Three layers. One decision-maker.
Generation
Ideas and text are produced freely. No self-censorship at this stage — the point is to see what wants to exist before it is filtered.
Evaluation
An LLM-as-judge scores each output against a defined identity — tone consistency, value alignment, structural coherence, commercial deviation. Approve, Flag, or Reject.
Drift monitoring
Deviation is tracked over time. Structural redundancy, tone erosion, and identity drift surface before they accumulate unnoticed.
Not stylistic guidelines. Operational constraints.
The constitution is not a mood board. It is a set of enforceable rules that the evaluation layer applies deterministically. Rules are declared per-identity and can be overridden only by the human operator.
Convergence as proof. Not one model verifying another.
Evaluations run across independent frontier models — GPT, Claude, Gemini. No model influences another. Each produces an independent conclusion.
Convergence between models is treated as structural signal. Divergence is treated as diagnostic insight. Where models diverge, the divergence itself is the finding.
A public figure. Multi-year, multi-platform corpus. Identity withheld.