METHODOLOGY

§ Arkheia · The Science

Mathematical certainty
in AI governance

Most AI evaluation asks “does this look right?” Arkheia asks “is this statistically consistent with how this model normally behaves?” The difference is measurable, defensible, and auditable.

§ How It Works

The detection pipeline

01

Characterise

Build a behavioural profile for each model family from controlled characterisation runs. Not generic thresholds — empirical data.

02

Observe

Intercept every invocation at the API boundary. Extract surface signals: token statistics, timing patterns, output structure.

03

Compare

Calculate effect sizes against the model's behavioural baseline. Cohen's d > 0.8 = actionable drift.

04

Decide

Three-zone classification: Safe, Review, Hold. Execution decisions, not confidence scores. Honest about uncertainty.

See the science
in action

Detection profiles built from empirical data. Per-model baselines. Statistical rigour you can explain to your auditor.