Automation you can
hold accountable.

Adekka develops foundational infrastructure for consequential decisions and operations. Its flagship platform, NUBAS, is being developed for trusted autonomous systems. It sits between automated workloads and consequential operations, authorizing actions before execution and creating evidence designed for independent verification afterward.

The problem

Authorization systems answer "may this occur?" What happens next is usually recorded somewhere else.

Automated systems now move money, transform sensitive data, communicate externally, and actuate physical infrastructure. Existing controls may authorize the action and log the execution, but those records are often produced by separate systems and are not bound into one independently verifiable chain.

What was authorized

A permission granted at a point in time against a particular policy state, often recorded separately from the system that later executes the operation.

THE GAP
where failures live
What actually occurred

The operation that ran, including its recorded inputs, transformations, outputs, and delivery result. Reconciling the two conclusively is often difficult.

NUBAS is being developed to close this gap structurally by binding each governed authorization to recorded evidence of the operation that followed.

How it works

Three steps. Every governed operation.

01

An action is declared

Before an automated system does anything consequential (move money, merge data, send a message, actuate a machine), the action is written down in a machine-readable form: who wants to do what, with which inputs, producing what.

02

Your rules decide

NUBAS checks the proposed action against signed, versioned policy and issues one of six deterministic verdicts: allow, allow with constraints, hold, escalate, deny, or emergency restrict. The governed execution path does not proceed without a verdict.

03

Proof is created

Two linked records are written into a tamper-evident ledger: one recording what was authorized and one recording the operation and its delivery outcome. The evidence-bundle design is intended to support independent verification by auditors, regulators, investigators, and other authorized reviewers without requiring access to the live operating system.

What that gives you

Authorization

Wrong actions can be stopped before execution rather than detected afterward. Decisions are deterministic, and every verdict includes a rationale a human can inspect.

Learn more

Evidence

You can demonstrate what your systems were authorized to do and what occurred afterward. The evidence architecture is designed to support independent verification without relying solely on the operator or requiring access to the live system.

Learn more

Accountability

When incidents occur, lineage identifies the affected operations, actors, inputs, and outputs. Structured recall and provable remediation are part of the product roadmap.

Learn more
NUBAS Platform

One platform for governing consequential actions.

Your automated systems run on top of NUBAS rather than governing themselves. The architecture is designed so automated workloads receive post-enforcement data and generate recorded operational evidence. In a conforming governed execution path, the workload must not bypass authorization or alter the evidentiary record.

Explore NUBAS
nubas · verdict streamSYNTHETIC
actionoperationverdictproofs
Product interface: in development · synthetic stream

Built for anywhere a wrong action is unaffordable

Additional ActionSpecs application profiles may extend the architecture into other regulated domains.

Our platform. Your rules.

NUBAS is in active development. The capabilities below describe the intended platform architecture derived from the public ActionSpecs protocol draft. Items marked Roadmap are planned for future versions and should not be interpreted as currently deployed functionality.

ledger · receipt chainappend-only
PROA 01J9ZK…HD3Mwhat was authorized
PROOD 01J9ZK…HD3Mwhat occurred
export evidence-bundle.taroffline verification design

Verdict Engine Enforcement

Deterministic policy decisions designed for machine-speed execution

  • Six verdict types mapped to any domain's dispositions
  • Deterministic: same input, same answer, fired clause cited; no probabilistic model in the path
  • Deployment architecture designed to support local and resource-constrained environments

Evidence Ledger Proof

Paired proofs of authorization and outcome, per operation

  • PROA ⇄ PROOD cryptographically linked in an append-only, hash-chained ledger
  • Lineage query design: intended to trace governed outputs through their recorded operation chain
  • Self-verifying export design: evidence bundles intended to support independent validation without access to the live system

Policy Control Governance

Rules as signed, versioned code, with controlled activation and governed change

  • Versioned policy activation designed to avoid enforcement discontinuity
  • Governance of the governor: rule changes are themselves verdicted and receipted
  • Governed override: attributed to an authorized actor and recorded as part of the evidence chain

Recall & Remediation Roadmap

When something goes wrong, prove the cleanup

  • Blast radius as a query: forward lineage is intended to identify affected downstream operations represented in the governed record
  • Compensating actions: every remediation step authorized and receipted
  • Honest reversibility: each operation classed as undone, contained, or irreversible, and provably so
Deployment architecture

Deploy where the action happens.

