The evidence chain, over thirty years
Development evaluation and evidence-based medicine each spent thirty years writing down how to trust a finding. Read their rule-books side by side and they split: one loosened its grip on the chain from claim back to source record while the other tightened it, and the loosening field is the one now handing its drafting and synthesis to AI. Every quote below is checked against the source.
Two fields that spent thirty years writing down how to know something
Two professions share a problem and almost never compare notes. Development evaluation asks whether an aid programme actually did what it promised. Evidence synthesis, the discipline behind evidence-based medicine, asks whether a treatment actually works. Each one ends in a finding that someone with power will act on, a funder deciding the next grant or a health system deciding the next guideline, and each spent the last three decades writing down how that finding is allowed to be made.
The rule-makers are the OECD’s Development Assistance Committee (the DAC, the donor-government body that sets the terms for foreign aid) and the United Nations Evaluation Group (UNEG, the UN’s own network of evaluation offices). They publish the criteria that define a good evaluation and the standards that govern how it is run, and a funder’s results desk treats them the way a court treats statute.
The rule-makers are volunteer networks of methodologists. GRADE (Grading of Recommendations Assessment, Development and Evaluation) rates how sure you can be of a result; Cochrane and the PRISMA reporting group set how the underlying review is run and written up so a stranger can rebuild it.
This audit is about the first of them. Development evaluation is the field whose documents a funder holds you to, and the one now moving AI into its drafting and synthesis, so it sits under the lens. Evidence synthesis is here as a control: a standard that gets revised shows what its field chose to weigh more heavily and what it quietly stopped asking for, and the only way to tell whether a direction was forced or chosen is to watch a neighbouring field meet the same problem. The medical one met it and went the other way.
In development evaluation, the oldest document is the strictest.
The 1991 DAC Principles required a structurally independent evaluation unit, reporting past line management to the minister or the board, and an explicit attempt at causal attribution. Twenty-eight years later, the field’s flagship revision of its criteria is silent on independence and treats attribution as already handled. A discipline that began strict on the chain of custody loosened its grip on it, at the moment AI moved into the workflow.
That could be read as simply how evidence standards age, except a neighbouring field ran the same experiment and got the opposite result. Over the same span, the standards for clinical and systematic-review evidence tightened, until a 2020 reporting guideline asks you to publish the data and the analytic code so a stranger can recompute your result. The clinical canon is the control: it shows the development trajectory was chosen rather than inevitable, because a neighbouring field given the same decades went the other way. On the medical side the strictest document is the newest; in development evaluation, it is the oldest.
This is the Evidence Chain Read method run across time, and it is a worked companion to the single-period audit of the evaluation standards.
Two fields, opposite directions
A qualitative reading of how firmly each canon required its evidence to be traceable, independently checked, and honestly attributed, from the early 1990s to now. The vertical axis is judgement, not a metric. The shape is the argument.
The evaluation line tracks operative requirements. The field’s glossary of terms (2002, updated 2023) is left off it: its causal definitions stayed sharp throughout, so the easing is in what the standards require, while the dictionary held.
A fair reading separates two things the word “standard” hides. Some of these documents define criteria; others set process. Criteria documents were never where chain of custody lived, so the 2019 Revised Criteria reading “weak” on traceability is partly a fact about its genre. The finding controls for that and still holds: the concrete independence and causality discipline the 1991 Principles carried was not renewed, the field’s 2018 to 2019 energy went into making the criteria more flexible, and its process standard has sat unchanged since 2010. On attribution, where the comparison is definition against definition, the regression is clean.
The discipline it had early, and let go
Eight documents from the DAC and UNEG, spanning the field’s whole codified history, from the 1991 founding principles to the 2023 second edition of the glossary and the 2025 UNEG addendum. Each column is one of the four questions an evidence auditor asks; each row is a document, in order of publication.
Strongest on independence and causality at the start, thinning as the canon modernised toward flexible criteria.
| Year & document | Source traceability | Independent verification | Attribution / certainty | Definitional stability |
|---|---|---|---|---|
| 1991DAC Principles for Evaluation | weak | partial | partial | partial |
| 2002DAC Glossary of Key Terms | weak | weak | strong | strong |
| 2005UNEG Standards, UN System | partial | partial | weak | partial |
| 2010DAC Quality Standards | partial | partial | partial | partial |
| 2016UNEG Norms & Standards | weak | partial | weak | weak |
| 2019DAC Revised Criteria | weak | weak | weak | weak |
| 2023DAC Glossary, 2nd edition | weak | weak | strong | partial |
| 2025UNEG Norm 11 addendum | weak | weak | weak | weak |
The two most recent entries cut different ways, and the distinction is the point. The 2023 second edition of the DAC Glossary did not soften the definitions. It kept the rigorous attribution language (“confounding factors … or external shocks”) and added counterfactual and contribution analysis to the vocabulary, sharpening the concepts. What it also did was formalise contribution as a legitimate substitute for strict attribution, writing the retreat the 2010 standards only gestured at into the dictionary itself. The 2025 UNEG addendum, for its part, added the numbering collision a citation can no longer survive without a year. So the field’s dictionary held its causal rigor while its operative standards did not renew theirs. A discipline can define chain of custody immaculately and still decline to require it, and that gap is what the Evidence Chain Read is built to find.
