Skip to content
Regulayer™Human Control for AI
Book a live demo

Technical overview · for engineers and licensees

An enforcement point at the moment of effect.

Regulayer™ runs outside the model, between an AI agent and the systems it acts on. Every proposed action is checked against the current authority of a named person before it takes effect, and every decision is written to a signed record that anyone can check offline.

Runs inside your environment. Your data does not leave it. No cloud service of ours.

At a glance

The engine, from the outside.

Placement
In the path, before a consequential action takes effect: a tool call, an API call, a write, a payment, a batch release. Outside the model and outside the agent.
The caller supplies
The proposed action and its target, the agent that proposes it, and the named person whose authority applies.
What comes back
Allow, send back to correct, hold, or stop. Every answer carries a reference to its record.
Failure mode
Fail closed. If the check cannot complete, the action does not run, and the record says stopped or unknown, never a success it cannot show.
Decision time
Decided inside your environment, before the action takes effect. No round trip to us.
Deployment
Inside your environment: on premises, on your own servers, or a sandbox for a pilot. Your data does not leave it. No cloud service of ours is required.
Models
Model-agnostic. Change the model and the check stays where it is.
Record
One entry per decision, allowed or refused, signed. Tamper-evident: any change after sealing fails verification.
Verification
In the browser or offline, with the open verifier. No account, no call to Regulayer™.

Integration

How it connects.

What it is: a deterministic enforcement point that runs in your environment, separate from any AI model. The same inputs give the same decision. It judges the action against the authority, so it works the same with any model and needs no retraining or fine-tuning.

IntegrationHow it worksBest for
Agent tool wrapperThe agent’s tool calls pass through the check before they execute. The tool runs only on allow.Teams building their own agents.
Fail-closed gatewayConsequential actions route through a gateway where the decision and the action happen in one step, so there is no gap between allowed and done.Validated and high-consequence workflows.
NVIDIA OpenShellTrusted middleware in the request path, outside the agent’s sandbox, the slot NVIDIA describes for identity and application-specific checks. How it fits the NVIDIA stackAgents built on NVIDIA’s stack.
MCPStatus, record export and verification over the Model Context Protocol.Agent platforms and tooling.
Embedded (OEM)The same control inside a vendor’s own product, under licence.Instrument, automation and software platforms.

Interoperability, enabled in a pilot on request. Records export as OCSF 1.9. An identity provider can withdraw a person’s authority through SSF and CAEP security events: withdraw only, never grant. A read-only MCP server gives status, record export and verification.

Architecture

In your architecture.

The model is treated as an untrusted candidate generator. The decision is made outside it, deterministically, at the boundary, before the action takes effect.

ConcernHow it is handled
Zero trustA policy decision point and enforcement point in one, in the terms of NIST SP 800-207. The AI proposes. The action runs only on allow.
Cloud placementThe same position in the request path as an API Gateway Lambda authorizer, a human-approval step in Step Functions or an Amazon Verified Permissions check: before the call takes effect. What it checks is a named person’s current authority over this action.
LatencyIt runs inside your environment, so there is no round trip to us. Authority is checked at the moment of effect: in the scale run, 7,457 in-flight actions were re-decided when authority changed.
Partitions and outagesIf the check cannot complete, the action holds and the record says stopped or unknown. A stop stays latched until a named person re-arms it, and the re-arm is recorded as its own decision. Availability and wrong refusals are counted, as the Cleanroom Standard requires (6.11). Plan the on-call path for re-arming as part of the deployment.
Records and data residencyEach record carries the decision, not your data. Payloads stay inside your boundary, behind your own network and access controls. Change one character of a record and verification fails.
Accelerated infrastructureOn HGX and DGX systems, where GPUs share NVLink and PCIe fabric, Class 3 isolation means the check runs on separate compute: its own node or a discrete controller. It shares no memory, no process and no storage with the AI it checks.
Edge and physical AIOn Jetson-class devices, robots and instruments, it holds or stops the next consequential action at the software boundary. It does not replace the machine’s safety interlocks or its real-time safety controller. It sits above them (Cleanroom Standard 1.6).
Drift and accelerator loadDrift is measured in the path, at the boundary, outside the model. The check does not need the accelerators serving the model.

Checking a record

From the checker’s side.

Drop the record into the open verifier. It checks in your browser, with no call to us.

Change one character and it fails.

Measured

Measured at scale.

69,199Agent actions evaluated in one run, by 320 agents against 80 protected resources.Scale test
7,457Actions re-decided mid-transaction when authority changed while they were in flight.Same run, 578 authority changes
428,521Records verified, all intact, in 7.2 seconds by separate code.Internal test, separate verification code

Standards

What it supplies evidence toward.

