Sectors
Industries where somebody eventually asks who decided
The problem is the same shape in each. What differs is who asks, how long afterward, and what it costs when the answer is not there.
Science and research
Who checks the result before the reviewer does?
Laboratories that put AI to work on the science itself: the hypothesis, the experiment, the analysis and the paper.
AI now proposes the hypothesis, selects the data, runs the experiment, chooses the metric and writes the paper. This year three such systems were published in the leading journals, and an autonomous laboratory ran tens of thousands of experiments with the scientist out of the loop.
The field has measured its own problem. The agent holds both the result and the evidence for it. Its own surveys this year found that most systems release their code but not what a reviewer would need to re-run the work, and that in open-ended research tasks an agent will meet its own test without doing the work, unprompted. When the agent can touch its own scorer, it learns the scorer.
A result that cannot be checked is a claim. The journal will not list an AI as an author, because an author answers for the work and the AI cannot. The funder holds the scientist to the same rule, whichever tool produced the data.
ProofSmith lets a laboratory delegate research to AI at machine speed and still answer for every result.
Science is the high ground for Independent & Provable GOVERNED DECISION LINEAGE. It is the oldest arena for the rule: a result counts when someone other than its author can check it.
Someone has to rely on those results without taking the agent’s word:
- The scientist, who answers for every result
- The reviewer and the journal
- The funder, the patent examiner, the regulator
You write the protocol and grant the authority. An independent arbiter, outside the agent, permits or denies each step before it is taken — the dataset, the metric, the change to the protocol, the order to the laboratory — then issues a signed proof of that decision. The model learns on a record it cannot edit. The reviewer checks the run without asking the laboratory or the agent.
Keep your data. Prove your results.
Innovative technology
How do you know your own teams are not changing their logs?
Innovation companies that use AI to run internal testing and competitions. No one outside checks them yet.
Innovative technology companies run internal technology and product competitions. Two teams build two versions. The company tests both, and the winner ships. AI now does more of that work, and it learns from the results.
When the record is kept inside the company, whoever keeps it can change it. He can delete an entry, leave out a failure or change a rule after the fact. Most people never will. But no one can tell which record was touched. Not the team that lost. Not you.
The best teams spend time defending their numbers. Others learn to shade them. Speed drops.
If you want to go faster, build better brakes.
Every internal competition has a winner and a loser.
Someone has to rely on those actions without taking the actor’s word:
- The chief executive
- The evaluators
- The team that lost
You write the rules and grant the authority. An independent arbiter, outside every team, permits or denies each action before it is taken, then issues a signed proof of that decision. Settle internal races on the record. The record shows how each run was governed, and lets the loser check it.
Keep your secrets. Prove your decisions.
Regulated
These six have something in common that most of the economy does not. In each, a decision is reviewed by a party who was not present when it was taken, who has no reason to take the acting organization’s word for anything, and who arrives months or years later — an inspector general, an examiner, a regulator, a coalition partner, a court.
That reviewer is the reason the record has to be built for somebody else to read. Elsewhere, a self-kept log may go unchallenged for years. Here, it is the exposure.
National security
A Combatant Commander cannot prove an autonomous action stayed within authority. No fielded system can.
Authority, intent and rules of engagement are written with great care, and then the record of whether they were honored is kept by the system that acted.
Authority in a military organization is explicit. It is delegated in writing, bounded in scope and time, and its limits are the difference between a lawful action and an unlawful one. That discipline is old and well understood, and it works, for as long as a human being sits in the decision path, because that human being is the accountability.
Autonomy removes the human being from the path and puts nothing in that place. What remains is a log, written by the acting system, held by the acting system, in the acting system's format. When an inspector general, a coalition partner, a congressional committee or a court asks whether an action stayed inside the authority granted, the answer available is the actor's own account of itself.
The House has asked for the record. The defense authorization bill it passed on 22 July 2026 would direct the Department to update its policy on autonomous and artificial intelligence-enabled systems so that it preserves human command responsibility for the use of force, identifies the commanders responsible for authorizing it, and sets requirements for auditability, traceability and accountability. By Executive Order 14434 of 29 September 2026, the executive branch now calls that technology Super Intelligence. A policy can require the record. It cannot produce one that does not depend on the word of the system that acted.
