- Entry date
- 15 April 2026
- Category
- Models
- Lead over the world
- ~2 years ahead
- Access
- 🔒 Protected
ResearchPartly proven
Everyone builds on attention. We built a sequence model out of wave equations — and made it prove it was actually running before we believed a single loss number.
- 1.114B
- params in the GPU smoke run
- 10.41→10.19
- loss descending over 20 measured steps
- RETRACTED
- the '1B WaveCore at loss EMA 2.69' entry — label aspirational, checkpoint gone
- 292.6M
- parameters in the largest artifact we actually trained
- 1,024+
- independent wave channels per layer
Honest evaluation
Partly proven
The architecture demonstrably learns (a real, measured 20-step descent on a 1.1B smoke run) — but a smoke is not a trained model, and the flagship speed claims are stated targets, not measurements.
Runnable proof — see it work
What we learned — a preview
The honest edges are open even when the engine is not. The full report — what it was, what we built, and the measured internals — is protected.
- The smoke run is a beginning, not a trained model: phases 2–9 of the architecture are gated and queued, not done.
- The aspirational step-time targets (single-digit ms vs a ~500ms baseline) are targets — the smoke measured descent, not speed.
- The '1B WaveCore trained to loss EMA 2.69' claim is RETRACTED, by us, on our own evidence. The '1B' label was aspirational and was never grounded in module arithmetic — our design document's own words. The largest artifact we actually trained is 292.6M parameters. The cited 4.5 GB checkpoint is absent: the directory exists and is empty. Even if it were found, the run was 5,000 steps on a synthetic corpus and the loss is unverifiable past step 2000, with no final gate. This is not a reproducibility bundle that is 'still owed' — it is a result that cannot be produced.
- A sibling label in the same family was separately corrected in our own records from '1B' to 15.7M actual parameters, with 9 of 10 checkpoint files untrained. Treat any parameter count from this line as provisional unless it names the checkpoint.
- This is architecture-IP: the internals are protected, so this report gives the thesis and the measured smokes, not the kernels.
Protected report
The deep body of “A model made of waves, not attention — the non-transformer architecture” is behind access.
We open the demos, the specs, and the method; we protect the engines, the model internals, and anything that touches commercial, safety-critical work. This report describes an engine — so its details are gated, and no client, company, or project is named. Members can read it in full; if you have a genuine reason to see it, tell us who you are and why.
Proofs & sparks
We demonstrate rather than assert. Each ✅ proof is a visible result with a hard figure.
- Wave-equation propagator + closed-form backwardkernel verified · smoke loss 10.41 → 10.19the wave-equation forward propagator and closed-form backward gradients run, gated by an upfront GPU smoke test — a measured smoke-run loss curve (one sub-claim, hysteresis-bounded path selection, still unconfirmed).
- Hierarchical Huffman-softmax LM head2.8× faster LM heada band-aligned Huffman tree on the wave-equation LM head — implemented and validated with a measured 2.8× gain.
- Speculative-loss-prediction lookahead headverified across a 50k-step runthe speculative-loss-prediction lookahead head is present and exercised end-to-end across a real 50,000-step training run — source-grep + run log.
- Confidence-gated layer-skip + dynamic reservoirpair, source-verifiedconfidence-gated layer-skip and a dynamic-frozen reservoir are both present and exercised on the wave recurrent stack — source-verified.
- Reversible Verlet backward + sparse AdamW4 mechanisms, source-verifiedreversible-Verlet backward, a fused propagator and sparse-AdamW paths are present in the wave backward/optimiser — source-verified, partial.
- Layer streaming + NVMe optimizer spillsource-verified, partialsliding-window layer streaming and NVMe optimizer-spill let a huge model train under tight VRAM — paths present, source-verified, partial.
- Activation-fingerprint fact attributiondesigned — not yet measuredan activation-fingerprint index for real-time fact attribution during training — fully specified, not yet implemented (an honest design claim).
- An honest correction — archivedarchived — math inverted vs the real kernela closed-form cell Jacobian / TBPTT-1 anchor / fused Triton cell — design-only claims later found to have an inverted derivation versus the real forward kernel; kept on the record as a correction.
- Ternary GEMM, bit-exact92.8% of cuBLAS, bit-exactthe ternary WMMA GEMM reaches 92.8% of cuBLAS and is bit-exact; int4 fused decode 2.9× at M=1, HBM traffic −8.7×.
- Flat at 128k context2× llama.cpp at 128kat 128k context the wave / SSM decode stays flat and runs 2× llama.cpp — where a transformer's cost climbs; sparse-attention 59.7× at 65k.
- A real LLM on our own stacktoken-for-token == llama.cpp · 151.7 tok/sa real Qwen2.5-0.5B generates end-to-end from QUANTA-emitted PTX on real GGUF weights — token-for-token identical to llama.cpp at 151.7 tok/s, zero cuBLAS/cudart.
- Does-it-stick: GOgo / no-go = GOassociative memory validated on real Qwen-0.5B embeddings — holds at cos 0.87 (≈ random 0.90), drift-bounded, no catastrophic forgetting; go/no-go = GO (needs dreaming; recall approximate).