EU AI Act · Article 12
High-risk AI systems must keep automatic, immutable records of events across the system lifecycle. ORMAS emits that record as a physical property of training rather than as a logging layer bolted alongside it.
ORMAS · STRUCTURAL INTELLIGENCE FOR NEURAL NETWORKS
A neural network that names its own broken part.
A training architecture that records what each component of the model did, while it learns — rather than a monitoring layer bolted on afterwards.
Deployed on your own infrastructure, in
A run dies at 60 percent. Today —
The loss curve says the model degraded around step 40,000. It cannot say which component caused it, because nothing computed that. So the run restarts from the last good checkpoint, blind, and the same compute is paid for twice. On a frontier run that is thousands of GPU-hours to arrive back where you already were.
The same run, trained under ORMAS —
The network names the component, the step it failed at, the size of the change, and the bound it stayed inside. Not an estimate produced afterwards by a second tool — a quantity the network computed about itself while training, which costs nothing to read back.
The same record answers the auditor asking which data shaped a decision, and the regulator asking what changed between two versions. One mechanism, three questions.
WHAT IT COSTS YOU TODAY
Nobody budgets for this. Everyone pays it.
Every error signal in a neural network touches every parameter at once. That is what makes these systems powerful, and it is why which part did this? has no available answer — not because the answer is buried, but because there is no part to point at. The bill arrives in a different currency in every market.
MarketWhere it breaksWhat it costs
AI TrainingA run degrades silently. The loss curve moves only once the damage is thousands of steps old, so recovery means discarding everything since the last checkpoint.$16,000–$24,000per two-hour recovery
Medical AIAn audit asks what the training changed. No record of it exists, because no architecture in clinical use produces one — so the model goes back.12–36 monthsadded to market entry
Regulated FinanceA model that cannot be audited component by component cannot be validated under SR 26-2. A model that is not validated is not deployed.The programmewritten off in full
Data ObligationA data licence terminates. The clause covers anything trained on it, and retraining from scratch is the only remedy that satisfies the clause.$100,000–$500,000per deletion request
Defense & Safety-CriticalA system modifies itself in the field and cannot account for it afterwards. That is not a finding a programme remediates.The deploymentended, not remediated
One property is missing from every architecture on that list, and every figure above is a consequence of its absence. ORMAS is the correction.
Everyone else reconstructs what happened. This records it.
When something goes wrong inside a model, every tool available today works the same way: it examines a finished system and produces its best estimate of what must have occurred. The estimate is often good. It cannot be verified, and it cannot be filed.
What follows is not a better estimate. It is a different kind of answer — one written while the work is happening, by the system doing the work.
The question you have to answerHow it is answered todayWith ORMAS
What failed, and when?
A performance chart. It moves only after the damage is done, and it cannot describe anything smaller than the whole run.
The component, the step it happened on, and how severe it was.
Is that a finding, or an opinion?
An estimate, produced afterwards by a separate tool that can itself be wrong.
A record, written at the moment of the change, by the system that made it.
Which data caused the damage?
Current methods cannot tell bad data apart from genuinely difficult data, so they discard both — and the difficult examples were the valuable half.
The two are told apart, and neither is discarded.
Can one source be removed later?
Only by retraining the model from the beginning. Six figures, every time it is asked.
What the model built from that source is a named, removable part. Remove it, and issue a certificate of what was removed.
What does accountability cost to run?
A second model, trained and maintained alongside the first.
Nothing. You deploy one model, the same size and speed as the one you would have deployed anyway.
Will an auditor accept it?
A written account of what the team intended to do.
Evidence of what the training actually did, signed and unalterable.
What is being compared
SHAP · LIME · Integrated Gradients
Explain why a finished model made one particular prediction. A good answer to a different question — they were never built to watch a model being trained, and do not claim to be.
Weights & Biases · MLflow
Excellent at recording what happened to a training run. They cannot record what happened inside the model.
Co-teaching · DivideMix
Run a second model to get a second opinion, then use it to decide which data to throw away.
Retraining from scratch
The only method that genuinely removes a data source from a model. Six figures, per request, and at scale it is not done at all.
None of this is a report produced for you afterwards. It is written by the system as it works — which is the whole difference between something you are asked to believe and something you can check.
ORMAS is a training architecture. What you buy is what it leaves behind.
Nothing about how you train changes: the same data, the same objective, the same model at the end of it. What changes is what exists alongside that model — parts that can be named, and therefore protected, repaired, accounted for, separated by source, and removed on request.
What you receive
01
The model
Deploys exactly as it would have. Same architecture, same inference cost, nothing about serving it changes.
02
The named parts
What the model learned from each source, as a separate object you can point at — bound to a record of the data that justified it, and removable.
03
The record
Every change the training made: which part, when, why, how much, and inside what limit. Signed, and diffable against the version you last approved.
Where it lands hardest
Those six operations can be combined in many ways. Seven combinations carry almost all of the value, because each one lands on a failure a specific market is already paying for — the same failures priced at the top of this page. They are set out in the order of what each asks of you: the first needs a file and never touches your model; the last needs a limit agreed with your reviewer before you begin. Other combinations exist. These are the seven worth naming.
The use caseWhat it doesWhat it asks of you
Attaches to a model you have already trained, without modifying it, and returns a ranked list of the labels it believes are wrong — kept separate from the ones that are merely difficult. Broken down by class, and by source where your data carries it.
