A black box that helps decide on care, finances or personnel does not earn my trust. I want to see how that decision is made. Not as a bonus. As a starting point.
Most AI platforms work exactly like that. You send data to an API, you get an answer back, and what happens in between stays hidden. That feels acceptable as long as it is a chatbot drafting an email. The moment AI helps decide on a care indication, a credit application or a job applicant, it is a different story. Then you want to know why.
The difference between a black box and a glass box
A black box gives you an answer. A glass box shows which data was used, which rules were followed and why the system made that particular choice. For a human-impacting decision, that difference is not cosmetic. It is the difference between trust and hope.
Why open source sits in every layer
At UPPR, open source is not a marketing tick. It sits in every layer of the system. Hermes as the agent runtime, so we can read what an agent intends before it acts. pgvector as the knowledge base, so we can see exactly which fragments were retrieved. n8n for workflow orchestration, so every step is visible and adjustable. Locally running models when privacy demands it.
That does not mean open source is easier. It is more work. No off-the-shelf dashboards, no vendor that solves everything for you, no SLA to hide behind. But it does mean you are never dependent on a company that unilaterally decides what you can see, what it costs and when it disappears.
What the AI Act actually asks
For high-risk AI, the European AI Act asks for logging, traceability, human oversight and explainability. With a closed API that is technically hard and legally risky. With open components it is not an extra project but a natural consequence of how the system is built. Audits become something you can simply show, not something you dread.
Transparency as foundation, not feature
Transparency is not a switch you flip. It is the way the entire system is put together. That is why organisations working with us can have their own IT, their own security and their own regulator look along. Not because they have to. Because it is the right thing to do.
What "open" really means here
Open source AI isn't a marketing label. It's a specific set of properties that, taken together, let you actually inspect what you're running:
- Open weights. The trained model is downloadable and you can run it without asking permission.
- Open architecture. The model design is documented so independent researchers can reason about it.
- Open evaluation. Benchmarks and limitations are published, not hidden behind a sales call.
- A licence that allows business use. Llama, Mistral and Qwen all allow commercial use within clear bounds.
Not every "open" model ticks every box. We pick models that score well across all four, because that is what makes audit, replacement and long-term continuity possible.
Why this matters for high-stakes work
When AI helps decide on a care indication, a credit application or a job application, "we trust the vendor" is not an answer. Three concrete things change when the model is open:
- You can audit. A team — yours, ours, or an external auditor — can trace why the model leans the way it does on a class of cases.
- You can swap. If a better open model comes out next quarter, you change one config line. No new contract, no new DPO review, no migration project.
- You can prove non-discrimination. Regulators are starting to ask not just whether your AI is "explainable" but whether the underlying model is inspectable. Closed weights can't pass that test.
The cost story people don't tell you
Closed frontier models charge per token, and the bill scales linearly with usage. For an organisation running an Agentic OS at meaningful volume, that bill grows fast. Open models, hosted on your own infra or in our EU cloud, have a fixed cost: the box runs whether you use it at 10% or 80%. Past a modest volume threshold, open is cheaper and more sovereign.
The other cost nobody mentions is portability. With a closed vendor, you are pricing in switching cost from day one — you just don't see it on the invoice. With open weights, that switching cost is close to zero.
"But isn't the frontier closed?"
Today, yes — the absolute frontier still sits with closed labs. But two observations matter. First, the gap between frontier-closed and best-open has been shrinking every quarter for two years. Second, almost no business automation needs the absolute frontier; it needs reliable drafting, classifying, summarising and tool use, and open models have been good enough at those tasks for a while.
For the small share of work where the frontier really matters, you can call out to a closed model from inside the OS, with explicit consent and logging. The default stays open and inspectable; the exception is conscious.
Closing
Transparency isn't a feature we add for show. It's the only honest answer to the question what is this AI actually doing inside my organisation? If you can't answer that, you don't really run the AI — it runs you.