The better workflow took twice as long and cost four times as much
Thoughtworks scored a specification-led development workflow against a direct baseline and published both halves: 2.75 points more correctness, double the time, four times the tokens. A Korean water utility reorganised around AI and found its own budget review was the constraint, and Dropbox audits consent on more than two hundred of its own pages by visiting them.
How we Organise
A utility put seven disciplines in one division and still waits two years for a project
Korea Water Resources Corporation created an AI Division in late 2025. It staffed the division from seven areas, among them administration, telecommunications and civil, mechanical, electrical and environmental engineering. Its first head is Ha-Dong Jung. He says a project there takes six months to a year to prepare, about a year to implement and six months more to close out. The models change throughout. He names securing staff and passing project and budget review as the obstacle he has not solved.
We tell a client that a team’s boundary only becomes real when it can get an idea into production without asking anybody.
A boundary is real only when a team can go from idea to running software without asking anyone’s permission.
The division owns the whole chain from dam to tap. It still cannot move at the speed of the thing it is adopting. The permission it waits on is budgetary, not technical. The account is one officer’s, and the two-year figure is his own rather than a measured distribution.
How we Build
The specification-led workflow scored 2.75 points higher and took twice as long
Thoughtworks built SpecBench to run a configured development workflow against defined tasks in four scenarios, assess the output independently and record what the run consumed. On its published Spec-to-Code figures a direct baseline reached a mean 92.25 per cent behavioural correctness. A workflow adding specification, planning, decomposition, implementation and convergence stages reached 95.00. That second workflow took 163.93 minutes against 83.74, used 1,781,930 tokens against 450,250, and saw its trajectory score fall from 60.52 to 48.79.
Our position is that a client measures a change to a prompt before it goes live, and keeps the evidence that produced the number.
Every revision to a prompt or a skill is scored against the suite before it is adopted, the score carries the sample behind it, and where a model does the scoring it is not the model under revision. Its agreement with a person’s reading is measured in the period by a chance-corrected statistic, reported with its reliability and its position bias.
Thoughtworks did that. The figure it got back endorses the workflow that nearly doubled the time. The hard part is that the two numbers point opposite ways, and the practice says nothing about which one decides. This is not the first source to bear on it. A study of twenty-one judges published in September cut the same practice on its scoring clause, and that reading stands unanswered. Thoughtworks states that each cell is a single evaluation group and that many frontier models are uncovered, so whether the extra correctness is worth the cost on any particular codebase may not be settled by four scenarios. If your arithmetic came out the other way, write to transform@dromologue.ai.
How we Assure
Dropbox opens its own pages as a visitor to see whether the banner held
Yasmin McDowell and Lawrence Good describe an auditor that opens a Dropbox page in a fresh browser session. It records what loaded before any interaction, declines non-essential cookies, reloads and checks the preference survived. It runs three times for every page, as a United States visitor, a European Union visitor and a Global Privacy Control signal. It covers more than two hundred surfaces and 22 languages. It acts on the consent elements and never on button text. The classifications sit outside the auditor’s source, so the privacy team can change them without waiting for a release.
We ask that a disclosure be verified on the deployed system rather than in its configuration.
Every surface a person meets discloses the machine, that person can reach a human with the authority to change the outcome, and before they rely on it they can read what the system is for, what it was tested against and what it must not do.
Dropbox reaches that rule on its own. It calls testing what actually happens one of the project’s most important decisions. The account covers consent and stops there. It says nothing about the statement’s other two clauses. It gives no count of what the auditor found, and nothing measures what it misses.