Helix — the verifiable execution layer. One page.
The claim: real, unchanged GPT-2 (124M → 1.5B) and a 2024 Llama-architecture model run on a software stack rebuildable from 299 hand-typed bytes, with outputs matching an independent oracle token-for-token (25/25) — gated fail-closed, reproducible, attested.
299 bytes → a compiler → 8 GPU kernels → verified AI
Every model, gated
| model | argmax | max score diff | tokens |
|---|---|---|---|
| GPT-2 124M · 12 L | id 262 exact | 2.59e-04 | 25/25 |
| GPT-2-Large 774M · 36 L | id 262 exact | 3.8e-05 | 25/25 |
| GPT-2-XL 1.5B · 48 L | id 262 exact | 4.4e-05 | 25/25 |
| SmolLM2-135M · 30 L · Llama arch | id 260 exact | 4.9e-05 / 49,152 | 25/25 |
Audit instead of trust
Bring your weights: the execution layer beneath your model becomes fully traceable — same 8 kernels from 124M to 1.5B, zero new ops at scale.
One command (scripts/reproduce_trust.sh, ~1 min, CPU-only) rebuilds the chain from raw and asserts the anchors: 9837db12 · 0992dddd · 84363adb.
fp32 · to PTX not SASS · single sm_86 GPU · ≈10 s/token live by design · base models, not assistants · oracle shares the spec.
Contact: ajdemarco10@gmail.com · linkedin.com/in/anthony-demarco10 · github.com/Questeria/helix — Web: 299bytes.com