The government's new playbook for testing policy on vulnerable people says it has tried not to be prescriptive. The one place it is, keeping those same people safe, it never says who is checking.
Estimated reading time: 7 minutes
In short. The Cabinet Office published its Test, Learn and Grow Playbook on 19 October 2026, a ten-page guide to piloting policy interventions that explicitly says "we've tried not to be prescriptive" [1]. Across all ten pages, there are eight sentences using "must" rather than "should" or "could". Every one of them is in the one section on research ethics and participant safeguarding, and not one names a panel, a gateway or a person outside the project team who checks any of it happened [2]. The playbook's own worked examples run live tests on domestic abuse cases and violence against women and girls in Islington, and on children's social care in Barnsley [3]. Separately, the playbook tells teams that "standard cost-benefit analyses assume an idea will work from the start, creating a false sense of certainty" [4], a description that does not match the mandatory uncertainty and optimism bias rules in the Green Book, the Treasury guidance those same teams are supposed to use. Two documented mistakes: four stars.
Picked from this week's new registry entries: a Cabinet Office playbook published yesterday, verified live on GOV.UK with no withdrawal notice, scoring on two criteria, a claim that does not hold against a document already torn down on this site and guidance with must-statements but no named enforcement or review clause of its own. This is a Tier 2 structured teardown, the document's own words tested against its own cited evidence and against the Green Book's.
A playbook that says it is trying not to tell you what to do
The Test, Learn and Grow (TLG) programme is the Cabinet Office's effort to spread an agile, pilot-first way of designing public services across government and local authorities. Its own introduction frames the whole exercise as deliberately loose: "Your local context will look different to ours. We've tried not to be prescriptive. However, learning and iterating are fundamental to the approach" [1]. The playbook is written for two audiences: "practitioners who are ambitious for change, and need practical support to achieve it" and "leaders who have the power to enable or get in the way of the test and learn approach" [1].
Nine of its ten published pages hold to that promise. Agile ceremonies, team structures, evidence-gathering, theories of change, decision criteria for whether to scale or stop an intervention: all of it is written as "should", "could" or "you may", the advisory register the document itself favours. Read all ten pages together and the word "must" appears eight times. Every single instance sits inside, or points back to, one page: "Conducting research safely" [2].
The claims, tested
| The playbook's claim | What we found | Verdict |
|---|---|---|
| "You must create a protocol that sets out what to do if a participant discloses abuse, harm or neglect." "You must complete an ethics assessment showing your research is justified and your plan has clear objectives." "If a participant is under 16 years old, you must get parental consent." [2] | The only accountability named for any of these musts is internal and self-selected: "Get in touch with your safeguarding team and ask them to review your protocol" and "if you cannot protect your participants, find a better method for your research" [2]. No page names a central panel, a sign-off gate, or any stated consequence for a team that skips the ethics assessment and runs the research anyway. | The one place this playbook turns mandatory is the one place it names nobody outside the project team who checks the mandate was met |
| "Standard cost-benefit analyses assume an idea will work from the start, creating a false sense of certainty. The bigger financial risk is spending years of effort and money on an idea that fails in the end." [4] | The Green Book, HM Treasury's guidance for appraising government proposals, states that "practitioners must consider risk and uncertainty as part of their balanced judgement on value for money" (para 6.61, p.41) and that "practitioners must account for optimism bias by making explicit adjustments at the outset of an appraisal" (para 6.79, p.44) [5]. It names "pilot studies" directly as one of the ways to reduce that uncertainty before relying on the numbers (para 6.63, p.41) [5]. | The standard method this playbook describes is not the standard the government's own mandatory appraisal guidance, used by the same practitioners, actually sets out |
Must, with nobody named to check
The "Conducting research safely" page carries real weight because of who the playbook expects to use it on. Its own examples are not abstract. A worked case study describes the TLG team running a workshop with domestic abuse survivors in Islington "to develop test ideas that would help tackle violence against women and girls", using a tree-shaped exercise to break down each idea with the group [3]. A second describes the same team working with Islington Council's Children's Services data analysts "to understand how domestic abuse cases move through Children's Services" [3]. A third describes a project with Barnsley Council testing "whether council data could track how children move through services" in the children's social care system [3]. These are the populations the safeguarding page exists to protect.
That page is explicit about the stakes: "For some projects, you'll need to research sensitive topics or work with people in vulnerable situations. This research can cause distress if you do not plan carefully. You must know how to protect both participants and researchers" [2]. It sets out real procedure: a safeguarding protocol for disclosures, an ethics assessment, a data protection impact assessment for "special category data" such as health status or sexual orientation, parental consent below the age of 16 [2]. None of it is wrong as a checklist.
