The problem

Regulators now ask for evidence of how a model was developed; this is that evidence. A record written after the fact is testimony, not evidence. It will not survive a hostile regulatory review.

What it delivers

Emits a signed, regulator-legible record of every structural modification made during training. It documents exactly which component changed, what diagnosis triggered it, what treatment was applied, at what step, and within what declared bound. This is generated natively during training. It is not written after the fact.

The record from an archived run, opened at the bottom of the replay: every correction with its component, diagnosis, step, magnitude and the declared ceiling it stayed inside. Generated during training, not assembled afterwards.

The same network, twice. One sealed, one open.

—

Loading the archived run…

A neural network is damaged partway through training. The version on the right names which parts broke, two steps later, and repairs itself. The one on the left cannot say anything about it, and never will.

Epochepoch 0 / 199

What each one can answer, at this epoch

The question	Standard CNN	ORMAS

The full replay page carries three further scenarios, including the adversarial injection where this architecture measures worse than the baseline.

Who it is for

A fit for

Regulated medical device teams
Medical AI governance boards
Financial model risk managers
EU AI Act compliance counsel

Typically triggered by a failed audit or an impending regulatory deadline.

How it works

The Diary is built from 3 of the six capabilities of the ORMAS architecture. Each is a property of how the network trains, not a tool bolted on afterwards.

01ACCOUNTEmit a bounded, timestamped, causally attributed record of every modification the network made to itself.
02PROTECTGate a component out of the forward path when its health falls, so a failing region cannot poison what follows.
03REPAIRApply a bounded, conserved, logged correction to a named component under a declared per-diagnosis bound.

Where it is bought

This is why ORMAS is one product rather than seven. The Diary is the same mechanism in every row below; only the customer’s own vocabulary changes.

Medical: PCCP submission
Finance: model-risk validation
Biotech: study documentation
AI lab: model card
SaaS: AI Act Annex IV

Markets using it

Regulated FinanceSegments by sectorlead
Medical AISegments by sector
Data ObligationSegments by obligation
Defense & Safety-CriticalSegments by sectorlead

Questions

Is this a separate product?

No. ORMAS is one product under one licence, and The Diary is one of the seven things it is proven at. An institution turns on what its situation needs; there is no separate purchase, no separate integration and no separate contract.

Does this only work in my industry?

The mechanism does not know what industry it is in. What changes between them is the noun, not the method. The table above is the same mechanism with each sector's own vocabulary substituted, which is why one product serves all of them.

Where does it run, and does our data leave?

On customer infrastructure. The licence is on-premise and per-institution: nothing is sent to a third party, there is no hosted tier, and no usage metering phones home. In a hospital enclave or a controlled environment, that is a requirement rather than a preference.

Pre-book ORMAS with The Diary in scope.

Every engagement starts with one specific problem: a run that died, an audit question that had no answer, a dataset nobody can vouch for.

We reply within a few days. Every message is read by a person.

Pre-book a deployment

One licence, no price list. Terms are agreed per institution and written into the contract, and pre-booking fixes them ahead of general availability.

Useful to bring

What is being trained, and roughly at what scale
The obligation driving it: SR 26-2, the EU AI Act, Article 17, a PCCP, an internal policy
What a failed run currently costs to diagnose

Pre-book a deployment