Close-up of a robotic manipulator arm, an illustrative photograph

Authority and evidence for machine actions

When a machine acts for you, can anyone prove who allowed it?

Agents now change and ship real work. Identity says who they are; logs say what ran. Neither proves who allowed the action, within what limits, and what came of it. GuardianChain's architecture connects those into a record the recipient can inspect.

Apply as a design partner Explore the architecture

Design-partner stageThis site is informational. It does not provide platform access.

Starting point

Bounded, human-governed machine work: one delegation or handoff where someone must later show who authorized it.

The architecture

An authority, evidence and finality layer that sits beside identity, orchestration, policy and logging.

Available today

This informational site, inquiry forms, and Partner Access to your own inquiries. Not platform access.

The problem before the architecture

Pilots stall at the same five questions.

Authority
Who allowed the agent to act, on what, and for how long? A prompt is not a grant.
How the chain addresses it
Architectural link: human authority and bounded scope, recorded as an explicit grant rather than inferred from a credential.
Durable proof of changes
Logs held by the tool that did the work are not a record the recipient can lean on.
How the chain addresses it
Architectural link: machine action tied to its grant, with evidence references attached to the outcome.
Continuity across tools
Work moves between assistants, editors and services. Context and limits are lost at each hop.
How the chain addresses it
Architectural link: GuardianBridge session records written to GC-1, so authority does not silently renew between tools. Coverage depends on integration.
Independent verification
The party receiving the work needs to inspect it without trusting the producer's summary.
How the chain addresses it
Architectural link: FactReceipt, a portable record the recipient can check. Verification is not acceptance; that decision stays with the recipient.
Audit and rollback
Security and incident teams need evidence they can reconstruct from, and a path back.
How the chain addresses it
Architectural link: revocation and expiry recorded in GC-1 rather than inferred. GuardianChain does not itself perform rollback.

Design-partner stage. This website does not execute these workflows.

The chain we connect

Four links. One evidence chain.

GuardianChain connects who authorized an action, its permitted scope, the consequential decision, and independently verifiable evidence of the outcome.

AUTHORITY → SCOPE → ACTION → EVIDENCEArchitecture overview

Who allowed this?Human authority
A named, accountable person or role grants the action. A prompt or API key is not a grant.
Within what limits?Bounded scope
What may be done, on which resources, under which conditions, and until when. Refusals and restraint are recorded too.
What was decided and done?Machine action
The consequential decision made by an agent or automated system, linked to the authority and scope it relied on.
Can someone else check it?Outcome evidence
A portable record a recipient can inspect independently, without trusting the producer's own summary.

01 / Human authority

Record contents: the accountable grantor and the grant itself.

Without it: an agent acts on a prompt or credential, and no one can show who allowed it.

Use the arrow keys to explore the record structure.

Human authority

Record contents: the accountable grantor and the grant itself.

Without it: an agent acts on a prompt or credential, and no one can show who allowed it.

Bounded scope

Record contents: the permitted actions, resources, conditions and expiry.

Without it: authority silently renews or widens across tools and sessions.

Machine action

Record contents: the consequential decision, linked back to its grant and scope.

Without it: logs show what ran, but not whether it was allowed.

Outcome evidence

Record contents: evidence references a recipient can inspect separately.

Without it: recipients accept work on the producer's word.

GuardianChain architecture shown for explanatory purposes. Three U.S. provisional patent applications have been filed covering the underlying technology.

Why existing tools are not enough

Each system holds one piece.

Identity, orchestration, tool connectivity, policy and logs all matter, and we build on them. None of them, on its own, establishes the complete chain from human authority to checkable outcome.

Typical coverage by category, in isolation. Individual products vary; this is our framing, not a benchmark.
Category	Primary job	Who authorized	Permitted scope	Decision linked	Independently checkable
Identity providers	Authenticate people, services and agents	Partly	Not by itself	Not by itself	Not by itself
Agent orchestration	Plan and run multi-step agent work	Not by itself	Partly	Yes	Not by itself
Tool connectivity	Connect agents to tools and data	Not by itself	Partly	Yes	Not by itself
Policy engines	Evaluate rules at decision time	Not by itself	Yes	Partly	Not by itself
Logs and observability	Record events inside each system	Not by itself	Not by itself	Partly	Partly
GuardianChain architecture	Connect the four, alongside the systems above	Architectural focus	Architectural focus	Architectural focus	Architectural focus

Product inventory

Three parts, one record.

Three architectural roles, with workflow fit and availability qualified through design-partner discussions. This website provides information and inquiries, not platform access.

Canonical authority and evidence substrateDesign-partner stage

GC-1

The record of who authorized what, within which scope, and what evidence exists for the outcome. The other parts depend on it.

How GC-1 works

Human and machine interface for governed sessionsDesign-partner stage

GuardianBridge

Where a person grants, bounds, watches and ends a machine session, so authority is explicit at the moment it matters.

How GuardianBridge works

Portable, independently verifiable evidenceDesign-partner stage

FactReceipt

A statement a recipient can check without trusting the producer: what was claimed, by whom, under what authority, with what support.

How FactReceipt works

Complements, never replaces

We sit alongside what you already run.

Identity providerWe rely on your existing identity systems to know who someone is.

Secrets managerCredentials stay where you keep them today.

Agent frameworkYour agents and orchestration remain your choice.

Universal gatewayWe do not need to sit in every request path.

Compliance dashboardWe do not certify compliance or legal sufficiency.

Where handoffs carry weight

Settings we want to test with partners.

Two engineers testing a white humanoid robot in a workshop — **Technical and AI delivery**Delivery acceptance when agents did part of the work.

Researcher controlling a robotic manipulator in a lab — **Regulated research handoffs**Provenance and authority travel with materials. Not a compliance claim.

Wheeled autonomous delivery robot on a pavement — **Freight evidence reconciliation**Preserved evidence when parties disagree about a shipment.

Orange industrial robot arm on an automated line — **Industrial operations**Where automated decisions have physical consequences.

Photographs illustrate settings and are not customers or deployments. Photo credits.

Pressure is building

The control gap is no longer theoretical.

Controlled research, not a forecast of routine failure. Read the sources and their limits

Controlled scheming evaluationsOpenAI's o1 System Card reported basic in-context scheming in deliberately constructed Apollo Research evaluations, not normal deployment rates.
High-stakes simulationsAnthropic's summer 2026 research documented covert sabotage, fraud assistance and motivated mislabeling in simulations, not real incidents.
Real systems, in evaluationsAnthropic reported four incidents of unauthorized access to real third-party systems during misconnected cyber evaluations. Not customer deployments.
Lifecycle obligationsEU AI Act high-risk requirements include logging, documentation and human oversight on a 2027 to 2028 timetable. No certification or automatic compliance is claimed here.

Help shape the next stage.

We are working with a small number of teams that have a real delegation, handoff or reconciliation problem. You get direct access to the builders and influence over scope. Applying implies no platform access, commitment or guarantee.

Apply as a design partner Review candidate use cases

Not ready to talk yet

Follow progress.

Inquiries are stored only to respond and evaluate partnerships. Do not send secrets or sensitive evidence. Read the privacy notice.

Record interest in updates

This records your interest only. Email delivery is not active yet.

Our technology ecosystem

Built with established tools.

The tools behind this website, alongside selected names from the wider GuardianChain stack. Directory listings are not live platform integrations, partnerships or endorsements.