What it lacks is a mechanism. "Ask your safeguarding team to arrange training for anyone who needs it" [2]: which safeguarding team, reporting to whom, is left to the reader's own organisation. "You must complete an ethics assessment showing your research is justified" [2]: assessed by whom, filed where, checked by whom before a project proceeds, the page does not say. "If you cannot protect your participants, find a better method for your research" [2] is an instruction to the same team that designed the research to mark its own homework. Nine pages of this playbook are consciously non-prescriptive because the document trusts local teams to adapt the approach to their context. The tenth page issues the only hard rules in the whole guide, for the work with the highest potential for harm, and extends exactly the same trust.
Standard cost-benefit analysis, misdescribed
The second claim sits in a different register: not about process, but about how the playbook justifies itself against the government's existing rulebook for spending decisions. Making the case for funding a test and learn project, the playbook tells teams: "Standard cost-benefit analyses assume an idea will work from the start, creating a false sense of certainty. The bigger financial risk is spending years of effort and money on an idea that fails in the end" [4].
The Green Book is the document that standard applies to. Its 2026 edition is explicit that appraisal is not supposed to assume success: "Practitioners must consider risk and uncertainty as part of their balanced judgement on value for money" (para 6.61, p.41) [5]. It goes further, naming the exact failure mode the TLG playbook describes, over-optimism, as a rule rather than a risk: "Optimism bias is the demonstrated systematic tendency for practitioners to be over-optimistic about key assumptions in appraisal, such as social costs, social benefits or project duration... Practitioners must account for optimism bias by making explicit adjustments at the outset of an appraisal" (para 6.79, p.44) [5]. And it names the TLG playbook's own method as part of the standard fix: "This uncertainty can be reduced by research, evaluations of previous interventions and pilot studies" (para 6.63, p.41) [5].
None of that means the Green Book's safeguards always work; this site found in September that HS2's own optimism bias adjustment was made exactly as the rulebook requires and still proved short by tens of billions of pounds. But the claim on trial here is narrower: that the standard method "assumes an idea will work from the start". The government's own mandatory standard does not say that about itself. It requires the opposite, in terms at least as strong as anything in the new playbook.
Credit where due
The playbook's citations elsewhere hold up. It twice points readers to "Chapter 5" and "Chapter 6" of the "Magenta Book Test and Learn Annex" for prioritising assumptions and choosing testing methods [6]: that annex exists, under that name, published alongside the Magenta Book, and its Chapter Five does set out a prioritisation framework for exactly the purpose described [7]. The safeguarding page itself is substantively careful where it matters most: it tells researchers to let participants skip questions, to allow pauses, to work in pairs with an observer when sessions involve people in vulnerable situations, and to keep disclosure records private and accurate rather than paraphrased [2]. And the playbook is candid about its own record: "We're proud of this work. We're also aware it has not always been perfect" [1], a rarer admission in government guidance than it should be.
Verdict
Four stars, from two documented mistakes. Neither is a claim that test and learn is the wrong way to design public services, or that the safeguarding advice on this page is bad advice. The first mistake is structural: a document that is deliberately light-touch everywhere except the one section protecting people in abuse, violence and care cases, and that, even there, names no one outside the project team to check its own must-statements were followed. The second is a citation problem dressed as a methodology claim: the playbook justifies its own novelty by describing standard practice in terms the standard itself does not use. A playbook that tells teams to "follow the evidence" should follow its own evidence about what the evidence base already requires.
The star score counts two documented mistakes: the safeguarding and research-ethics section's eight must-statements (protocol, ethics assessment, parental consent for under-16s), none of which names a sign-off gate, panel or external check anywhere in the document's ten published pages, tested against the playbook's own case studies naming domestic abuse, violence against women and girls, and children's social care as the subject matter; and the "Getting started" page's description of standard cost-benefit analysis as assuming "an idea will work from the start", tested against the Green Book's mandatory risk, uncertainty and optimism bias provisions (paras 6.61, 6.63 and 6.79). Two falls in the 1 to 3 band: four stars; the bands are on the ratings page. Checked directly against all ten pages of the Test, Learn and Grow Playbook and the Green Book's 2026 edition, fetched 20 October 2026. The document is tracked in the Daily Register; a revision will get a diff entry.
Sources
- Test, Learn and Grow Playbook: Introduction, Cabinet Office, published 19 October 2026
- Test, Learn and Grow Playbook: Conducting research safely, Cabinet Office, published 19 October 2026
- Test, Learn and Grow Playbook: Using evidence to scope test ideas (phase 2), Cabinet Office, published 19 October 2026
- Test, Learn and Grow Playbook: Getting started, Cabinet Office, published 19 October 2026
- The Green Book (2026 edition, PDF), HM Treasury
- Test, Learn and Grow Playbook: Testing, learning and iterating (phase 4), Cabinet Office, published 19 October 2026
- Magenta Book: Test and Learn Annex (PDF), HM Treasury / Cabinet Office Evaluation Task Force