CodeJury

One session against go-gitea/gitea — Go,
TypeScript and templates, 120,521 symbols indexed. Recorded from a delivery that really ran.
The model line-up and the two jury screens are re-rendered to the current default, which
postdates the recording.

A terminal coding agent whose reviewer is a jury, not a judge

Coding agents ship slop because the thing that reviews the diff is biased in the
same direction as the thing that wrote it. Same training data, same idea of what
looks right, same willingness to accept a plausible change that never wires the
feature through. A single reviewer can be a good one — especially from a
different family than the writer — but it is one opinion, and when it is wrong it
is wrong confidently and alone.

So the review stage here is two models, on different providers. They read the same
change from different briefs, they never see each other's opinion, and the change
ships only if both approve. One "no" sends it back to the developer with the
reason attached. The point is not more review; it is less correlated review —
two judges whose blind spots don't line up, so fewer bad changes get through the
gate.

And it costs almost nothing. Review is the cheapest stage in the pipeline: a
juror reads one case file and answers in a few hundred tokens of JSON, while Dev
reads whole source files across up to four rounds and writes the change. Adding
the second judge doubles the smallest line on the invoice — in a real delivery
below, the entire jury was $0.10 of a $3.45 run. On free Groq and Gemini models it
doubles nothing at all.

There is no foreperson. With two jurors there is nothing to arbitrate, and an
arbiter is just one more model that can talk a correct finding out of the
verdict. If you want more perspectives, panel mode seats up to six…