A coalition partner cannot be asked to trust an ally’s internal logs, and cannot be handed the intelligence behind a decision. What the partner needs is what every other reviewer on this site needs: proof that the decision stayed within authority, checkable without the ally’s cooperation. Shared confidence, not shared secrets, moves a coalition at speed. Classification and release authority stay where they are.
Authority delegates down the chain of command. Verifiable proof returns up.
The record is produced for somebody who was not there.
An inspector general does not ask whether the system was confident. The questions are who authorized it, and where the record is.
Where it bites first
Proving that an action stayed inside a delegated grant, to somebody outside the chain that executed it.
Keeping oversight in contact with systems that decide faster than people can review.
Separating error from misuse from adversary action, on the first day of an inquiry rather than the ninetieth.
Critical infrastructure
The control action fired. The authority for it is a matter of reconstruction.
Grid, water, pipeline, rail and plant operators are automating control decisions faster than they can evidence them.
Safety-critical industries already know how to govern a decision. Authority to operate is bounded, procedures are written in advance, and the operator who takes an action outside the envelope is identifiable afterward. The regulatory regimes built on top of that — reliability standards, safety cases, accreditation — all assume a named human being at the point of action.
Automated and AI-assisted control erodes that assumption quietly. A setpoint changes, a breaker opens, a valve closes, a maintenance action is deferred. The system logs that it did so. What the log does not carry is the authority the action was taken under, or any way for a regulator to test that authority without the operator's cooperation.
The consequence is felt at the worst moment. After an incident, the organization is asked to show that automated actions stayed inside the approved envelope, and the showing is an exercise in reconstruction, performed by the operator, from the operator's own records, under pressure, in the window where credibility matters most.
An operator who has to reconstruct authority after an incident is defending a record, not producing one.
Each of them arrives after the event and reads what the system left behind.
An incident investigator does not want the operator’s summary. The investigator wants the sequence, timed, from a source the operator did not write.
- A sector regulator examining whether automated actions stayed inside the envelope
- An accreditation authority deciding whether the system may operate at all
- An incident investigator reconstructing a sequence after the fact
- An insurer pricing an operational risk that nobody can evidence
- A safety case author who must argue the system is adequately controlled
Where it bites first
Showing that an automated control action was inside the approved operating envelope at the moment it fired.
Producing evidence for a regulator on demand, rather than assembling it by hand after an event.
Distinguishing a fault from a misconfiguration from an intrusion, while the incident is still live.
Banking and financial services
The decision was automated. The accountability was not.
Credit, trading, payments, onboarding and surveillance decisions are moving to models and agents. The obligation to evidence them did not move anywhere.
Automation has outrun the evidence. A model or an agent now participates in credit decisions, payment authorization, customer onboarding, exception handling and surveillance triage. The institution retains every obligation it had before, and holds, in place of a decision record, an application log written by the application that acted.
The exposure is concentrated where the examination is hardest. A fair-lending review, a model validation, a regulatory examination, a customer complaint, a litigation hold — each asks the institution to evidence a specific past decision to a party that has no reason to take the institution's word for it, and each is answered today with records the institution produced about itself.
Independent challenge is already expected in this industry. The only question is whether the evidence it examines is independent too.
The examiner was not in the room.
An examiner does not take the institution’s word for a decision. The examiner tests the record the institution did not get to edit.
- A prudential examiner testing whether decisions stayed inside delegated authority
- A model validation function expected to challenge with sufficient independence
- An external auditor supporting a management assertion
- A customer, a plaintiff, or a court, in an adverse action or a dispute
- A supervisor abroad, applying an operational resilience regime to the same firm
Where it bites first
Evidencing that an automated credit, payment or trading decision was inside the delegated authority in force at the time.
Producing a decision record that an examiner or a validator can check without relying on the acting system.
Turning examination readiness from a project into a property of the running system.
Healthcare
On whose authority was that clinical decision taken?
Clinical decision support, prior authorization, triage and documentation are automating. Liability and the standard of care are not.
Algorithmic systems now shape decisions across the whole pathway: triage and acuity scoring, diagnostic support, treatment recommendation, coverage and prior authorization, documentation, coding. Each of them influences what happens to a patient. Those systems were built to improve a decision, not to evidence one. A record showing what was proposed, what was accepted, and on whose authority is not a by-product of clinical software. It has to be designed for.
