Drift Harness

Every AI system drifts. It softens a position under pressure, drops a constraint it was holding a moment ago, or states a guess as if it were certain. The correction is almost always manual: a human notices, pushes back, forgets, and corrects the same failure on the next turn. Nothing remembers, and nothing scales. Drift Harness makes that loop automatic — it intercepts the exchange before the user has to act, logs what failed and why, and builds a structured record that drives correction at scale. The system fans a single exchange across thirteen specialist agents, each checking one slice of behaviour: constraints, antipatterns, voice, quality, identity, alignment, gap analysis, profiling and question generation. Every agent returns the same five-field verdict — agent, status, rule, excerpt, severity — so one shape holds across every layer. Its core idea is how it represents certainty. Rather than a percentage, which is just a token prediction dressed up as a probability, each agent commits to one of three states — violation, uncertain, or clean — always tied to an exact excerpt from the reasoning that triggered it. The excerpt is what makes the label mean anything. Under the hood, the logger mints a UUID4 per exchange and classifies each turn in Python before the model runs. Findings write to a FastAPI and SQLite backend; agents communicate over a shared Band session; a C++ coordinator handles multithreaded fan-out. The full stack runs live on a Hetzner VPS under pm2, with a dashboard at dashboard.malecsystems.com. We proved it end to end: one misaligned input fanned across every live agent produced six confirmed findings, written straight to the backend. All thirteen agents are deployed and the dashboard is live. The harness is the asset. The agents are the mechanism that fills it. Every AI system drifts — this one notices, records it, and turns a manual habit into infrastructure.