Episteme Advisory
Epistemic validation for the AI that matters
Post-cognition infrastructure for organisations that need to trust their AI outputs, not merely comply with a rule.
Risk does not wait for the rule. The EU simplification package has moved the obligations for high-risk systems to 2 December 2027. The obligation has moved; the problem has not. Unvalidated outputs are entering decisions today, and the ability to say why an output can be trusted is not built in the six months before a deadline.
Why the delay is a reason to start, not to wait (essay in Italian)
The problem
AI systems produce assertions without establishing what kind of assertion they are producing. This is not a lack of power: a model can falsify a causal premise, tell a correlation from a mechanism, and recognise that a normative judgement depends on the frame in which it is made. What it does not do is decide, before answering, what kind of thing it is about to assert. And the internal certification of that step cannot be checked from outside.
What is needed is a post-cognition infrastructure: a validation protocol specified independently of the model and applied to outputs after generation, which reconstructs from the output itself the ontological commitments the system does not represent. The second pass can be run by a machine; the protocol that governs it is external. It is not the model watching itself, and that is why it works. Human-in-the-loop alone is not enough: a human reviewer without a protocol telling them which question to ask first inherits exactly the same category error.
Services
Epistemic audit
A structured assessment of the quality of AI outputs in your decision processes. A map of cognitive risks and dependencies.
Validation infrastructure design
Design of the post-cognition cycle: a structured second AI pass plus a protocol for qualified human intervention.
EU AI Act compliance
Documentation of validation processes for high-risk systems: acceptance criteria, audit trails, records of human intervention. It serves compliance when it arrives, and the defensibility of decisions in the meantime.
Why us
AI vendors will tell you their models are reliable. The large consultancies will sell you generic frameworks. We start from the theory of knowledge (semiotics, Bayesian epistemology, philosophy of language) applied to concrete validation problems. We do not sell technology; we build the method for assessing the technology you already use.
Who it is for
General Counsel, Chief Risk Officers and compliance leads in organisations that use AI systems in critical decision processes. Cultural foundations facing the AI transition with institutional responsibility.
The thinking behind Episteme
Some of the theoretical groundwork for this project, in narrative form, is in the series La condizione epistemica (in Italian): essays on how reality resists our conceptual schemes, one on the difference between an LLM and an embodied mind, one on the crisis of epistemic authority in contemporary politics.
The lab
The method Episteme applies comes from independent research, conducted in my own name and published in full in the Lab: taxonomy, validation data, protocols and open problems, with downloadable artefacts. The Lab is not Episteme’s shop window. It is the bench where the method is put to the test, including the times it does not hold.
Go to the LabIf the problem is yours and you do not know whom to ask, write to me: the first conversation is to understand whether there is work to be done.
Let’s talkEpisteme Advisory is being built and is looking for skills and capital: who we are looking for (in Italian).