Assay

Layer 2 — Verification Substrate

A new layer in the AI stack.

Models generate. Assay verifies. Deterministic checks on probabilistic output — a verification substrate that sits below models in the trust hierarchy.

L4
Applications
L3
Models
L2
Verification Substrate — Assay
L1
Compute

Deterministic > Probabilistic — verification must sit below models

Generate
Extract Claims
Verify Each
Remediate
↑ loop until verified

Two pipelines

Forward Pipeline

Submit code. Assay extracts every implicit claim, verifies each one independently, and returns a detailed report with remediation.

Code in → Claims extracted → Each verified → Report out

Reverse Pipeline

Describe what you need. Assay generates code, then automatically verifies it before returning — only shipping code that passes.

Task in → Code generated → Auto-verified → Verified code out

Deployment options

GitHub Action

Drops into any CI/CD pipeline. Runs on every PR. Free for open source.

REST API

Two endpoints. Forward to verify, reverse to generate. Bearer token auth.

LVR Service

Full codebase audit and reconstruction. Three verification loops. Enterprise.

Why this layer is permanent.

We trained models on verification signal (RLVF). More data made them worse. Verification can't be baked into models — it requires a separate, deterministic layer.

  • Hybrid formal verifier: deterministic checks at zero marginal cost. Pure TypeScript, no LLM, can't hallucinate.
  • Domain-agnostic: code today, legal/medical/financial tomorrow. Same verification loop, different claim types.
  • Provisional patent #63/980,048 (filed February 2026)
  • Published research: DOI 10.5281/zenodo.18522644