adapterOS
Air-gapped inference with verifiable execution receipts.
An offline-first runtime in development, designed for regulated environments where cloud access is not an option. The goal is signed runs, versioned adapters, and artifacts an auditor can verify. It is tuned for Apple Silicon and other UMA-style hardware profiles. The evidence layer is what makes AI usable in compliance.
In development. Early access is limited and by invitation.
MLNavigator develops and commercializes adapterOS for regulated deployments.
What It Is
Local-First Runtime
adapterOS is designed to run on your hardware. Models, adapters, and configurations load locally. Inference is intended to run offline in controlled environments.
Adapter Management
Adapter management focuses on loading, switching, and versioning adapters (LoRA, QLoRA) with provenance tracking. Hashing and signing are part of the design.
Proof Generation
Each run is designed to produce a proof pack: hashed inputs, model state, configuration, and outputs. Signed with Ed25519 and chain-linked to previous runs.
Deterministic Execution
Determinism is handled as policy: fixed seeds, pinned versions, and declared tolerances for repeatability within documented numerical bounds.
How It Runs Offline
No Network Calls
The design target is zero outbound connections during inference. No license checks, no telemetry, no updates.
Bundled Dependencies
Dependencies are intended to be vendored. No package manager calls. No CDN fetches.
Air-Gap Compatible
Designed for disconnected networks. Transfer via secure media. Verify on isolated systems.
Hardware Profile: Unified Memory
adapterOS is designed to exploit unified memory systems where CPU, GPU, and NPU can share a coherent memory space.
Offline operation is the primary design target. Network connectivity is treated as an avoidable risk in high-assurance environments.
Planned Artifacts
Planned outputs include structured artifacts for audit, configuration control, and verification.
proof.jsonHash chain linking input, model, adapter, config, and output. Ed25519 signature.
manifest.jsonDeclaration of what was supposed to run. Versions, hashes, configuration.
audit.logTimestamped event log. Hash-chained entries. Deletion-detectable.
energy.jsonOptional Joules/token measurements when energy profiling is enabled.
What It Does Not Do
No Data Exfiltration
adapterOS is designed without endpoints for outbound data transfer.
No Tracking
No usage analytics or behavioral telemetry in the core runtime.
No Auto-Updates
Updates are intended to be explicit and verified. No background downloads.
No User Accounts
No cloud accounts required. No license servers in the core runtime.
Interested in adapterOS?
Pilot deployments are scoped to regulated environments with defined compliance requirements. Contact us to discuss constraints and audit needs.