A POSITION FROM LICET · SEPTEMBER 2026
The Warrant
The decision that a system may act deserves to be as durable as the record of what it did.
A classifier removes a post. A ranker changes what a person sees. An age-assurance tool determines which experience someone may access. Each is an action taken by an institution, through an automated system, under a claim of authority.
Somebody considered the policy. Somebody reviewed the system. Somebody decided it could go live. The question is whether that decision exists in a form the institution can still produce, explain, and test against what has changed.
The authority exists. Can you produce it?
Ask who authorized a classifier and on what basis. Trust and safety may describe the policy it enforces. Legal may describe the obligation the policy answers to. Engineering may describe the launch review it passed.
Each answer contributes something essential. Yet the complete decision may live across a policy document, a launch checklist, a legal memo, a message thread, and a person’s recollection of a meeting. The pieces exist. The record has not been assembled.
That becomes expensive when the people change, the system changes, or an examiner asks a question the institution should already be able to answer. Reconstructing a decision is work. Recording it at the point of authorization makes that work useful more than once.
What a record must contain.
The record must name an accountable person who accepts the residual risk. That acceptance should have a scope, a date, and conditions. A list of reviewers is useful; a clearly identified acceptor answers a different question.
Every condition must be traceable. What authority prompted it? What did the institution decide that authority required? Which policy expresses that interpretation? Who approved the condition? Which control is declared? What evidence is required to assess whether it is in place? Requirements and collected results must be distinguishable.
A missing link should be visible before the record issues. Otherwise, the document gives the appearance of completeness without the substance.
Judgment stays with the institution.
The record must distinguish an institutional interpretation from a statement about what the law permits. Software can assemble a decision and make its basis examinable. The institution still owns the interpretation and the acceptance of risk.
A licet record captures your institution’s approved interpretation of its obligations. It records what you decided and who accepted it. It does not state what the law permits.
A decision cannot stay current by standing still.
Models are replaced. Policies are revised. Capabilities are added. Obligations begin applying. The question is whether the decision made for yesterday’s system still covers the system operating today.
An authorization therefore needs a boundary and a means of revalidation. When a proposed action exceeds that boundary, the system needs a new decision. When evidence or assumptions change, the record needs to be examined again.
The original record should remain intact. A later review should explain what changed, who considered it, and what decision followed. Revalidation should add to the history, rather than erase it.
The record belongs in the workflow.
Enforcement platforms capture decisions a system makes. An authorization record captures the decision that it may make them. The two should meet where action happens.
The licet prototype issues that record, records permission decisions and links them to their authorization. Production executor integration and operational-evidence assessment are further work. The goal is a decision that can be inspected without reconstructing the room in which it was made.
The ask is one sentence.
Name the executive who accepts the risk for each system.
Start with one system already in production. Record its authority, conditions, and accountable acceptor. Put the result in front of the people who need to rely on it. Learn what the record must become by using it.
That is the work we are beginning with design partners.