StandardWhat it asksWhat Regulayer™ supplies
Court & litigation
FRE 901(b)(9)Evidence describing a process or system and showing that it produces an accurate result.A record of each decision, made by the process that decided it, that anyone can check.
FRE 902(13)Certified records generated by an electronic process or system.Designed to support self-authentication, with the certification of a qualified person.
FRE 902(14)Certified data copied from an electronic device, storage medium or file.Designed to support self-authentication, with the certification of a qualified person.
FRE 903An attesting witness is needed only where the governing law requires one.A signed record of each decision, with no attesting witness involved.
Pharma & medical
21 CFR Part 11 §11.10(e)Secure, computer-generated, time-stamped audit trails of operator entries and actions.A time-stamped, signed record of each AI action, made as it happens.
21 CFR Part 11 §11.10(g)Authority checks, so only authorized individuals can use the system, sign, alter a record or perform the operation.Each action checked against a named person’s current authority before it runs.
21 CFR 211.22(c)The quality unit approves or rejects procedures and specifications.That a named, authorized person in the quality unit made the decision, and when.
21 CFR 211.68(b)Controls over computer systems, with changes made only by authorized personnel.Automated actions that run only under a named person’s authority, each one recorded.
21 CFR 211.188(b)(11)The batch record identifies the person performing or checking each significant step, including a step performed by automated equipment.The named person whose authority covered each automated step.
EU GMP Annex 11Audit trails, access control and data integrity for GMP computerised systems.System-generated records of GMP-relevant AI actions, with the person responsible.
EU GMP Annex 22 (draft)AI in GMP, under human review and oversight.The AI output and the human decision on it, recorded together.
FDA and EMA Good AI Practice, Principle 6Documentation of data and decisions that is detailed, traceable and verifiable.A traceable record of each AI decision that an outsider can check.
ALCOA+Data that is attributable, legible, contemporaneous, original and accurate, and complete, consistent, enduring and available.Attributable is the first letter: each action carries the named person whose authority it ran under, recorded as it happens.
EU AI Act
EU AI Act, Article 12Record-keeping: automatic recording of events over the lifetime of the system.An automatic record of every decision.
EU AI Act, Article 14Human oversight, including the ability to intervene or stop.Designed to support it: a named person’s authority, checked before each action, which that person can withdraw.
Insurance & financial services
NAIC AI Model BulletinSections 3 and 4A written program for AI systems used by insurers: risk controls, oversight of third-party systems, documentation.A signed record that the controls operated, action by action, with the named person whose authority applied.
NYDFS 23 NYCRR Part 500§500.6Audit trails, and records of what authorized users did.A signed record of each AI action and the authorized person it ran under.
EU DORARegulation 2022/2554ICT risk management, and the detection and reporting of ICT-related incidents.A tamper-evident record of each AI action, ready for incident reconstruction.
Standards
ISO/IEC 42001An AI management system, with operation and monitoring records.Operation-time evidence: who authorized each action, and what happened.
ISO/IEC 23894AI risk management, with the process recorded.Per-action evidence of the control applied, and its outcome.
NIST AI RMFGovern, Map, Measure and Manage AI risk.Evidence for each action: the authority it ran under, and the outcome.
EN 18286A quality management system for AI Act purposes. Cited in the Official Journal, 14 October 2026.Records the quality system can draw on.

Regulayer™ supplies the evidence. Your auditor, regulator or court makes the determination.

The laws and standards

Research

Sources.

  1. Li et al., Harvard, COLM 2024. Measuring and Controlling Instruction (In)Stability in Language Model Dialogs. Instructions drifted within eight rounds, even with two chatbots talking to each other.
  2. Qin et al., 2024. SysBench: Can Large Language Models Follow System Messages? In multi-turn sessions, the share of GPT-4o sessions that followed every rule fell from 84.8% after turn 1 to 33.7% after turn 5.
  3. OWASP, 2025. Top 10 for LLM Applications 2025, LLM01: Prompt Injection.
  4. Related: Panavas et al., July 2026. HANDBOOK.md: A Benchmark for Long-Context Agentic Instruction Following. Agents working to long policy documents: under strict grading, the best model passed 36.2% of trials, and most frontier models stayed below 25%.

Measurement and availability

Evaluate the release you will deploy.

The speed and scale figures on this page are our own test runs. They are not a service promise or an outside audit. Ask us for the release, the hardware, the workload and the verification files behind each one.

It connects as a tool wrapper or a gateway, around the actions you name. Which environments, and what counts as passing, is written into the scope.

Licensing

Embed it in your own stack.

Enterprises, AI platforms, OEMs, agentic-system providers and regulated operators license the engine and run it on their own infrastructure. SDKontrol carries the same control inside a vendor’s own product.

U.S. patent claims allowed by the USPTO. Patent pending issuance. U.S. Patent Application Publication No. US 2026/0342586 A1.