist architecture.
What's Here
Checkpoints
| Model | Size | Description |
|---|---|---|
| h4_fullscale_final.pt | 94MB | 24M ternary H4 LM, PPL 10.0 on TinyStories |
| h4_cross_encoder.pt | 98MB | 80% R@1 cross-encoder for reranking |
| olympus_code/final/ | 116MB | SmolLM3-3B LoRA, code specialist (loss 0.768) |
| olympus_math/final/ | 116MB | SmolLM3-3B LoRA, math specialist (loss 0.235) |
| olympus_qa/final/ | 116MB | SmolLM3-3B LoRA, QA specialist (loss 1.39) |
Verified Results
| Result | Value |
|---|---|
| H4 attention scan ratio | 3.1% at T=2048 (O(log t)) |
| Rust ChamberTree speedup | 10.6x at 65K keys, 98.3% recall |
| Ternary quantization gap | 0.003 bpb |
| Language generation | PPL 10.0 (beats 33M baseline) |
| Router accuracy | 100% on 50 test cases |
| Compiled arithmetic | 30/30 exact |
| Transformer-VM | 10.7K tok/s with OpenBLAS |
| Code verifier | Catches DP backtracking bugs via property checking |
Mathematical Discovery
Galois Conjugation Theorem (machine-verified in Lean 4): For every E8 lattice vector v, the H4 and H4' projected norms are Galois conjugates in Q(sqrt(5)). The E8 theta series decomposes as a Hilbert modular form of weight (4,4) with palindromic coefficients.
Formal proof: GSMLean/GaloisConjugation.lean
Architecture
Three-tier compute: transformer-vm (exact, 10.7K tok/s) > compiled arithmetic (fallback) > specialist LLM (language). Code verification via property checking (sprint contract pattern).
Quick Start
Research Plan
See docs/GEOMETRIC_INFERENCE.md for the roadmap to Opus-level AI at 50 tok/s on consumer CPU via five multiplying geometric optimizations.
License
Apache 2.0. Transformer-VM integration uses Percepta's Apache 2.0 code.