“Impartiality and independence will best be achieved by separating the evaluation function from the line management … a central unit responsible for evaluation reporting directly to the minister or the agency head … or to a board of directors.”
1991 DAC Principles, §16 · verified. The oldest document holds the firmest line on independence.
“The ascription of a causal link between observed … changes and a specific intervention … taking account of other interventions, (anticipated or unanticipated) confounding factors, or external shocks.”
2002 DAC Glossary, “Attribution” · verified. Attribution as causal inference, naming confounders.
“valid attribution is embedded in all the criteria.”
2019 DAC Revised Criteria, Impact / Box 6 · verified. The same idea, demoted to a background property.
The discipline it built, and kept building
Three landmarks anchor the other side. GRADE (2008) put a certainty rating on every finding. Cochrane’s risk-of-bias tool, revised to RoB 2 in 2019, made the judgement of how reliable a study is structured and repeatable. PRISMA (2009, expanded in 2020) set what a review must report, and where its data has to live.
Certainty attached to every finding, dual-independent extraction, registered protocols, and public data and code.
| Year & document | Source traceability | Independent verification | Attribution / certainty | Definitional stability |
|---|---|---|---|---|
| 2008GRADE, rating quality of evidence | partial | partial | strong | strong |
| 2009PRISMA statement | partial | partial | partial | partial |
| 2019Cochrane RoB 2 | partial | strong | strong | strong |
| 2020PRISMA 2020 | strong | strong | strong | strong |
“High quality — Further research is very unlikely to change our confidence in the estimate of effect … Very low quality — Any estimate of effect is very uncertain.”
GRADE, 2008 · verified. A four-level certainty rating that travels with each estimate.
“an algorithm that maps responses to signalling questions to a proposed risk-of-bias judgement for each domain.”
Cochrane RoB 2, 2019 · verified. The bias judgement itself made structured and reproducible.
“Report which of the following are publicly available and where they can be found: template data collection forms; data extracted from included studies; data used for all analyses; analytic code; any other materials used in the review.”
PRISMA 2020, item 27 · verified. Publish the code, so an outsider can recompute the result.
One field made regenerability a checklist item. The other never wrote it down.
PRISMA 2020 asks a systematic review to state where its data and its analytic code are publicly available, so a reader can take the materials and recompute the result without asking anyone. That is the hard floor the Evidence Chain Read looks for: a claim regenerable from a preserved record by someone who was not in the room.
Across nine evaluation documents and thirty-four years, no equivalent requirement appears. The nearest, the 2010 Quality Standards, asks that sources be described and listed, and opens by saying it binds no one. The evaluation canon assumed the chain. The clinical canon wrote it into the form.
Same problem, two eras, opposite answers. The gap is not oversight. It is a choice about what a standard is for.
The soft field is the one automating first
The dashed line and the AI marker meet at the worst possible place. Development evaluation reached its loosest grip on the evidence chain at the exact moment models moved into the drafting, the coding, the translation, and the synthesis. Every soft word the canon leaned on, “sufficient,” “described,” “explained,” is a word a model can satisfy without touching a source record.
The clinical canon reached for the opposite instrument. It made the reviewer’s bias judgement a structured, algorithm-mapped, documented artefact, and it required the data and code to travel with the finding. Those are the controls that survive automation, because they ask for a record rather than a reassurance. The field that built them is not the field now most exposed to the thing they defend against.
Your results sit on the softer side of this chart.
If your evaluation, results framework, or evidence synthesis answers to the OECD-DAC or UNEG standards, it inherits their assumptions about the chain, and the AI in your workflow inherits the gaps. The Evidence Chain Read runs the six-seat council across your document and returns a hardening memo, judged by a human, before a funder’s results desk reads it for you.
- 1991 DAC Principles for Evaluation of Development Assistance (OCDE/GD(91)208, endorsed 3-4 December 1991).
- 2002 DAC Glossary of Key Terms in Evaluation and Results Based Management.
- 2005 UNEG Standards for Evaluation in the UN System (UNEG/FN/Standards(2005)).
- 2010 DAC Quality Standards for Development Evaluation (ISBN 978-92-64-08390-5).
- 2016 UNEG Norms and Standards for Evaluation; 2019 DAC Revised Evaluation Criteria; 2025 UNEG Norm 11 addendum.
- GRADE (BMJ, 2008); Cochrane RoB 2 (BMJ, 2019); PRISMA 2020 statement and checklist (BMJ, 2021).
Documents read as published, retrieved July 2026 from primary sources (the 1991 Principles and 2002 Glossary via the Internet Archive; clinical papers via open-access repositories). The trajectory is a qualitative reading of the operative text, not a metric, and this is an audit of the documents, not a claim about any institution’s conduct. It is not evaluation, clinical, or legal advice.