Keeping a Source Register for Policy Desk Research
Keep a traceable source register for policy desk research, recording versions, scope, extraction notes and the claims each source supports.
A source register is a working record of the documents a desk researcher has relied on, kept in enough detail that a colleague can open the same text, find the same passage and see where the evidence ends and interpretation begins. It differs from a bibliography in three respects: it records location and version rather than citation style; it keeps source text and researcher inference in separate fields; and it stays usable when the documents themselves are revised, withdrawn or replaced. Build it while reading, not after drafting, because the register is where a claim acquires its audit trail.
Why the register belongs to the research, not the write-up
Citation lists are written for readers. Registers are written for the next researcher — often a colleague six months later, or the same analyst returning to a brief after the file has been reopened. Two failures recur in desk research that skips this step. The first is a locator that cannot be followed: a bare URL to a landing page, a PDF that has since been replaced, a statistic quoted without a table number. The second is a conclusion that has drifted from its source, so that a sentence describing what a document says is in fact describing what the analyst inferred from it.
A register addresses both by making the distance between document and judgement visible. It also imposes a useful discipline at the point of reading: if a passage cannot be located precisely enough to log, it is not yet evidence.
The Magenta Book, HM Treasury's guidance on evaluation for UK government, includes companion guidance on quality in qualitative evaluation (Quality in Qualitative Evaluation). Its scope is the assessment of qualitative evidence quality; a register is one practical way to make that assessment inspectable for a specific piece of desk research, without treating the framework as a checklist to be completed mechanically.
What another researcher must be able to locate
A locator is not the same as a citation. The test is whether someone with the same access rights can arrive at the same sentence. That requires more fields than a reference list normally carries.
| Field | What it should hold | What goes wrong without it |
|---|---|---|
| Identifier | Short stable code used in claim records | Claims cite sources by author and year, which collide when a body publishes several papers |
| Publisher and author | Issuing body, and the named author or unit where one exists | Attribution defaults to the website, which may be a host rather than an author |
| Title and document type | Full title plus status: consultation, interim, final, technical annex, written answer | An interim finding is later treated as settled |
| Publication date and version | Date as printed on the document, plus any version number | Undated guidance gets cited as current |
| Access date | The date you read it | A revised page silently changes the basis of a finding |
| Persistent location | Direct link to the document, not the section landing page | The reader lands on an index and cannot tell which item was used |
| Passage locator | Page, paragraph, table or figure number | Quotation cannot be verified in a long report |
| Jurisdiction and scope | Which administration, period and population the document covers | A policy position in one nation is applied to another |
| Held copy | Where a saved copy sits, if one may lawfully be kept | The link dies and no substitute exists |
| Access conditions | Paywall, registration, embargo, restricted circulation | A colleague cannot tell whether absence of access explains an apparent gap |
Two conventions are worth adopting early. Keep the identifier in the form ORG-YYYY-NN so entries sort predictably and do not need renaming when a second document from the same body appears. And record the passage locator at the moment of reading: reconstructing page numbers from notes is slower than logging them once.
Separating source text from inference
The central discipline is a three-way split, held in three fields rather than one.
Locatable passage. The words the document actually contains, in quotation marks, with the passage locator. Short extracts are enough; the register is not a transcription exercise. Where the source is a table or chart, record the figure and the row or column heading.
Faithful summary. A restatement in your own words that adds nothing. Test it by asking whether the author of the document would accept it as accurate. Evaluative verbs belong elsewhere: "the report states that uptake was uneven" is a summary; "the report shows the scheme failed" is not.
Inference. Your judgement, with the reasoning that produced it, a confidence marker, and the specific step that would change your mind. Inference is legitimate and necessary in desk research; the error is leaving it unlabelled in the summary field, where it acquires the appearance of a source claim.
A useful convention is to make the inference field answer three questions in a sentence or two: what do I now think, on what basis, and how would I know if I were wrong. This also makes disagreement between sources productive rather than obstructive, because each source's contribution is recorded separately instead of being averaged into a single narrative.
Handling documents that change
Policy documents move. Consultations become government responses, statistics are revised, guidance is withdrawn, departmental pages are restructured. Three kinds of change need distinguishing, because they call for different responses.
- The document changed. A new version exists. Add a new register entry, mark the earlier one superseded, and note whether the change touches any conclusion that rests on it.
- The location changed. The content is the same but the link no longer resolves. Update the locator, keep the access date of the original reading, and record where a copy is held.
