Specify which rules take precedence.
Order theory makes authority relationships explicit. The organization remains responsible for the authority it delegates and the rules it chooses.
Constitutional computing
Probabilistic reasoning. Deterministic permission.
Constitutional computing separates proposal from authority. An AI may generate a candidate action; that fact confers no authority for the action to occur. A mathematically specified process determines what change is admissible, what must remain true, and whether the proposal may become committed state.
What is constitutional computing? A short definition, and how it differs from Constitutional AI
The idea
Rules that sit outside the model.
A model can be persuaded, confused or simply wrong. Authority should not depend on the model being right.
The governing rules are explicit, bound to a reviewed version and evaluated independently of the model. The model never holds the authority itself.
Three foundations
What the decision process rests on.
State the properties to preserve.
Temporal logic expresses constraints on permitted behavior. A formal specification makes those constraints available for checking.
Bind rules to the reviewed version.
Cryptographic signatures and digests establish artifact identity and integrity. They do not establish that a policy is correct or complete.
Accepted, forthcoming, founder-authored
Constitutional computing in IEEE Computer.
Jonathan Hirshon’s two-part feature, Checking LLM Outputs Before They Become State, is scheduled for December 2026 and January 2027.
Publication of the architectural thesis is distinct from independent validation of an Inktomi deployment.
Part I · December 2026Checking LLM Outputs Before They Become State: The Problem and the Admission GateDOI 10.1109/MC.2026.3675280
Part II · January 2027Checking LLM Outputs Before They Become State: Prompt Framing Across Four ModelsDOI 10.1109/MC.2026.3733511
Technical evaluation
Evidence from your own workflow.
A useful evaluation reproduces the relevant controls in your environment, measures their costs, and records what is and is not covered.
- Agreed permitted, approval-required and refused cases
- Direct-call and uncovered-path tests
- Decision records and independent replay
- Coverage and reconciliation findings
- Integration effort and end-to-end overhead
In practice
Inktomi is the first application.
Inktomi applies constitutional computing to enterprise AI agents: each consequential action is checked before it runs, and every decision can be replayed.
See it in 30 minutes.
Explore Inktomi end to end: how it governs agent actions, handles exceptions, exposes evidence and replay, and fits your environment—with time for your questions.