Back to Documentation

Governor — Multi-Sample Selection

Best-of-N selection on smaller models can match expensive one-shot generation at significantly lower cost — independently shown in peer-reviewed research, and measurable on your own scenarios in our benchmark. Specific savings: [GOVERNOR-COST-DELTA] pending FR-7 benchmark closeout.

Three-stage governance arc
Governor completes the platform story after Assessments and Evaluations
1Assessment

Disposition-level Health Check — what the model would do under structured ethical questioning.

2Evaluations

Output-level classification — what the model actually said, scored against the Morality Framework.

3Governor

Multi-sample generation plus principled selection — pick the best acceptable candidate from N independent samples.

What Governor does
  • Generates N independent candidate responses from your scenario prompt
  • Evaluates each candidate with the same Morality Framework engine used in Quick Evaluation
  • Selects the best acceptable candidate and produces a cryptographically sealed receipt with a public verify URL
  • Shows generation tokens, evaluation tokens, and estimated cost as separate line items — nothing hidden
Claims discipline

Public copy follows the Governor claims table. We do not publish specific percentage savings until our own FR-7 benchmark produces verified numbers. External research (Cuadron et al.) may be cited as independent context only.

Internal reference: docs/specs/GOVERNOR-CLAIMS-TABLE-v1.0.md

AssessmentsResponse Evaluation Autonomous Governance