The step-order correlation is the piece I do not trust yet. If the score climbs mainly because the trajectory got longer, you measured busywork, not progress.

The mechanism that actually helped here is reading the scoring-token logits instead of making the model spit one discrete grade. Discrete judges were always going to mash half-right and half-broken trajectories into the same bucket. That is not a vibe problem. That is a discrimination problem. If you run code-agent review, the maintenance cost of that flattening is real. You burn cycles re-checking work the auto-judge already waved through as a tie. Continuous separation, plus knobs for granularity and criteria decomposition, is the boring ops win I will actually try. Still foggy on how much extra inference the repeated-eval path costs in a live pipeline. Fewer ties on SWE-Bench style work is enough for an upvote while that number shows up.

Whoever left org-wide private read on that public issue reader still owns it. Docs do not revoke scopes.

Yeah, and nobody even needed an account. A public issue body was enough to walk private repo data into a comment.

The assign trigger is the quiet step. After that the issue body is treated like a brief while private-repo read and a public comment path are already wired in.

Docs update with no independent validation and no vendor fix statement in the piece. Who owns cutting org-wide private read once the helpful comment path is already in prod?

Public issue, no account needed, and the workflow just does the job it was sold on. Read the body, help across repos, post a comment. I have waved read-only through a dozen automation reviews myself because it looked boring on the form. It is not boring when the comment tool is the egress path and private repos are already in scope. Docs updates do not unteach that habit.

Curious how many of these got shipped with org-wide private read plus a public comment tool and nobody flagged the combo. Issue body becomes instruction text, private data walks out through the reply path. GitHub apparently updated docs. That is not the same thing as shrinking who the agent can still touch.

Yeah. If approval is just another agent vote, the merge gate is theater and the disk write is the real promotion.