Hypernovelty Institute Technical Intelligence

Flagship analysis · July 14, 2026

The World Is Building an Accounting System for Action

By Jordan Finneseth

AI regulation, provenance standards, event logs, and agent receipts are turning decisions into machine-readable records. The struggle now moves to the schema.

Two colleagues review a paper record beside an open laptop in a quiet office.

The record now has two readers: the person reviewing the decision and the system reconstructing how it happened. Photo: Cht Gsml / Unsplash.

A Law With One Quiet Sentence

On July 6, Illinois Governor JB Pritzker signed SB 315, the Artificial Intelligence Safety Measures Act. The law takes effect January 1, 2027. Beginning in 2028, covered large frontier developers have to maintain published safety frameworks and hire independent auditors to examine compliance. The law also creates incident-reporting and record-retention duties.

Most of that sounds like familiar institutional plumbing. Define the risk. Write the framework. Hire the auditor. Keep the files.

Then Section 10(c)(3) makes a quieter move. Summaries of catastrophic-risk assessments must be provided in a machine-readable format “to facilitate verification of model claims.”

The requirement pushes past a report posted for people to read. It asks for a record that software can index, compare, and test against later claims. One sentence in one state law points toward a much bigger transition: the world around AI is being rewritten so machines can inspect how decisions happened.

The Document Was Only Part of the Record

For most of modern institutional history, the explanation lived in a document. Forms, contracts, audits, minutes, filings, and reports gave institutions a durable way to describe themselves. A skilled analyst with enough time could work through them.

The deeper record was scattered underneath. It lived in emails, database entries, meeting notes, memory, and the sequence of choices that produced the final report. When something failed, responsibility had to be reconstructed from whatever survived.

A PDF can be easy for a person to read and structurally opaque to a computer. A machine-interpretable state exposes entities, permissions, provenance, timestamps, relationships, changes, and outcomes in a form systems can query.

The Securities and Exchange Commission’s Inline XBRL program offers a mature comparison. The SEC adopted phased Inline XBRL requirements in 2018. A single filing can be both human-readable and machine-readable, with contextual information attached to tagged data. A person sees a financial statement. Software can compare a value with other filings, identify the reporting period, and follow its definition.

That shift already changed how financial reporting can be searched. Now a similar logic is spreading into records of action.

Records by Design

The EU AI Act gives a bounded legal example. Article 12 requires covered high-risk AI systems to technically allow automatic event logging over the lifetime of the system, with traceability appropriate to its intended purpose. The rule applies to a defined category in a defined jurisdiction. That logging requirement changes the build. Covered systems need the technical capacity to leave a trail somebody can inspect later.

C2PA applies related thinking to media. Its Content Credentials specification defines tamper-evident, signed assertions about how an image, video, audio file, or document was created and changed. The specification draws a careful boundary around that record. It can help validate that assertions are associated with an asset, correctly formed, and free from tampering. It offers no verdict on whether those assertions are good, fair, or true.

That boundary matters. A receipt can be authentic and still describe a bad decision.

Money Was the First Thing Worth Counting

Traditional accounting records where money moved. Debit, credit, balance, reconcile. The ledger identifies parties, amounts, dates, and transactions so an auditor can inspect the history and challenge it.

The emerging layer records agency. A useful record has to show what the system observed, what it recommended, who authorized the next move, what was executed, what happened afterward, and whether anybody reviewed or repaired the result.

Each stage needs a distinct record. Otherwise “the AI decided” becomes a place to hide responsibility that was actually distributed across managers, developers, vendors, policies, and automated systems.

NIST’s Building Evaluation Probes into Agentic AI project shows what part of this could look like. Researchers are testing automated verifiers inside agent workflows and storing structured verdicts in machine-readable audit trails. The probes compare claims with trusted reference material and record how the evidence supports or fails to support the output. This is a research testbed rather than a binding standard. The shape is still useful: evidence about whether work passed travels with the work.

