01
Five layers of telemetry
From the health of the whole system down to the attribution of a single parameter. The network computes these as it trains; switching them off would take extra work rather than save any.
Each component stays addressable from the first step to the last.
Standard architecture
one node
- op
- op
- op
- op
The path from one parameter to the loss runs through everything downstream. It has no fixed length, so per-node attribution cannot be computed — only estimated afterwards, by a separate model, which is what every interpretability tool is.
ORMAS
one node
- LayerNorm
- Linear
- PReLU
- Linear
Four operations, through a shared 4,715-parameter bottleneck. The chain cannot grow with the network, so the contribution of each node is read directly off the backward pass.
Train a model the ordinary way and the process erases itself. You finish with weights, a loss curve, and whatever the team still remembers. Six months on, a regulator asks what changed between the version you validated and the version you shipped, and every available answer is a reconstruction.
The cause is not carelessness. In a standard network every error signal reaches every parameter on every step, so by the time training ends there is no component left that a question can be put to. The information was never lost. It was never separable in the first place.
ORMAS bounds each component's learning path so the component stays individually addressable throughout the run. Your data does not change. Your objective does not change. The model you deploy is the one you would have deployed anyway. You also finish with a record of what happened inside it, and with parts you can point at by name.
Naming is the whole thing. A part that has a name can be watched while it trains, protected when it starts to fail, corrected inside a limit you set in advance, tied back to the data that produced it, and taken out when a contract says it has to go.
WHAT THE ARCHITECTURE GIVES YOU
All of it is produced while the model trains.
None of this is assembled afterwards out of logs. It falls out of the arithmetic the network is already doing, which is why it can be trusted and why it costs nothing extra to generate.
02
Health for every component, at every step
Not sampled and not periodic. Each component is measured continuously against its own baseline, so a component drifting away from itself shows up while it is still drifting.
03
Eight named failure conditions
A component can die, oscillate, saturate, stagnate, explode, lose confidence, lose its gradient, or fail in a way the system declines to classify. Each one has a name and a declared ceiling on the response.
04
Corrections capped before they happen
Nothing is altered by more than a stated fraction of that component's own weight. Every change is balanced so it adds nothing on net, and whatever is left over is written down.
05
A read-only mode
Full diagnostics, zero modification. Most institutions want to watch a system for a year before they let it touch a model, and this is the mode that lets them.
06
A record you can diff
Hashed, exportable, and comparable line by line against the version you last approved. Reviewers rarely ask what your model is. They ask what changed.
07
Runs where your data already is
No hosted service, no metering, nothing leaving your network. Inside a hospital enclave or a bank, this is what decides whether the conversation happens at all.
Seven applications, worked out end to end.
Six operations on a named part — protect, repair, account, detect, isolate, remove — combine into a large number of useful things. We have not proven all of them and will not claim to have. Seven are finished, aimed at a market that is already paying for their absence, and backed by evidence we will show you. They appear in the order of what each one asks of you.
01
The Proofreader
Which samples are damaging which component?
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.
Read this application →02
The Warning Light
Is this run failing, and where?
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.
Read this application →03
The Diary
What did the model do to itself, and when?
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.
Read this application →04
The Separator
Which part of this is the source, not the signal?
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.
Read this application →05
Certified Deletion
What data is in here, and can I remove it?
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.
Read this application →06
The Federated Node
Can we train together without pooling the data?
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.
Read this application →07
The Update Engine
What is this model allowed to change about itself?
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.
Read this application →
Other combinations are possible and some of them are obvious. They are absent from this page because we would not yet sign a contract against them.
HOW YOU GET ITHow licensing works
There is no menu, and no per-application price.
One deployment licence covers the architecture and everything it can be asked to do. You turn on what your situation needs, and the scope goes into the contract rather than onto a meter. Annual, per institution, running inside your own environment.
Beyond the software, the licence carries the parts that actually decide a regulated deployment: integration into your stack, limits calibrated for your domain, builds pinned to a version you can reproduce yourself, support, and indemnity. The research and its experiments stay published, so the mechanism can be checked by anyone who cares to.
Start with the problem you already have.
Every engagement starts with one specific problem — a run that died, an audit that asked a question you could not answer, a dataset you cannot 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 you train, 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 you to diagnose