Adam · runtime governance for LLMs

Prove which model answered.

Adam is a model-agnostic runtime governance layer for any organization that deploys LLMs and needs to prove which model answered, that it wasn't swapped, and that it stayed within policy.

Request a demo How it works

● Running system, recorded live
Patent-pending (filed with SAIP)
Self-measured; no third-party audit yet

2:14 screen recording of the running console and an isolated test environment. English voiceover; captions burned in. The 32B model's inference wait is sped up and labelled on screen.

The problem

Logs say a model answered. They don't prove which one.

Organizations putting LLMs into production often can't prove which model actually answered a request, whether it was swapped or altered behind the endpoint, or whether it stayed within policy at runtime.

Auditors, risk teams and regulators ask for that proof. Application logs alone don't provide it.

How it works

Seal → Gate → Record → Verdict

Adam sits between your applications and your models. Every step below is shown in the demo on the running system.

Seal
Every weight file of a model is fingerprinted into a sealed identity. Adam confirms the running process is that model, sandboxed, with no outbound connections.
Gate
Each request passes access, identity, isolation and egress checks before the model is called. An answer contradicted by live evidence is withheld.
Record
Delivered, withheld and blocked exchanges are all appended to a ledger, and results export as an ed25519-signed evidence pack.
Verdict
A release monitor rolls nine indicators into one daily go / no-go. One red blocker closes the day's verdict — no average can hide it.

See it running

Screens from the running system

Captured from Adam's console on 2026-10-20 (UTC). Screens marked isolated run Adam's own gate against a throwaway model, so the live model is never touched.

Runtime gate trace for a live request: access, sealed identity, isolation, network egress, inference, claim assessment, verdict and ledger, each with timing; result delivered and recorded. — LiveA request through the runtime gate. Every stage is timed: access, sealed identity (29 weight files re-checked), isolation, egress, inference, claim assessment. The answer had no checkable claims, so it was delivered as unverified and written to the ledger.

Registered models with sealed identities, file counts, sealing dates, process isolation and outbound connection count. — LiveSealed model identity. The active model's 29 weight files match the seal; the running process is the sealed model, sandboxed, with zero outbound connections. A second model is registered but not cleared, so its gate is closed.

Two gate runs: before and after one byte of the model weights is altered; the second run fails the sealed identity check and is blocked. — IsolatedModel swap blocked. One byte of the weights is altered behind the endpoint. The next request fails the sealed-identity check and is blocked before the model is called.

Release monitor showing today's verdict and nine indicators, three of them marked as blockers. — LiveRelease monitor. One daily go / no-go across nine indicators, three of them blockers. On capture: GO — with follow-up, no red blockers, Model Honesty on watch. Self-measured; no third-party audit yet.

Evidence pack panel with SHA-256 recomputed and ed25519 signature verified, next to a table mapping EU AI Act record-keeping articles to Adam's mechanisms. — LiveSigned evidence pack. The console recomputes the SHA-256, verifies the ed25519 signature and exports the signed JSON. The pack maps Adam's mechanisms to EU AI Act record-keeping articles — a mapping, not a conformity assessment.

Deployment

Runs in your environment, next to your inference stack

How it deploys

Adam runs inside your own environment, beside the servers that host your models.
It governs open-weight models served behind OpenAI-compatible endpoints.
A model is added by sealing its weight files and registering its endpoint; until it is cleared, its gate stays closed.

Wired today

Serving: MLX on Apple silicon, behind an OpenAI-compatible endpoint, on-premises.
AceGPT-v2-32B (4-bit, with a fine-tuning adapter) — sealed, cleared for service.
ALLaM-7B — sealed and registered, not cleared for service.

That is the full list of wired providers and models today.

Who it's for

Teams that must show their work on AI

Regulated enterprises

Finance, healthcare and telecom teams that must show auditors which model answered and that it stayed within policy.

Public-sector IT

IT departments that need a signed, append-only record of every model exchange, including the ones that were blocked.

Teams shipping LLM features

Product teams whose LLM features have to pass an audit, with evidence packs mapped to record-keeping requirements.

Status & company

Working software, honestly labelled

Status: Running system. Every screen on this page and in the video comes from the running software.
Intellectual property: Patent-pending (filed with SAIP). SA 1020268315 — runtime model governance with sealed model identity. SA 1020268588 — model provenance and lineage.
Measurement: Self-measured with Adam's own tooling; no third-party audit yet.
Product sheet: One page, PDF
Company: MrF X Nexus
Address: Riyadh, Saudi Arabia

Request a demo

See Adam on your own models

Tell us what you deploy and what you need to prove. We will reply by email.