NUBAS places the verdict and evidence path inside the environment where consequential operations occur. It supports a local single-node runtime for development and evaluation, then deploys across managed, customer-controlled cloud, sovereign on-premises, and edge environments. Policy authority, enforcement boundaries, key custody, and evidence custody are configured according to each deployment model.

Managed deployment

Managed for enterprise

Managed NUBAS infrastructure for enterprise applications and agent workloads, with defined tenancy, policy, and evidence controls.

Customer-controlled cloud

Your VPC, your boundaries

Deploy within your VPC or private cloud, with customer-defined network, data-residency, key-custody, and evidence-custody boundaries.

Sovereign on-premises

Inside your perimeter

Deploy within organization-owned infrastructure for regulated, restricted, and sovereign environments. Operational dependencies, policy authority, cryptographic keys, and evidence custody remain locally controlled. Evidence bundles support authorized review outside the operating environment without access to the live system.

Edge & disconnected

Operate without continuous connectivity

Run local verdict evaluation, enforcement, and evidence capture close to the governed operation. Records synchronize through an authorized connection when connectivity becomes available.

Local development

Build and evaluate on one machine

Run the NUBAS verdict and evidence path as a single-node local runtime for development, integration, testing, and policy evaluation before deploying into managed, customer-controlled, sovereign, or edge environments. The local runtime follows the same ActionSpecs semantics and is the target environment for protocol conformance testing.

The same ActionSpecs verdict semantics, evidence requirements, and policy-governance model apply across deployment modes; local policies and application profiles can vary by environment. Deployment modes describe the NUBAS product architecture in development; validated operating envelopes will be published as implementations are tested.

Public working draft

Built from ActionSpecs. Open to examine and implement.

The public working draft defines the intended lifecycle, verdict semantics, paired-evidence model, and conformance requirements. It is published under CC BY 4.0 for independent examination, implementation, and contribution.

actionspec.jsondeclared → verdicted → receipted
{
  "action_id": "01J9ZK3W7Q4R8V2N6P0XA5HD3M",
  "principal": { "id": "agent-operations-7" },
  "operation": "data.release",
  "inputs": [{ "ref": "dataset:4471", "digest": "sha256:…" }],
  "proposed_output": { "destination": "partner-enclave-2" },
  "policy_scope": ["information.releasability"]
}
// verdict: allow-with-constraints (releasability rule R-114)
// constraints: redact restricted fields · bind output label
// PROA issued · PROOD issued after delivery
"

A consequential operation must not occur without a policy verdict, and no verdict or outcome may exist without a receipt.

The ActionSpecs invariant: the foundational rule NUBAS is built to enforce
The company

Adekka is a research-driven AI infrastructure company.

Adekka develops foundational ideas into open protocols and deployable infrastructure. We formalize the architecture, publish the protocol so it can be independently examined and implemented, and build platforms that apply it in real operating environments. NUBAS is the first platform being built through this process.

Public working draft

ActionSpecs

Our foundational research output: a public protocol draft for authorizing and evidencing consequential automated operations, available for independent examination and implementation.

Flagship platform

NUBAS

Adekka's trusted-action platform, the reference platform being developed from ActionSpecs. A verdict before each governed action; an evidence architecture designed for independent verification afterward.

Continuing work

What comes next

Application profiles extending the protocol into new domains, future protocol versions, and further infrastructure for consequential decisions and operations.

Read our position: the hard questions about autonomous action

Governing something that matters?

We're selecting a small number of design partners deploying agents and automation where wrong actions have real cost. Partners get early access, direct influence on application profiles, and evidence-verification capabilities designed for independent review.

Talk to us