VRAM-resident quant pipeline (honest CPU fallback today; GPU path ROADMAP)
Three orthogonal risk signals per bar — PCA-Risk · TDA-Fracture · HJB-Kelly — each a SIGNED receipt, honestly labeled SAMPLE_SIGNAL · NOT_LIVE · NO_BACKTEST_VALIDATED. Not a trading instruction.
backend=CPU pure-Python reference · gpu_reachable=False · scenario=calm
CPU pure-Python referenceLedoit-Wolf shrinkage intensity ρ and the MP noise edge λ⁺ are computed exactly per the cited formulas on a SAMPLE synthetic universe. cuML/cuPy GPU acceleration is ROADMAP (CPU fallback ran here). NO backtest.
β0 (components) is exact via union-find; β1 is the 1-skeleton cycle rank (E−V+C), an honest fast proxy for the genuine Vietoris-Rips H1 that the GPU giotto-tda/Ripser++ path (ROADMAP) computes exactly. Synthetic SAMPLE windows; z-score baseline is illustrative. NOT calibrated on real data.
Weights auto-compress when fracture/anomaly fire (derisk_ratio<1) — the elegant property of the TDA-Kelly channel. BUT γ and κ are free, UNcalibrated hyperparameters: this is MODELED architecture, NOT a backtested strategy. NEVER a live-trading instruction.
szl-gpu-quant-v0.197973642d2e4884f…DSSE envelope over the canonical receipt — REAL ECDSA when the cosign key is in the runtime, else an explicit UNSIGNED honesty marker (never a fabricated signature).
vLLM --tensor-parallel-size 2 shards ONE larger model across a-11-oy.com GPU + RTX 4000
e.g. Qwen3-32B comfortably, or a quantized Nemotron-3-Super across combined VRAM
main GPU = primary agent model; RTX 4000 = dedicated governance/draft GPU: Auto-Review CLASSIFIER + speculative-decode DRAFT (Qwen2.5-Coder-1.5B) + embeddings
keeps the main GPU from stalling on inline review/draft — best fit for our agent+Auto-Review arch
Route via build.nvidia.com NIM through our LiteLLM/RouteLLM gateway
Ultra (550B-A55B) needs ~768GB VRAM (4×GB200-class) — CANNOT run on 2 local GPUs. NEVER claim local Ultra. Verify NVIDIA claims on OUR τ-bench+J/token harness.
| Claim | NVIDIA datasheet | SZL-MEASURED (signed) |
|---|---|---|
| Nemotron speedup vs prior frontier | up to 5× | — |
| Reasoning/accuracy uplift | +30% | — |
| Benchmark accuracy | 91% | — |
| Long-context retrieval | 1M-token retrieval | — |
| cuML PCA speedup (quant Layer 1) | 10–50× (S&P 500 scale); ~100× genomic | — |
| Ripser++ persistence (quant Layer 2) | up to 30× vs CPU Ripser | — |
The 'NVIDIA datasheet' column is the vendor's published claim (cited, not endorsed). The 'SZL-MEASURED' column stays null/ROADMAP until we run OUR τ-bench + J/token + J/bar harness and SIGN the result. measured > datasheet, always. Never fabricated.