How we work.
Anyone can call the same model. Nobody else holds what each rollout builds. Six choices make that true.
01 Position
The harness, not the model.
A language model is one component. What an organisation gets from Inora depends on the harness around it: how it builds context, consults sources, calls tools and checks its own answer.
- Why
The same model scores very differently on the same tasks depending on the harness. In a benchmark of May 2026, the aggregate score on 106 tasks ran from 52.4 to 76.2 across six harnesses and eight models. It is a preprint with caveats, and the same sources reject the strong form, that the model does not matter. Read the note.
- What it costs
We cannot swap a model and be done: we test every model and every change on the organisation’s own questions. For tasks in ordinary language the effect of the harness is smallest, so the measurement for our kind of work is still ours to make.
- What we did
On 18 October 2025 we rebuilt Inora from the ground up on open building blocks, so that no platform decides which models and capabilities we can use. Read the piece.
- How to test it
Ask a vendor to run the same questions on two models in one harness, and on one model in two harnesses. Ask us the same.
02 Position
The maker implements.
We implement Inora ourselves. What shows on the floor comes back into the product, and that only works when whoever builds it also puts it to work.
- Why
Whether a system works shows only on the floor. A builder who does not implement hears late what goes wrong.
- What it costs
Every implementation costs our own time and our own engineers. We grow more slowly than a company that leaves implementation to partners.
- What we did
In 2026 we decided that our own engineers carry out every implementation and that we use no implementation partner.
- How to test it
Ask who sits at your table in the first weeks, and whether they are the same people who build the product.
03 Position
Where a wrong answer counts.
We work where a wrong answer has consequences: care first, then other regulated professions such as pharmacy, social work, compliance and public administration. What aviation and spaceflight have learned about error, we take over where it is proven, and no further.
- Why
Aviation and spaceflight have worked for decades on one problem: how an error stops spreading. Four mechanisms are well documented: checklists run by a fixed method, confidential reporting without punishment for the reporter, observation of normal flights, and treating failure as organisational.
- What it costs
Evidence that this works in care is largely missing. The one tested transfer to nursing, a ‘do not interrupt’ vest during medication rounds, had no effect on administration errors. So we claim a design discipline, never aviation results.
- What we did
On 9 October 2026 we published our note on what aviation can teach care, with the source for each mechanism and what remains unproven. Read the note.
- How to test it
Check every source in the note. Ask us which mechanism Inora applies and how we would show it.
04 Position
Any model, anywhere.
Inora is tied to no single model or environment. It can run anywhere, on an organisation’s own servers included, with the language model the organisation chooses.
- Why
Models change faster than organisations decide. A product tied to one model inherits its price, its limits and its direction.
- What it costs
Staying free takes work. We test every model on the organisation’s own questions before staff work with it, and we avoid capabilities only one vendor can supply.
- What we did
On 18 October 2025 we rebuilt Inora because the platform the first version ran on decided which models we could use, and when. Read the piece.
- How to test it
Ask for the setup you need, in your own cloud or on your own servers, with the model you choose, and ask what changed to make it possible.
05 Position
The professional decides.
Inora supports professionals with their organisation’s knowledge. It does not diagnose or decide on treatment; the professional decides. Its intended use keeps it outside medical-device scope.
- Why
A language model does not reason: it picks the most likely next text and does not know what it does not know. So the model writes, the sources decide what is true, code does what must be exact, and the professional decides.
- What it costs
Some users want the system to take the decision for them. We do not build that. Where a question holds a decision about a client, Inora should say who decides it.
- What we did
On 14 September 2026 we withdrew the word ‘reasoning’, which we had used for Inora. Read the piece.
- How to test it
Ask any supplier of AI for professionals for its intended use in writing, and for who decides what happens to a client. Ask us the same.
06 Position
The knowledge stays yours.
An organisation’s knowledge belongs to that organisation. What it puts into Inora stays its own and trains no model.
- Why
What compounds over time is the knowledge. An organisation’s protocols, work agreements and handbooks are its property and its responsibility, and they stay so.
- What it costs
We do not improve our models with what an organisation entrusts to us. That is a deliberate choice, and it makes learning from use harder for us.
- What we did
On 4 July 2025 every organisation in Inora got a closed space of its own, with its documents in their own structure and with their own authority. Read the piece.
- How to test it
Ask what happens to your knowledge if you stop: in what format you take it with you, what is deleted, and whether it trains anything. Ask us the same.