Everything that makes the architecture work is released on publication and stays released: the mechanism, the telemetry, the correction system, the paper and all 383 experiments behind it. The manuscript is under peer review; the release follows acceptance and the terms are already fixed. From that day a researcher, a student, a model-risk validator or a regulator can read it, run it, reproduce every result and publish about it, permanently and without asking us. That commitment is why any of this is checkable rather than merely asserted.
Read the full licence text — a working draft, published now so a university or an enterprise legal team can clear it before the release lands rather than after.
Running it in production is licensed. The distinction is deliberate and it is not hidden anywhere in a footnote: research and teaching are free, and an institution putting this into a regulated deployment takes a licence.
Published code has no counterparty. You cannot put it in a filing, you cannot ask it to stand behind a declaration, and there is nobody to hold to it.
A licence gives you the thing the repository structurally cannot: a named party who calibrated the bound, validated it against your domain, and signed it. In markets that buy on auditability, that is what is actually being purchased.
The rows are in order. The free column runs out partway down, and where it stops is the product.
The training architecture itself
The mechanism, in full. Run it, modify it, ship what you build with it.
The telemetry channels
All five layers, emitted natively, exactly as the paper describes them.
The correction mechanism and its bounds
Eight diagnoses, their triggers, and the per-diagnosis ceilings.
The paper and every experiment behind it
383 controlled runs, each regenerating from seed, with the reproducibility checklist.
Teaching, publishing and independent evaluation
Use it in a course, cite it in a paper, or evaluate our claims against it. No permission needed, and no notification to us.
Submission-grade record generation
The record in the shape a reviewer expects to receive, rather than a research log.
The diff engine
Two records compared line by line. The only question a reviewer actually asks.
Cryptographic signing and a frozen reader
A hash over a canonical body, a detached signature, and a reader binary you keep.
The bound, calibrated for your domain
Derived against your data and your risk profile, then validated. Not a default.
Integration into your stack
Bindings for your training infrastructure, your federation, your imaging pipeline.
Version-pinned reproducible builds
Hash-pinned, with the means to verify the binary you run is the one validated.
Support against an agreed response time
With the person who built the architecture, not a tier-one queue.
Indemnity
Contractual liability, which is the thing an open licence structurally cannot offer.
A counterparty who signs the bound
Somebody who will stand behind the declaration under questioning.
Free for research, teaching and evaluation. Permanent.
The architecture, the telemetry, the correction mechanism and every experiment behind them, released in full on publication. The manuscript is under peer review and the release follows acceptance — a commitment written into the licence rather than an intention. From that day a researcher, a student, a model-risk validator or a regulator can run it, reproduce our results, teach from it and publish about it without asking us and without paying us. Evaluation access is available before then under agreement. Production use sits under the deployment licence.
Purpose: the mechanism can be checked.
Free for research. Licensed for regulated commercial deployment.
The record generator, the diff engine, the signing and versioning machinery, and the templates mapped onto the frameworks these markets file under. Readable by anyone, because a format nobody can inspect is a format nobody adopts. Converts to a permissive licence on a date fixed in the licence text.
Purpose: the format, not the fee.
Annual, per institution, on your infrastructure.
Integration into your environment, the bound calibrated and validated for your domain, reproducible builds, support, and indemnity. This is the licence, and essentially all revenue is here.
Where the commercial relationship lives.
Priced per engagement.
Calibrating a bound for a domain nobody has calibrated before, drafting a change-control plan, authoring the validation report, and appearing to defend it. Low volume, high value, and the reason a deployment licence renews.
The layer that cannot be copied at all.
A rate card would price the software. What an institution is buying here is the removal of a specific, quantified risk, and that is worth a different amount to a consortium than to a bank. The figure is set per institution and written into the contract. Four things determine it.
What the problem currently costs you, which is the only anchor that means anything.
Which applications you turn on, and how much of your model portfolio is in scope.
Whether the bound for your domain already exists or has to be calibrated and validated from nothing.
What you need us to sign, and therefore what we are carrying.