- DETECT
A model file and a sample of your labels
Watches every component of a model while it trains and raises an alarm when one crosses its own baseline — naming the component, not just the run. The operator rolls back hundreds of steps instead of thousands.
- ACCOUNT
One line around your training loop
03
The DiaryWhat did the model do to itself, and when?
Regulated Finance · Medical AI · Data Obligation · Defense & Safety-Critical
Produces a signed, timestamped record of every modification made during training: the component, the diagnosis, the treatment, the step, and the limit it stayed inside. Diffable against the last version you approved.
- ACCOUNT
- PROTECT
- REPAIR
Training on the architecture
Separates what a model learned about the world from what it learned about where the data came from. You ship one model that works anywhere, a named removable part for each source, and a record of which data justified each one.
- ACCOUNT
- DETECT
- ISOLATE
- SUPPRESS
Training on the architecture, with source labels
05
Certified DeletionWhat data is in here, and can I remove it?
Regulated Finance · Medical AI · Data Obligation
Removes a named source's contribution and issues a signed certificate of exactly what was removed, with a bounded statement of what changed. No retraining.
- ACCOUNT
- DETECT
- ISOLATE
- SUPPRESS
The Separator, plus your deletion policy
06
The Federated NodeCan we train together without pooling the data?
Regulated Finance · Medical AI · Data Obligation
Runs inside a federation you already have. Each party's contribution is recorded before the averaging step destroys it — so when a round drops six points, you can say which site.
- ACCOUNT
- DETECT
- ISOLATE
- SUPPRESS
A federation that already exists
07
The Update EngineWhat is this model allowed to change about itself?
Regulated Finance · Medical AI · Defense & Safety-Critical
Retrains a deployed model inside a limit declared and signed beforehand, and produces the evidence that every change stayed inside it. Built to the shape of a predetermined change control plan.
- ACCOUNT
- PROTECT
- REPAIR
A limit agreed with your reviewer before you begin
One product, one licence, on your own infrastructure. You turn on what your situation calls for and the licence is scoped to that — one integration, one contract, and nothing priced per use case.
See the product in fullTHE EVIDENCE
Three results. Every condition stated.
Each figure comes from a controlled experiment against a parameter-matched standard network: same parameter count, same data, same schedule.
Read the full evidence80.3%
Recovery from total structural collapse
A layer was destroyed mid-training, on a network that had reached 85.1%. The standard network stayed at 10% — chance level — on every seed, permanently. ORMAS diagnosed the failure within one epoch and climbed back to 80.3%.
94.6%
Retained through a shift in the task
Trained on one task, then another, with no replay buffer and no task identifier — the conditions under which a deployed model is normally retrained. The standard network kept 47.3% of what it knew.
22,014
Stable where standard training failed outright
A fifty-node network at 30% label noise across 200 epochs. Standard training returned NaN and stopped. ORMAS stayed stable, and every one of those corrections is on the record with its component, magnitude and limit.
383 controlled experiments across four architecture families. All results on CIFAR-10 and CIFAR-100 — the paper carries the full conditions, the error bars, and the one result that went against us.
The mechanism is the same. The vocabulary is yours.
A hospital calls it the site. A bank calls it the region. A biotech calls it the batch. An AI lab calls it the corpus. Every market below asks the same question in its own words, and each is answered by one of the seven use cases.
AI Training
SCALEIs this run failing, and where?
The Warning Light gives you that signal — per component, every step, while the run is still recoverable.
Regulated Finance
INDUSTRYWhat changed, and can a validator verify it?
Attribution per component means validation per component. The record answers SR 26-2's validation and change-control expectations with evidence rather than with a memo.
Medical AI
INDUSTRYWhich site caused this, and can we show it?
Each site's contribution is recorded before the averaging step destroys it. When a round drops six points, the record says which site, and on what evidence.
Data Obligation
OBLIGATIONWhat data is in here, and can I remove it?
Certified Deletion routes the memorisable contribution into a named, removable structure, and hands the requester a signed certificate.
Defense & Safety-Critical
INDUSTRYWhat did the system do to itself in the field?
Degrading to 80.3% rather than collapsing permanently to 10% is a different safety category — and every modification the system made to itself is bounded, timestamped and reviewable afterwards.
The record is becoming an obligation, not a preference.
Three separate regimes now require of a trained model something standard architectures cannot produce: evidence of what changed inside it, attributable to a part. The institutions that can generate that record are not merely compliant — they are setting the format everyone else will be measured against.
SR 26-2 · SS1/23
The interagency guidance that replaced SR 11-7 in April 2026 keeps the disciplines that matter: model inventory, independent validation, and documented change control. A model that cannot be audited component by component cannot be validated, and a model that is not validated is not deployed.
GDPR · Article 17
The right to erasure is being enforced against trained models, not only against databases. The compliance industry has already named the category it needs — unlearning-ready architectures — ahead of any supply existing.
Pre-booking is the window in which the product is still shaped around you.
There is no price list. The figure is agreed per institution and written into the contract, and pre-booking fixes those terms ahead of general availability. There is no limit on how many are accepted and no countdown attached to it — the reason to do it early is that the requirements are still open, not that the slots are scarce.
Early partners get flat-rate, enterprise-wide terms, their own compliance workflow built into the baseline architecture, and direct engineering access rather than an account manager.
Pre-book a deploymentORMAS: Neural Architectural Transparency Enables Autonomous Self-Correction
- Experiments
- 383
- Status
- Preprint under review