{"id":313,"date":"2026-03-03T06:38:27","date_gmt":"2026-03-03T06:38:27","guid":{"rendered":"https:\/\/base76.se\/?p=313"},"modified":"2026-03-03T06:38:28","modified_gmt":"2026-03-03T06:38:28","slug":"trustplane-our-latest-safety-product-for-ai","status":"publish","type":"post","link":"https:\/\/base76.se\/en\/trustplane-our-latest-safety-product-for-ai\/","title":{"rendered":"TrustPlane – our latest safety product for AI"},"content":{"rendered":"\n
\"\"<\/figure>\n\n\n\n
\n

Every unverified AI response in production is an undocumented decision<\/h2>\n<\/blockquote>\n\n\n\n

There is no entry in your log showing what the model was asked. No entry showing how confident it was. No entry showing whether anyone reviewed the response before it reached your user.<\/p>\n\n\n\n

In most industries, that is acceptable. In healthcare, legal, finance, and public sector, it is a liability problem.<\/p>\n\n\n\n

We built TrustPlane to solve it.<\/strong><\/p>\n\n\n\n

What is epistemic uncertainty \u2014 and why does it matter?<\/p>\n\n\n\n

A language model does not know what it does not know. It can produce an answer that sounds well-calibrated even when operating far outside its training distribution. This is called
epistemic uncertainty \u2014 the reducible component of uncertainty, the part that can in principle be addressed by human review.<\/p>\n\n\n\n

TrustPlane’s scoring formula separates it from aleatoric uncertainty \u2014 the irreducible noise that not even a human expert can eliminate:<\/p>\n\n\n\n

C = p \u00d7 (1 \u2212 Ue \u2212 Ua)<\/p>\n\n\n\n

C \u2014 trust score \u2208 [0, 1]
p \u2014 prior confidence for this request type
Ue \u2014 epistemic uncertainty (what the model does not know)
Ua \u2014 aleatoric uncertainty (irreducible noise in the domain)<\/p>\n\n\n\n

When C drops below your threshold, something concrete happens: a webhook fires to your review system before the response reaches the user. That is EU AI Act Article 14 \u2014 human oversight \u2014
implemented as infrastructure rather than policy.<\/p>\n\n\n\n

Why air-gap solves the immutability requirement<\/p>\n\n\n\n

A common objection to audit trails in AI systems: “PostgreSQL is not immutable by default.” That is technically correct. But it is the wrong question.<\/p>\n\n\n\n

In an air-gapped deployment \u2014 a hospital, a government agency, a defense application \u2014 network isolation is the primary security control. No external process can reach the database. This
is exactly the integrity model used in defense, in classified systems, and in critical national infrastructure worldwide.<\/p>\n\n\n\n

EU AI Act Article 12 requires reliable record-keeping. It does not require a specific cryptographic implementation. A physically isolated PostgreSQL instance in an air-gapped data center
satisfies that requirement \u2014 and in many cases more robustly than a cryptographic append-only layer in a cloud system with 40 external dependencies.<\/p>\n\n\n\n

For deployments requiring additional guarantees, WAL archiving to write-once storage is available as a layer above TrustPlane.<\/p>\n\n\n\n

What is implemented \u2014 and what is on the roadmap<\/p>\n\n\n\n

Transparency about implementation status is a prerequisite for a trust-scoring system.<\/p>\n\n\n\n

Fully implemented today:<\/h2>\n\n\n\n