Annual, per institution, on-premise. No per-seat component and no consumption meter.
01
On termination you keep a perpetual, irrevocable right to read, retain and reproduce every record generated during the term, together with a frozen reader binary.
Why Your obligations outlive our commercial relationship. A customer who suspects they could lose their audit trail by not renewing will never sign in the first place, so this clause costs us nothing and makes the contract signable.
02
The declared bound and modification set are fixed when the licence executes. They change only by a documented, mutually signed amendment that generates its own record.
Why If we can quietly alter the bound, the bound is worth nothing and every filing that references it is void. Our own product goes under change control, and we volunteer that rather than waiting to be asked.
03
You receive hash-pinned builds and the means to verify that the binary running in your environment is the exact artefact that was validated.
Why Technical documentation requirements and software provenance expectations both point here. It costs us a build pipeline and it removes an entire class of question from a security review.
Every commercial deployment is on-premise or inside your own cloud environment. There is no hosted tier and there will not be one. Model weights and training telemetry leaving a controlled network is a disqualifying finding in a hospital enclave, a bank, or a classified environment, and those are the markets this exists for.
The practical consequence is that you hold the artefact. Nothing checks in, nothing reports usage, and nothing stops working if we do. That is a constraint we designed around rather than a feature we chose, and it shapes every commercial term above.
No per-seat pricing
The buyer is an institution carrying an obligation, not a team with users. Charging per seat would price the wrong thing and cap the deal at headcount.
No metering that phones home
In a hospital enclave or a bank network, outbound telemetry is a security review we would not pass — and should not.
No free tier touching regulated data
Unbounded support exposure and real liability, in exchange for a lead. The research release is free; a free production tier is not the same thing.
No relicensing of the research core
Publishing and then closing after adoption would destroy the credibility every commercial claim here rests on. The research release is permanent once made, and that is a commitment written into the licence rather than a default.
No delivery-date commitments
We will tell you what is built, what is being built, and what is not started. We will not put a date on the third category.
Because a claim nobody can check is worth nothing to the people who buy this. A validator, an assurance lab or a regulator can take the published release, run all 383 experiments, and confirm every number on this site independently. A closed reimplementation of a published method is worth less to that audience, not more — which is why publishing is a commercial decision rather than a generous one.
Research, teaching, benchmarking and independent evaluation are free and always will be. Production is licensed. If you are a university group, a student, or a validator checking our claims, take it and go — you owe us nothing and you need no permission. If it is going into something your institution operates, that is a licence, and the licence is where the calibrated bound, the reproducible builds, the support and the indemnity live.
No, and the licence text says so. Free research and teaching access is permanent. Publishing a method, waiting for academic adoption and then withdrawing it is a move that works exactly once, and it would take the credibility of every commercial claim we make with it. The published release is the reason the rest is believable, so it is not something we would trade.
When the review period on the paper closes. Peer review sets that timing, not us, so there is no date on this page. If you want the mechanism before then, ask — we send the preprint and the supplementary material to serious evaluators on request.
No, and the licence says so explicitly. You keep a perpetual, irrevocable right to read, retain and reproduce every record generated during the term, plus a frozen reader binary you keep permanently. Your regulatory obligations outlive our commercial relationship and your access to your own audit trail has to outlive it too.
Nothing. No telemetry, no usage meter, no licence server, no call-out of any kind. In a hospital enclave, a bank network or a classified environment, anything that reaches outward fails the security review before anyone opens the commercial terms — so the product is built for the environment it is sold into rather than for our convenience.
Two documents, and you get both before you engage rather than after. The research licence, which permits reading, running, reproducing and teaching, and reserves production use. And the deployment licence, an ordinary commercial agreement carrying the three clauses set out above. Neither is a copyleft licence and neither imposes any obligation on code you write alongside it.
Scope is set with your counsel, because a consortium and a bank are not carrying the same exposure and a single template would serve neither well. What does not vary: the bound we declare is one we will defend under questioning, and we do not sign declarations we are not prepared to stand behind.
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.
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