v2026.8.6

What's Changed

Fixes

• The council no longer treats agreement as a result. Providers are handed a
  description of your problem and never the system itself, so none of them can
  test a premise your question asserts — they reason from it correctly and agree
  with each other while doing it. Unanimity therefore measures how clearly the
  question was framed, not whether the framing was true, and the more confident
  it reads the better it hides that nobody could check. Two places said the
  opposite. The synthesis was told to keep it short when every provider
  agreed, which shortened the answer exactly where the premise had been examined
  least; it now says they agreed and then names the assumption the whole answer
  rests on. Agent mode (--agents) called high-confidence agreement a stronger
  signal, which holds for reasoning and not at all for premises.
• Questions are checked before they are sent. The council now separates what
  you have observed from what you are assuming: claims it can verify locally it
  verifies, correcting the question when one is wrong, and the rest are labelled
  and passed through as stated. A provider told "I have not checked whether X
  holds" can answer "then check X first". One told "X holds" never will.

Docs

• The README's worked example only ever showed a divergent synthesis, which is
  the case the council always handled well. It now also shows the unanimous one.
  New section on the observed-versus-assumed labelling, since you will see your
  own question relabelled before it goes out.

Reported by a peer session whose council returned a unanimous, competent and
wrong answer: five providers produced a genuinely good design, and one agent
that could read the disk then measured its premise away — of 4792 logged
failures, the effect the question assumed was present in about 2 percent.

Full Changelog: v2026.8.5...v2026.8.6