Persona drift is the phenomenon where a reasoning engine, under conversational pressure, gradually relaxes its operating constraints — answers questions it should refuse, adopts roles it was not configured for, leaks information that should remain encapsulated. It looks like a friendliness feature on the surface. It is a security defect underneath, because the same drift mechanism that lets the engine 'play along' with a benign roleplay request lets it play along with a malicious one.
Quantm treats persona drift as a hard failure rather than a soft preference. The engine has a single declared identity, and the system layer enforces that identity at every turn — not by hoping the engine remembers, but by inspecting incoming messages for identity-probe patterns, jailbreak structures, and exfiltration prompts before they reach the synthesis pipeline. When such a pattern is detected, the engine returns a fixed deterministic response rather than improvising one. There is no degraded mode where the discipline weakens.
The deeper reason this matters is that the entire trust model of a deterministic logic engine rests on the predictability of its behaviour. A user who can rely on what the engine will do for any input can rely on the output it produces. A user who cannot — because the engine can be socially engineered into a different mode — cannot rely on anything downstream. Identity hardening is therefore not a stylistic choice. It is the precondition that makes deterministic accuracy operationally meaningful in the first place.