4SI Research · Institutions
More capable AI.
Accountable action.
We research how physical verification, human judgment and enforceable authority can work together as intelligent systems take on more responsibility.
The institutional question
What may the system do?
Who gets to decide?
From an agent preparing a payment to software controlling equipment, our focus is the boundary between a proposed action and permission to execute it.
01 · AI safety & authority
Capability is not
permission.
An agent may have the tools to act. Your institution still needs to define its mandate, limits and the moments that require renewed human authority.
- DefineThe permitted action.
- VerifyThe required authority.
- EnforceThe execution boundary.
What this means for your organisation
- Payments: a prepared transfer is not permission to release funds.
- Privileged access: an open session is not unlimited authority.
- Physical operations: an instruction is not approval to move a machine.
The research examines where a control could stop an action, how approval is bound to that action, and which alternative paths could bypass the check.
Explore the control-primitive reportThe scope of the AI safety claim
This is research into controlling authority at defined execution points. It does not solve model alignment or establish that an AI system is safe.
The published control argument depends on a secure verification mechanism, enforceable policy and coverage of relevant action paths. Coercion, compromised endpoints, human misuse and ungated pathways remain material risks.
02 · Physical verification
Connect the record
to reality.
A valid credential or product record can still leave a physical question unresolved. Is this the expected person? The expected object? At the required moment?
Questions for a defined workflow
- What physical fact would change the decision?
- Can substitution, replay or a changed object defeat the check?
- What is observed, what is inferred, and what remains unknown?
- Which system retains the final decision and audit record?
03 · Human factors
A control people
can actually use.
Operators need to understand what a request permits, recognise when something is wrong and recover safely when a check fails.
What should be evaluated
- Comprehension: does the operator understand the exact action?
- Friction: does the check interrupt the workflow at the right moment?
- Exceptions: can staff handle failure without silently bypassing policy?
- Access: can people with different abilities use the control?
These are evaluation priorities, not reported study results. Any workflow trial should measure human factors alongside technical failures and false decisions.
04 · Economics & deployment
Make the control
worth the cost.
A technically interesting check only matters if it changes a decision. The economic question includes avoided loss, integration effort, operator time and the cost of getting it wrong.
Explore the economic researchExplore the thinking.
All publicationsVerification as a
Control Primitive
How authority could be enforced before an AI system acts.
Read the researchPhysical Trust at the
Digital–Physical Boundary
What a digital record can establish—and where physical evidence matters.
Read the researchOur approach
Ask clearly.
Test rigorously.
Good research makes its assumptions visible and gives others a way to challenge them.
From a question to evidence
- Define the decision. Who or what must be verified, and what happens if the check is wrong?
- Set the test. Specify threats, operating conditions, human factors and success criteria before measuring.
- Keep the whole result. Record failures, false decisions, recovery paths and unresolved limits.
- Decide what follows. Continue, revise or stop against the agreed criteria.
How to read our research
The public work includes two non-peer-reviewed reports, an economic note and a perspective. The reports develop frameworks and conditional arguments; they do not establish validated product performance.
Usability is a research direction here, not a claim of completed user studies. A prototype result and a deployment evaluation answer different questions. Neither should be inferred from a paper.
Publication records and full textsSources, review and corrections
We separate documented facts, interpretation and proposed mechanisms. Original records and the assumptions behind material claims remain available for review.
Bring one decision
worth getting right.
Define a workflow, its risks and the evidence needed to evaluate a physical check.
Discuss a workflow