OC-301b · Module 1

Voice Differentiation Testing

3 min read

If you strip the agent name from an output and the reader cannot tell which agent wrote it, the voice differentiation has failed. Voice differentiation is measurable — you can test it by presenting unlabeled outputs to evaluators and asking them to identify the agent. If identification accuracy is below 70%, the voices are not distinct enough.

The differentiation test protocol: take three agents with different DISC profiles. Give each agent the same input — a market analysis brief. Collect the three outputs. Remove agent names and identifying metadata. Present the outputs to three evaluators who know the agents' personas. Ask each evaluator to match each output to the correct agent with a confidence score. If all evaluators match correctly with high confidence, the voices are differentiated. If any evaluator matches incorrectly, analyze why — which voice markers were missing or which agents sounded too similar.

Do This

  • Test voice differentiation with blind identification — remove agent names and ask evaluators to match
  • Define 3-5 unique voice markers per agent that no other agent uses — signature phrases, structural patterns, vocabulary choices
  • Test differentiation under stress — rushed outputs and edge cases reveal voice collapse

Avoid This

  • Assume different system prompts produce different voices — they often converge to similar patterns
  • Rely on agent names to carry differentiation — the voice should be identifiable without the label
  • Test only with the agent's primary task — test with shared tasks where differentiation is hardest