Issues

• The cost multiple is not comparable across lineups, and the <=8x guardrail is missed by 2-3x

  known-limitationmeasured-finding
  JoaquinDG/quorum#5 · Filter by author JoaquinDG opened on Aug 16, 2026

• Seven of twenty rubrics cannot be made deliberation-neutral, and that may be a property of judgement questions

  known-limitation
  JoaquinDG/quorum#4 · Filter by author JoaquinDG opened on Aug 16, 2026

• A provider outage flatters the council, because a council degrades and a single model does not

  harness-integritymeasured-finding
  JoaquinDG/quorum#3 · Filter by author JoaquinDG opened on Aug 16, 2026

• Every completion budget tuned before reasoning models fails toward a flattering answer

  harness-integritymeasured-finding
  JoaquinDG/quorum#2 · Filter by author JoaquinDG opened on Aug 16, 2026

• The blinding leaks by 9.7 points, and the leak is not evenly spread

  known-limitationmeasured-finding
  JoaquinDG/quorum#1 · Filter by author JoaquinDG opened on Aug 16, 2026