Data access
Which data the model may see. Sensitive fields are masked before it does.
Craft · Deterministic layer
Reliability is built around the model, not into it.
A deterministic verification layer wraps the stochastic model, checks, quality gates, guardrails and audit trails that don't promise reliability, security and compliance, but prove it.
§01 Craft
The quality gate proves, it doesn't trust.
Every workflow ends at a quality gate in the arc42 sense: an algorithm that checks whether the requirements you defined up front are met, not a human skimming output that merely looks convincing. The verdict is binary and reproducible: the same result passes or fails the same way every time. What doesn't clear the gate doesn't reach production.
- 01 Input & context
- 02 Model, stochastic
- 03 Checks & guardrails
- 04 Quality gate: pass or fail
- 05 Production, with audit trail
Human AI Algorithm
§02 Craft
Hard rules, not polite requests.
Guardrails are boundaries enforced in code, not a wish in the prompt: which data the model may see, which tools it may call, what never leaves the context. Sensitive fields are masked before they reach the model; forbidden actions are blocked, not logged after the fact. No prompt can talk its way past a hard rule, and that is what separates an experiment from a system in production.
Tool permissions
Which tools it may call. Everything else is blocked, not logged.
Context boundaries
What never leaves the context. Enforced in code, not in the prompt.
§03 Craft
An audit trail you can defend.
Every decision is logged: which input, which model, which rules, which result, traceable and reproducible. That trail carries an audit under FINMA, GDPR and nFADP: you can show what happened, when and why, instead of asserting it. This algorithmic gate is exactly what turns a clever demo into a system a regulated company is actually allowed to run.
FINMA
Traceable decisions and provable controls for the Swiss financial context.
GDPR
Provable, logged handling of personal data.
nFADP
Swiss data protection, designed into the architecture from day one.
You can show what happened, when and why, instead of asserting it.
§04 The problem
What happens without the layer.
Without this layer, a system doesn't fail loudly. It fails quietly, until it counts.
- 01 The same prompt returns a different answer next week. Drift
- 02 Outputs look convincing and carry silent errors. Unchecked
- 03 Sensitive data ends up in the wrong tool. Risk
- 04 No one can show what was decided, when or why. Unauditable
A deterministic layer makes every one of these visible, and blocks it before it reaches production.
§05 Read on
More on the deterministic layer
How verification, guardrails and audit trails make a stochastic model checkable, for FINMA, GDPR and nFADP, we go deeper in the blog and in the open mema whitepaper.