Filum ships a pre-drafting assistant. It runs on your Anthropic key, it is disabled with one checkbox, and with no key configured the product works in full. Above all: nothing that constitutes evidence goes through it — traceability, signatures, the chained audit trail, the inspection archive and the 11 statement quality checks are deterministic and run offline. That is what makes them qualifiable, and 46 cases do it.
Model called today: claude-opus-5. That name is written into every retained draft, and it is never rewritten retroactively.
This answers a data protection officer's question: what leaves, when, and where does it go.
Each names the place where you can see it for yourself. A guarantee you cannot verify is not a guarantee.
No text produced by a model enters a record without a named person having accepted it, and that acceptance carries their name, its date and its entry in the audit trail.
An electronic signature engages a person's accountability (21 CFR 11.50). Text slipped into an approved document without passing through somebody would make that signature untruthful.
↳ Document screen: the draft is displayed separately, with its “accept” and “reject” buttons.
The system instruction forbids inventing a value, a result or a reference, requires quoting each source's exact code in brackets, and requires writing “to be completed” where information is missing.
An invented figure has no source, and it is the source that holds a design file together. A gap filled by a guess is worse than a gap: it can no longer be seen.
↳ The instruction is in the delivered code, readable as is. We will show it to you on request.
It is explicitly forbidden to conclude on the conformity or non-conformity of anything. It states what the data shows; the decision belongs to the signatory.
A regulatory verdict binds the organisation before an authority. It is not delegated to a tool, least of all a tool whose output varies from one call to the next.
↳ Same system instruction. A draft containing a verdict can be rejected like any other.
Filum ships no API key. It is YOUR Anthropic key that is used, stored in your local database, never in our code and never in our repository. With no key configured, no call goes out — and the product works in full.
A vendor key would route your design data through our account, under our terms, at our discretion. No design file should depend on a supplier's commercial relationship with a third party.
↳ Settings → AI assistant. The key's source is displayed there: your organisation, the server environment, or none.
An administrator disables the assistant for the whole installation. No call goes out then, even if a key remains configured.
Some organisations forbid any external call from systems holding design data. They must be able to enforce it, not ask their users for it.
↳ Settings → AI assistant → “enable” checkbox.
Traceability, signatures, content freezing, the chained audit trail, the inspection archive, the statement quality checks: none of them calls a model. They are deterministic, they run offline, and they are qualified.
You cannot qualify what does not return the same result twice. Everything carrying regulatory proof is therefore, by construction, outside the model's scope.
↳ The validation file: every case runs without an API key, and the whole suite runs offline.
The part brochures leave out. An auditor asking for the contract would find the gap in one question — better we write it ourselves.
Filum calls the API with YOUR key. The applicable terms — retention, use for training, processing location — are those of YOUR contract with Anthropic, not ours. We cannot commit on a third party's behalf, and a promise we do not hold would serve you nothing before an auditor. Read that contract, and if your policy requires it, leave the assistant disabled: the product works in full without it.
A language model returns plausible text, not true text. The instructions reduce the risk of invention; they do not remove it. That is exactly why nothing enters a record without human reading, and why the draft retains the list of sources it received: so that checking is possible, and quick.
Two identical calls do not necessarily return the same text. That is a property of the model, not a flaw in Filum — and it is why no qualified function depends on it. What is retained, by contrast, is perfectly reproducible: the accepted draft is frozen, and its digest enters the record like any other content.
We do not validate Anthropic, and nobody can in the ISO 13485 sense. What we validate is the SOFTWARE AROUND IT: the locks above are within the qualification scope, the model is not. If an auditor asks you for the AI's validation, that is the honest answer — and it holds, because nothing that constitutes proof goes through the model.
Yes, and that is the default state: with no key configured, no call goes out. Every regulatory function — traceability, signatures, audit trail, archive, statement checks — works identically.
No. No path in the product lets a model write into the graph. The requirement rewrite is displayed next to the field; you are the one who copies it back.
Each retained draft carries its origin — the model's name — the list of sources it was given, the person who accepted or rejected it, and the date. Whatever did not go through that was typed by somebody.
No. A draft keeps the name of the model that WROTE it, even after an update. Recomputing that field on display would say what we use today instead of what wrote the text — and a record is not rewritten.
A thirty-minute demonstration, on fictional data. We open the system instruction's code if you ask.