The quota ran out and the code review stopped
A spend limit working exactly as designed stopped a code review, a missing intrusion monitor at one Swedish supplier reached most of a country's municipalities, and inside an agent harness working out what broke was worth twice as much as guessing where.
How we Organise
The quota ran out and the review did not run
GitHub changed how Copilot code review is paid for on 8 October, and a request now bills by default to the requesting member’s own Copilot entitlement. If that member’s quota is exhausted, the review fails. Organisation owners can now bill the repository’s organisation instead, which needs paid AI Credits enabled and takes an optional budget. A second setting stops anyone outside the organisation using a personal licence to ask for a review.
We tell a finance team that a spend limit counts only when it runs without anyone deciding to enforce it, and when it has already stopped something.
The spend limit is enforced without anyone intervening, and the enforcement has been exercised.
The member-billed default meets both halves of that exactly. The thing it stops is a quality check. A limit can satisfy us completely and still be pointed at the wrong target. GitHub has now shipped the setting that moves it. If you read an exhausted quota failing loudly as the correct behaviour, tell us why: transform@dromologue.ai.
How we Build
The step that worked out what broke was worth twice the step that chose where to look
Four researchers at Illinois and Penn State paired each file, configuration and test in a repository with its own resident model. Every component then answers for its own code. It also tells the files that call it what it needs. Across four benchmarks the arrangement beat matched harnesses by 12.4 points. Spending a larger model only on diagnosing the failure took one benchmark from 27.6 to 46.1 per cent. Spending it only on choosing which components to touch reached 36.7.
Our position is that verification is what limits this work.
Verification, not generation, is the constraint.
The two earlier sources we hold on that both measured human teams shipping more pull requests. This one measures the inside of an agent. It puts the same gap at roughly two to one on the same budget. It is a preprint. Every figure in it is a benchmark score rather than a production outcome.
How we Assure
One missing monitor at one supplier reached 2.2 million people
Sweden’s data protection authority fined Miljödata i Karlskrona SEK 1.8 million for breaching Article 32. The decision was taken on 22 September and published on 1 October. An attack in August 2025 took personal identity numbers, contact details and records of sick leave and school incidents. The attacker published them on the darknet. The regulator named two failings: no adequate check when new software was installed, and no automated monitoring to detect an intrusion. The supplier’s customers include most of Sweden’s municipalities. The breach reached 2.2 million people.
We ask a board to size what one failure could reach across everything the firm depends on.
The enterprise is the unit of exposure, not the system.
IMY has since opened separate reviews of two of those municipalities and one of the regions. The authority is not treating the supplier’s own system as the boundary either. The full reasoning is published only in Swedish.