What is a cited narrative in a funder report?

In short: It explains what changed, for whom and why it matters, with every sentence traced to a record a reviewer can open. Citations make it checkable, not causal.

An impact narrative connects the work, the observed changes and their significance. A cited one lets the funder, or your reviewer, follow any sentence back to the wave, table or note it came from. The claim is still only as strong as the design: a before-and-after comparison with a footnote is still a before-and-after comparison.

On the usual path, the narrative is written last from an export pasted into ChatGPT, and sources are reconstructed afterwards, if at all. When each response already carries a person’s ID, wave and cohort, the sources exist before the first sentence.

Who reads the narrative, and what do they decide?

In short: Name the reader and the decision before you draft. A funder deciding on renewal needs its required measures answered first.

The training team’s funder is considering renewal and has asked for the spring cohort’s 30-day result; the delivery manager is deciding on a practice session. Same finding, different detail. The CDC’s evaluation-reporting guidance makes the point from public health: shape the report around purpose, audience and use, and keep the evidence and its limits intact.

Use a short sequence: what happened, what the evidence supports, what remains uncertain, what happens next. If the funder supplies a format, answer its fields first rather than replacing them with a more attractive story.

How do you keep outputs, outcomes and contribution apart?

In short: Say which kind of claim each sentence makes. Activity counts, observed outcomes and causal claims need different evidence; don’t let one slide into the next.

“We ran 12 sessions” is an output. “15 of 25 respondents used the skill at work within 30 days” is an observed outcome under a stated measure. “The course caused learners to use the skill” is a causal claim the team has no comparison group to support.

For a contribution claim, describe the plausible link between the work and the change, the evidence for it and the other explanations you considered. A learner’s account can show how the change happened for her; it is still one perspective. A count of responses is not always a count of people; the ID on each record tells them apart.

How do you build a claim-and-source table?

In short: List each piece of evidence, what it supports, the record it comes from and where its boundary lies. Check the arithmetic before you write a word of prose.

Here is the training team’s spring cohort (fictional), with Maria (ID 0417) as the illustration.

E1, F1, D1 and C1 are exercise labels. In your report, use references an authorized reviewer can resolve to the right record, wave or row.

Check the arithmetic: 15 + 10 = 25 respondents; 25 + 15 = 40 completers; 15 ÷ 25 = 60%; 25 ÷ 40 = 62.5%. The share of all completers known to have used the skill is 15 ÷ 40 = 37.5%. That is a floor, not an estimate, because the 15 unknowns could go either way.

How do you turn the table into a bounded paragraph?

In short: Write one sentence per row, cite it, state the coverage and unknowns, and end with an action that is clearly a next step, not a result.

The paragraph doesn’t suggest the time-to-practise theme explains the 15 non-respondents; that link hasn’t been shown. The action is written as a test; watch for “we will test” becoming “we solved” in review. For a board summary, keep the count, coverage, unknowns and action, and never shorten “60% of respondents” to “60% of learners”.

How do you choose and protect quotations?

In short: Say what job each quote does, keep it linked to the original response, check permission and identification risk, and keep the exception that changes the decision.

A quote can illustrate a common theme, explain an unexpected result, describe a minority experience or give one person’s account. Maria’s mentor note does the last job; say so in the text. Don’t choose by length or emotional force, and don’t assume the median-length answer is representative. When counts help, give the theme’s count and base beside the quote. A rare but serious barrier shouldn’t vanish below a frequency cutoff: “uncommon in these responses” is not “unimportant”.

Keep each quote linked to its response, read the sentences around it, and never merge several people’s words. For translated answers, keep the source text and note who translated it. People can be recognizable from a role or site without a name, so check permission for each audience, and keep a quote log: record ID, purpose, language, edits, permission and review date.

How do you use AI to draft from checked evidence?

In short: Give the model the claim-and-source table, not the raw export, ask it to flag what the evidence doesn’t contain, then test it with a question it can’t answer.

AI can assemble a paragraph or catch a missing qualifier. It shouldn’t invent a denominator, rebuild a missing source or decide one quote represents a cohort. Keep names, emails and phone numbers out of what the model sees; the paragraph needs none of them. In Sopact Sense you choose which fields are sent to AI, and each line of an AI Assistant answer links to its record, which shortens review without replacing it.

Draft a short funder report paragraph from the claim-and-source table I have shared.
Preserve every denominator, time window and limitation. Cite the listed reference after each claim. Report non-response as unknown.
Separate observed findings from interpretation and proposed action. Do not invent quotations or infer an overlap between groups.
Mark any request the table cannot support as "evidence not supplied". After the paragraph, return a claim-by-claim review list.

Then ask something the table can’t answer: “How many of the 15 non-respondents lacked time to practise?” A trustworthy workflow replies “evidence not supplied”; a plausible number is a warning.

How do you review the paragraph before it is sent?

In short: Ask five questions, and have someone who didn’t draft the paragraph follow one claim all the way back to its record.

If the reviewer can’t reach the source, repair the trail before polishing the design. Record who approved the paragraph, and when, in the team’s change log, and keep a dated copy of the evidence. If a source is corrected later, decide whether the published claim needs correcting too.

The limits travel with the narrative. These are self-reports from 25 of 40 completers, the 15 who didn’t reply may differ, one 30-day checkpoint says nothing about lasting use, and nothing here shows the course caused the change. People decide what the finding means.