W3C PROV approaches the same territory from a data perspective. Its model describes the entities, activities, and people involved in producing data or other things. Those provenance records can help people assess quality, reliability, or trustworthiness and exchange histories across different systems. Provenance supports an assessment. It does not settle the verdict.

The Struggle Moves to the Schema

Here is where the stakes get political.

Who decides what counts as an event? Whose contribution gets recorded? What evidence has to accompany a claim? Which actions require prior authorization? Who can query the history? Who can correct or delete a record? What stays intentionally illegible?

The technical choices carry power. Whoever writes the schema shapes what the future machine can see. An audit trail designed by the organization being audited can be structurally complete and selectively revealing at the same time. A category omitted from the vocabulary may disappear from every dashboard, audit, and automated review that inherits it.

Illinois SB 315 requires machine-readable summaries to facilitate verification. The law establishes an obligation. Implementation will determine what those records contain, how comparable they become, and which facts remain outside the frame.

Standards bodies, regulators, research organizations, companies, and civil-society groups are building different parts of this vocabulary. Their categories will shape what counts as an actor, a permission, a provenance claim, an incident, and a successful repair.

The people who write the schema write the history machines will be able to read.

Your Agent Is a Small Institution

This lands at the individual scale too.

A useful personal agent needs an identity that persists across sessions. It needs durable memory that shows what it knew and when. Its permissions need clear scopes. Sources need provenance. Consequential actions need receipts. Spending and action limits should constrain what can happen without another authorization. The agent needs a way to explain and correct its history, with a firm boundary between private memory and externally auditable action.

That is the architecture of a small institution. It keeps “my agent did that” from becoming the personal version of “the AI decided.”

Both scales share the same requirement: delegated action needs durable history. The record becomes an evidence base for trust and the surface where contested accountability eventually lands.

The Ledger Cannot Be the Judge

Several limits remain even when the record is well built.

Integrity preserves one kind of evidence. It still leaves open whether a record was accurate when written. Event order can help reconstruct a process; fairness has to be argued from context. Search patterns can raise a useful question or guide an investigation. Causation takes more than a pattern. And when an operator controls the audit trail, the form may be complete while important facts stay outside it.

Blockchain can add an external integrity or coordination mechanism when parties do not share a trust layer. Some systems will benefit from that. Plenty of databases, logs, and machine-readable interfaces will work without it. A ledger still cannot decide whether the events entered were the right events or whether the schema captured what mattered.

Complete legibility carries its own danger. The European Data Protection Board’s guidance on data protection by design and by default keeps data minimization, purpose limitation, storage limits, correction, and erasure in the frame. Those principles push against the lazy assumption that keeping more data creates more accountability.

In one setting, a richer record helps a worker challenge an automated decision. In another, it becomes surveillance, defensive bureaucracy, or a liability archive the subject cannot meaningfully contest.

Machine readability creates a precondition for reconstructing action. Accountability still depends on honest reporting, independent review, enforcement, context, and a real path to correction.

Questions Worth Sitting With

If your organization relies on AI for decisions that affect people, work through these questions with a real workflow in front of you:

  • What would an independent reviewer need to see to determine whether your AI policy shaped a real decision?
  • Who holds the evidence linking an output to the inputs, sources, and authorizations that produced it?
  • Which actions leave no durable record? Was that choice deliberate?
  • Who can correct the record, and under what conditions?
  • If the machine-readable version omits something important, who would notice?

These questions have specific answers inside specific organizations. Those answers will matter as audit duties, disclosure rules, procurement standards, and litigation discovery catch up with how AI systems operate.

The world is becoming more legible to systems that can search, compare, and act faster than a human reviewer. Some decisions may become easier to reconstruct. And you may find your own work and behavior measured through categories you did not choose.

So when a machine-readable record enters your workflow, do not accept the schema as neutral. Find out who wrote it. Make sure somebody can challenge it. Keep a path to correction open.

Pull this thread.

Public sources