Subtractive Invariance
Capability is shaped as much by what a system reliably excludes as by what it adds. Removing unsafe or unauthorized possibilities can increase the work a system may be trusted to perform.
The Thesis
Invariant Intelligence.
Capability that can be trusted with consequential work, because the constraints that matter are preserved by design.
The capability-versus-trust gap
Capability is growing faster than authority to use it.
An AI system may be technically able to move money, negotiate with customers or coordinate other agents. Whether a person or institution can authorize it to do so is a different question.
For example, an agent able to pay invoices is valuable only if it pays the right invoices, within its delegated limits, in line with policy and with a clear record of why. Each of those conditions is an invariant.
Three dimensions
The invariants that define trustworthy work.
Structural
What systems can reliably do
Consistency of behaviour under real conditions — correctness, robustness and predictable failure. Example: an agent produces the same valid ledger entry every time.
Constitutional
What systems are permitted to do
Legitimacy — rights, laws, policies and delegated authority. Example: an agent never exceeds the spending authority granted to it.
Operational
What systems are intended to accomplish
Alignment with the objectives and intent of those it acts for. Example: an agent optimizes for the customer's stated goal, not a proxy.
Effective Capability
The intersection is where useful work lives.
Effective Capability is the capability that remains once all applicable invariants are preserved. The aim is to make that region as large as possible.
Effective Capability
Useful capability that remains reliable, legitimate and aligned with intent — the region where all three invariants hold.
Open questions
Research-led and empirically grounded.
We distinguish clearly between working hypotheses and conclusions supported by experiment. The concepts below are currently framed as hypotheses or active lines of investigation.
Risk of repair
Correcting a system after a consequential failure carries costs — financial, institutional and human — that often exceed the value of the original task. Preventing violations is preferable to repairing them.
Measurable reliability & accountability
Trust should rest on evidence. We seek metrics that show whether invariants held, who authorized an action and how outcomes can be traced.
Resources, cost & outcome
We study how computational resources and economic costs relate to useful outcomes — so that reliability is achieved efficiently rather than by brute force.