← All Projects

AdapterOS

Deterministic inference runtime + verifiable execution receipts for regulated AI deployments.

wip 2024-present Creator
Deterministic AI Cryptographic Verification Inference Research

What It Is

A deterministic inference runtime designed for high-stakes environments. AdapterOS ensures that identical inputs produce identical outputs across different execution environments. It generates cryptographic execution receipts to bind inputs, routing decisions, and outputs into a single verifiable artifact.

Why It Matters

In regulated and offline deployment, "usually works" is a failure mode. You need proof of what happened. You need systems that behave the same way under the same constraints.

Key Features

  • Deterministic Inference Paths - Eliminating numerical divergence across runs.
  • Cryptographic Execution Receipts - Verifiable proofs for every inference step.
  • LoRA Hot-swap Adapters - Auditable multi-model routing.
  • Fixed-point Quantization (Q15) - Optimizing for stability over floating-point variance.
  • Canonical JSON Serialization - Stable hashing for model state and receipts.

Technical Identity

AdapterOS implement the core concepts of verifiable execution. It utilizes BLAKE3 for high-speed content addressing and HKDF-SHA256 for secure seed derivation in non-deterministic layers.

AdapterOS runtime architecture showing application layer, runtime components, and execution receipt output APPLICATION LAYER (User prompt, model selection) ADAPTEROS RUNTIME Deterministic Inference Engine • Fixed-point (Q15) • Canonical serialization LoRA Adapter Manager • Hot-swap • Version lock Receipt Generator • BLAKE3 content addressing • HKDF-SHA256 seed derivation • Token-level audit trails EXECUTION RECEIPT (OUTPUT) Input hash • Routing log • Output hash • Timestamp + version