- Your reading changed. Nothing in the document moved, but later evidence altered your interpretation. Record this against the claim, not against the source, so the source record stays a record of what it said.
Where licence, confidentiality or file-handling rules permit, keep a dated copy of anything load-bearing and record its location in the register. Where they do not, record who holds the copy and under what conditions it can be seen. Never let the register become a store for restricted material that should not be circulating.
A suggested register format
A two-tab workbook is sufficient for most desk exercises, and easier to maintain than a database. Treat the layout below as a working format to adapt to the brief; it is not a validated instrument, and no commissioning body requires it in this shape.
Tab one — sources. One row per document, using the fields in the locator table above, plus a status column with a small controlled vocabulary: current, superseded, withdrawn, unverified.
Tab two — claims. One row per claim you intend to make or consider making.
| Column | Content |
|---|---|
| Claim ID | C-01, C-02 and so on |
| Claim as it might be written | Plain sentence, no hedging yet |
| Supporting source IDs | One or more identifiers from tab one |
| Contradicting source IDs | Sources that pull the other way |
| Passage locators | The specific passages relied on for this claim |
| Status | supported, limited, contested, not supported |
| Confidence and follow-up | What would strengthen or overturn it, and the next practical step |
The status column does most of the work. Where a claim is marked supported, at least one locatable passage should say what the claim says. The limited status means the evidence permits the claim only in narrower terms. Marking a claim contested is not a problem to be smoothed away; it is a finding about the state of the record, and often the most useful sentence in the brief.
The register also connects naturally to how a research decision is framed: a desk review and a conversation with a practitioner answer different questions and carry different evidential weight, a distinction set out in the related guide on written research versus expert calls.
A fictional worked example
The following is an invented illustration. The programme, documents and figures are fictional and are used only to show the format in operation.
Suppose a fictional regional authority ran the Northbridge Skills Retention Programme from 2021 to 2024, and a fictional desk researcher is asked why participants left before completing it. Four documents are found: an interim evaluation report published in March 2023; the funding body's annual report, which gives completion totals; a written answer in the fictional legislature, giving headline numbers only; and an undated provider newsletter. The records are incomplete and the sources disagree — the interim evaluation attributes early leaving mainly to travel costs, while the newsletter credits participants with moving into work.
The register extract might look like this:
| ID | Source and locator | Passage | Faithful summary | Inference | Status and follow-up |
|---|---|---|---|---|---|
| NBR-2023-01 | Interim evaluation report, March 2023, s.4.2, p.31 | "Travel costs were the reason most frequently given at exit interview." | Exit interviews most often recorded travel costs as the reason for leaving. | Suggests cost was a factor for those who were interviewed, but exit interviews reach only those who left through a recorded route. | limited. Check the technical annex for the exit-interview response rate. |
| NBR-2023-02 | Funding body annual report, 2023, table 6 | Completion figure for the programme year | Reports completion totals for the year; gives no reasons. | Supports a claim about the level of non-completion, not its causes. | supported for the level only. |
| NBR-2024-03 | Provider newsletter, undated | "Many of our learners moved straight into work." | States that participants moved into work; no figures, no date, promotional context. | Weak corroboration at best; may describe a subset. | unverified. Search for provider monitoring returns. |
The claim entry that follows reads: the record supports a statement about the scale of early leaving; it does not establish a cause. Status: limited. Follow-up: locate the technical annex, check whether a final evaluation was published after the interim, and seek any published monitoring return that separates leaving routes. A brief written from this register would say what the numbers show, note the competing explanations, and state that the causal question remains open. That is a defensible conclusion, and a more useful one than a tidier answer the evidence does not carry.
Where the register stops
A register documents the basis of a judgement; it does not validate the judgement. No fixed number of sources proves a proposition, and convergence across several documents is not the same as statistical representativeness — a point that matters when qualitative sources are added together and presented as if they described a population. Completeness of the register is not completeness of the evidence, and a well-kept register can create false confidence if the limited and contested columns are under-used.
The format also assumes a purpose. Where documents are confidential, where access is uneven across the research team, or where the brief requires a public-facing evidence annex, the fields will need adjusting and some material will belong in a controlled location rather than the shared file. State those conditions in the register itself so the next reader knows what they are looking at.
Before the next brief opens, create the two tabs and log the first five documents already read. The gaps that appear in the inference and follow-up columns usually indicate where the search needs to go next; where the work is commissioned rather than self-directed, the arrangements and expected outputs are best set out in writing at the outset, as described on the for organisations page.