DORA asked 4,867 people and published no result
DORA's 2025 instrument carries four questions about whether a team can change its own system without waiting for another team. It put them to 4,867 people and published no result for any of them. GitHub reports repositories using stacked pull requests merging 9 per cent more code, and the AI Security Institute released the tool it needed once transcripts outgrew what a person can read.
How we Organise
The question was asked of 4,867 people and the answer was never published
DORA’s 2025 instrument carries a four-item construct it calls Loosely coupled teams. One item asks whether a team can make large-scale changes to its own system without depending on other teams. The survey ran from 13 June to 21 July 2025 and 4,867 people answered it. The report publishes no result for any of those items, and the phrase appears once in the whole document, in a footnote pointing at DORA’s own capability page. What the report does publish is that 17 per cent of respondents sit in a cluster it names Constrained by process.
We ask a team to own one capability, and to draw the line where it would need another team’s permission to go further.
The team owns exactly one capability, bounded as what it can take end to end without another team’s consent.
The check we run against that reads a register. A register records that somebody drew the boundary. DORA’s four items ask the team whether it can move inside one, and those are different questions by design. The only population where the second was put has gone unreported. Our test has therefore never been held against the thing it stands for, and a team can pass it and still lose a week waiting on somebody else. DORA disclaims causality and its sample is snowball plus a purchased panel, so the published half settles nothing either. If you run a boundary like this and have a way to tell whether it holds, we would like to hear it at transform@dromologue.ai.
How we Build
Nine per cent more code merged once the change arrived in pieces
GitHub made stacked pull requests generally available. The release landed on 6 October. A stack splits one large change into smaller pull requests. Each is reviewed on its own, and they merge together. Since the public preview GitHub says repositories using stacks have merged 9 per cent more code than their peers, and that over two thirds of the top 1 per cent of repositories now use them.
Our position is that work should arrive in units somebody can describe in a sentence.
Each unit of work is scoped to one judgement, statable in one sentence.
This is the first figure we have seen put against that. It is GitHub’s own, and it compares adopters with peers over a period it does not state. The top 1 per cent of repositories were working differently before the feature existed, so part of the 9 per cent may be who adopted and not what adopting did. The number counts merged code and says nothing about whether review got better or worse.
How we Assure
A transcript can now run to billions of tokens, and nobody is reading it
The UK AI Security Institute released Transect on 7 October. It is an open-source package that puts a whole agent evaluation on one timeline, indexed by turns, so an evaluator can decide what to examine closely. AISI gives the reason without hedging it: as agents work more autonomously on longer tasks and coordinate with each other, transcripts can run to billions of tokens, which is more than human experts can review in full.
We have held for some time that judgement belongs at the level of the population.
The unit of judgement is the population of acts, not the act.
AISI got there by running out of capacity to read them. The tool is a release and not a study, so there is no measure of what an evaluator catches with it that they would otherwise miss, and AISI says reliable conclusions still rest on the evaluation design and on expert judgement. The aggregate chooses where to look and does not deliver the verdict.