More than two billion dollars in penalties. More than a hundred firms. Not for misconduct — for gaps in the records they were required to keep. Every one of those was a record of what a person did.
The work is now being handed to AI agents. The record is still being asked for, but with greater oversight and increased fines.
The firms in that number were not accused of lying to customers. They were penalized because, when regulators asked, the record was incomplete — conversations held where nobody was keeping the book.
The sweep that produced those penalties has wound down. The enforcement has not: examinations continue, and firms were still being fined for the same failure in 2026.
Companies are now handing real work to AI agents — service, sales, claims, underwriting, operations — and the machines that follow will act in the physical world.
When one of them is challenged, the question is the same question it has always been, and it is historical:
Every one of these rules creates a duty to hold a record. None of them defines what makes a record believable.
On a date that is already set, someone — a regulator, an auditor, an insurer, a customer’s lawyer — will ask your company what your AI knew and what it did.
You will answer with something.
The question is whether they have to take your word for it.
A log says what it contains. A provable record can show what happened to it. Most systems can claim the first. Almost none can demonstrate the other three.
Not asserted. Measured on the engine, and enforced by the test suite on every build — so a claim that stopped being true would fail the build before it reached a customer.
A permanent, tamper-evident record of what AI systems knew and what they did — built to answer two questions that regulated industries already ask about people, and nobody can yet answer about machines.
The first market. Financial-services agent channels, customer-facing AI that makes binding representations, healthcare administration, insurance claims — any enterprise that must answer a regulator about what its AI did.
The physical wave. A robot is an agent whose acts are physical, and TIE needs no new machinery to serve one — manufacturing, surgery, security, wherever a machine’s act applies knowledge an expert or regulator would examine afterward.
Standards bodies, equipment makers and expert firms licensing their knowledge to fleets as sealed artifacts, with use accounting that doubles as the royalty basis.
The knowledge a machine consults before it acts — a profession’s rules and standards, versioned and citable, so what the system held on any past date stays answerable.
The record of what the machine declared it did — sealed page by page, added to but never rewritten, holding what it was told, what it did, and what it left out.
Proof that works on strangers. A customer, an insurer or a court can verify the record against the interest of whoever handed it over.
If nothing is ever thrown away, does the bill grow forever? It is the right question, and the answer is measured rather than promised.
The compression is lossless and deterministic: nothing is summarised, sampled or thinned with age, the record restores byte-identical, and every seal still verifies afterward — proven on every build, over every record.
The single limit is at the working edge, on the page still being written, so that nothing accumulates where the record is still forming. Everything sealed leaves for the permanent store, and the permanent store has no practical ceiling. How much one customer’s memory can hold is a storage bill, never a system limit.
Scope: the figures describe the structured record. Prose-heavy records land at the low end. Attachments — images, PDFs, audio — are already compressed and do not shrink further, so a media-heavy deployment is quoted a blended figure. Measured on the real engine, September 2026; re-measured on real records as deployments begin.
It cannot block what an AI does and never sits in its control loop. TIE witnesses; it never gates. We do not judge the conduct — we make it examinable, by a regulator, an auditor, an insurer or a court. That refusal is exactly why the record can be believed.
It contains no AI. No model, nothing to hallucinate, drift or go down. It does not retell the record — it returns it, so the same question produces the identical answer today, next year, and in front of a tribunal. AI systems are its customers, never its components.
A record you have to trust the vendor about is the failure this exists to end. So the proof is built to work on a stranger — and against us.
TIE holds none. The root of the record’s signatures lives inside your own cloud’s hardware boundary — the key service your security team already audits. It can sign. It can never be exported. And every signature it makes lands on your audit trail, not ours.
Every touch of your record is itself a sealed record, and you hold its chain. An operator who alters or deletes an access entry is caught by your own verification tooling — with nothing in the detection path that the operator controls. That operator includes us.
Two single files — one in Node, one in Python — importing nothing of ours, run on your machine, offline. Four independently written implementations, sharing no code, agree on every value. The chain closes to its root by the root’s signature, not by its name.
A security page that lists no limits has limits you have not found yet.
These are ours, stated here rather than discovered later. Each is a boundary of the design, not a defect in it.
TIE’s property is integrity, not secrecy — a record that can be shared with whoever is approved and still proven by anyone who holds it. On legal compulsion the answer is architectural: we hold nothing.
Every AI company is racing to make machines act.
We are building the record that proves what they knew — and what they did.
And the record outlives everything. Including us.
briefing@memoryofrecord.ai