By Team Faylo1 min read
Tokens replace names and numbers — but the same value always gets the same token. That’s why the model keeps context and reasoning quality without ever knowing who it is.
“Pseudonymisation” sounds technical, but the idea is simple: replace each identifiable value with a neutral label. “Jansen Bouw B.V.” becomes, say, [[ORG_4f9c2a]], and “P. van den Berg” becomes [[PERSON_9a3d]]. The word “deterministic” is where the quality comes from.
Deterministic means the same value always gets the same token within a session. If “Jansen Bouw B.V.” appears ten times in a file, the model sees exactly [[ORG_4f9c2a]] ten times. So the model still knows it is the same company throughout — the relationships in the text stay intact, only the identity behind them is gone.
Contrast that with random replacement, where every mention would get a different label. The model would lose the coherence and reason less well. Determinism prevents exactly that.
Not everything is tokenised. Non-identifying figures — revenue, salary, balances, percentages — pass through unchanged. That is precisely what good analysis needs. The model computes and reasons on the real numbers; only who is behind them has been replaced by a token.
The value-to-token mapping lives in a vault inside your own tenant. When the model’s answer returns — still tokenised — Faylo resolves each token locally, so your team reads a clean answer about the real client. The provider never saw the original values, and no one outside your organisation can reverse a token.
The net effect: you get the reasoning power of a leading model, on real files, without identifiable data ever leaving your environment.