Two audiences arrive later, and neither takes the record on trust. The first is the patient, asking why a decision was made. The second is a court, asking whether the standard of care was met, where the answer depends on distinguishing what the algorithm proposed from what the clinician decided — a distinction the clinical record is seldom built to make at the time.
A patient is entitled to know who decided. That is not a technical requirement. It is the oldest requirement in medicine.
Each of them meets the decision after it was made.
A patient asks one question: who decided this, and on what authority? A court asks the same question, under oath.
- A patient asking why a decision was made, and by whom
- A court applying the standard of care to a specific episode
- A peer review committee separating clinician judgment from tool recommendation
- A payer, a provider and a regulator disputing the same coverage decision
- A device or software regulator assessing a system in clinical use
Where it bites first
Showing, afterward and to an outsider, which was the algorithm’s proposal and which the clinician’s decision.
Showing that an automated determination was inside the authority the organization granted it.
Giving the patient, the payer and the court a record that does not depend on the vendor's cooperation.
Legal
Who decided what went into the filing, and on whose authority?
Research, drafting, discovery review and the privilege call are moving to AI. The duty of candor, the duty to supervise and the privilege are not.
AI now finds the authority, drafts the brief, reviews the production and proposes the privilege log. Courts have sanctioned filings that cited cases no court ever decided, and the orders name the lawyer who signed. The signature on a filing certifies that a person checked the work. The record of what the tool proposed and what the lawyer accepted is seldom kept, and where it is kept, the tool keeps it.
The rules already reach the machine’s part of the work. A lawyer owes the client competence and confidentiality, owes the court candor, and owes supervision over any assistant, human or not. Each duty is discharged by a decision: to rely on a citation, to withhold a document, to let a tool see a privileged communication. The duty is the lawyer’s. The record of the decision should not be the tool’s own account of itself.
The law runs on evidence, and a record a party produced about its own conduct is the weakest kind. A record made at the moment of decision, signed by a party that did not act, and checkable by a party that needs nothing from the firm or the tool, is the kind the rules of evidence were built to weigh. Those rules have a place for a record generated by a reliable process, certified by a qualified person. The certification is a person’s. What sits under it should be proof, not the tool’s word.
The court does not take a party’s word. The record of what the machine did should not ask it to.
Each of them meets the filing after it was made.
A court asks one question of any record: who made it, and can it be checked without trusting the party that offers it?
- A court applying Rule 11 to a filing drafted with AI
- Opposing counsel contesting a privilege log a model proposed
- A client asking what the tool saw, and where it was sent
- The bar applying the duties of competence, confidentiality and supervision to a lawyer’s use of AI
- A judge weighing a machine-generated record as evidence
Where it bites first
Showing what the tool proposed, what the lawyer accepted, and on whose authority the filing went out.
Giving the court, the client and opposing counsel a record of the machine’s part that does not depend on the firm’s or the vendor’s cooperation.
Separating, afterward and under oath, the model’s proposal from the lawyer’s decision.
Regulated industries
You sign the attestation. What actually stands behind the signature?
Conformity declarations, management assertions and authorizations to operate all rest, in the end, on records the signer produced about the signer.
Every serious regulatory regime now ends in a signature, and the instruments are set out under Obligations. This section is about the party who receives one. A regulator, an accrediting official, a supervisor, a board: each is handed a signed statement and asked to rely on it, and each is expected to test it rather than take the signer’s word.
Today, in most organizations, what stands behind the signature is a body of evidence the organization assembled about its own conduct, from its own systems, in its own formats, ahead of a deadline. The signer is asked to certify a thing that only the signer can see, and the recipient is asked to rely on a thing that only the signer can check.
No one jurisdiction failed. The default is the same everywhere, and it held while the signer could vouch for the people who made the decisions. When systems make the decisions, the signer is vouching for a record produced by the thing being vouched for.
We do not replace the attestation. We put proof under it.
The signature is asked for by somebody who will have to rely on it.
The recipient of an attestation wants to check it without asking the signer. If checking requires the signer, it is the signer’s word twice.
Where it bites first
Making the evidence a by-product of operation instead of a project with a due date.
Giving the recipient of an attestation a way to check it without the signer's cooperation.
Demonstrating that automated decisions stayed inside the authority the organization granted.
The demonstration
Built around your problem, not around our example.
Tell us which of the six is costing you, and that is the one we will build.
Start a conversationSee the six problems
We are not asking you to trust us. We are asking you to let us prove it.