The editorial council, one finding at a time

Perpetūra’s /council skill convenes an editorial panel to review a story. I described the shape of it in the voice post; here is one finding traced all the way through, because the loop matters more than the diagram.

The five lenses

A review runs five lenses, each its own agent: romance (does it land for the audience), craft (literary quality, house tics), architecture (route design and choice hygiene), canon (continuity and spoiler discipline), and series (consistency across stories). I build each lens an evidence pack first — quotes with file-and-line references — so the agents argue from the text, not from vibes. Then an external bench of non-Anthropic models weighs in, so the panel isn’t one model agreeing with itself.

One finding, end to end

During Reine’s review, the craft lens flagged two chapter closings. Its note, lightly trimmed:

The two session-added closings state the route VERDICT outright. This is the exact “states its own morals” tell the revision plan was written to subtract. A closing may carry FEELING and ACTION but never the VERDICT.

That’s the finding. Here’s the loop it went through:

The panel debated it — the romance lens partly disagreed, arguing one of the two closings earned its directness. That disagreement was preserved, not resolved, because the council reports and the author decides.

An adversarial verify pass then took the surviving finding and tried to kill it against the actual prose: is the closing really stating a verdict, or just an emotion that reads as one? It held up on one closing, fell on the other.

I adjudicated and rewrote the one that held. The council’s own record tags that fix with the commit that applied it, so the review doc is also a changelog.

Why bother, for a solo project

Because the failure mode of writing with an AI is uniform competence — prose that is fine everywhere and alive nowhere. A single model reviewing its own output smooths the same way it wrote. Five lenses with different priorities, an external bench, and a skeptic pass produce disagreement, and disagreement is where the real notes live. A machine caught me moralizing my own ending. I’d have defended that closing to a human editor. I couldn’t defend it to five of them at once.