Skip to content

Bias guard & confidence

Peer review is useful only if it resists style bias and accidental self-voting. Online escalation stays manual until a real dissent metric exists.

Hiding provider names isn’t enough. Models recognize their own style. The peer review stage therefore enforces:

  • Randomized order per reviewer: no positional bias toward the first/last answer.
  • Uniform output format (style normalization): strips stylistic tells that leak which seat wrote what.
  • No self-voting: a seat may not rank its own answer.
review_bias_guard:
randomize_order: true
forbid_self_voting: true
normalize_style: true

Online escalation is, for now, user_opt_in only. Automatic low_local_confidence routing is deliberately not enabled yet:

Auto-routing on “low confidence” risks sending data to the network unintentionally. It will only ship once a concrete dissent measure (a council-internal disagreement metric) is defined.

This is one of the open questions. Until then, every online step goes through the explicit privacy guard.