Quant-Qual Impact Narrative

Impact communications leads often hold two incomplete halves of the same story: dashboards of participation, retention, and outcome metrics, and a folder of client stories that never quite land in the same document. A quant-qual impact narrative use case closes that gap. An LLM drafts a single cohesive narrative that ties verified quantitative results to consented qualitative evidence, in language suitable for funders, boards, and community partners.

Tools in the Sopact class already organize indicators, survey responses, and story evidence in structured form. The model’s job is not to invent impact. It is to propose how those approved inputs read as one argument: what changed, for whom, by how much, and what the stories illustrate without overclaiming.

Staff remain the authors of record. The model drafts. Humans select which stories appear, confirm consent and anonymization rules, and approve every causal or outcome claim before anything ships.

Related workflows: Funder Report Section Drafting, Report Accuracy Adversarial Review, and Agentic Report Assembly.

Feed only material that already passed your organization’s verification and consent gates. Typical inputs include:

• Indicator tables or exports for the reporting period (counts, rates, change vs. baseline or prior period), with definitions and denominators intact.
• Disaggregations you are willing to publish (site, cohort, service line), when those cuts are methodologically sound and privacy-safe.
• Short, consented story records: quote or paraphrase, role or context label, date or program stage, and any required redactions.
• Mapping notes that already link a story to an indicator or outcome theme (for example, “illustrates housing stability at 6 months”).
• Audience and tone constraints: funder report, annual impact brief, board packet, or public web summary; reading level; banned phrases; required disclaimers.
• Claim policy: what you may state as association vs. attribution, how to handle small samples, and how to mark provisional or incomplete data.

Do not pass raw case notes, unreviewed survey free text, or stories lacking a recorded consent decision. If the metrics package is empty, or if no consented stories are available for the themes in scope, return empty output rather than filling gaps with generic copy or composite anecdotes.

Run the workflow after metrics are locked for the period and after story selection is complete (or explicitly scoped as “metrics-only sections,” in which case the qualitative weave is omitted and empty output is returned for story-dependent passages).

1. Normalize the evidence pack. Parse metrics into claim-ready facts: measure name, value, comparison, period, and caveats. Parse each selected story into attribution-safe fragments: allowed quote, theme tags, linked indicators.
2. Propose narrative architecture. Outline sections that mirror how readers scan impact writing: context and population served, core outcomes with numbers first, story illustrations nested under the claims they support, limits and next measurement steps.
3. Draft with strict grounding. Every quantitative statement must cite an input row or field. Every story beat must cite a consented story ID. Prefer adjacent placement: metric, then the story that illustrates it, then a short bridge sentence that does not escalate correlation into causation unless your claim policy allows it.
4. Flag risk language. Surface hedging needs for small n, missing baselines, selection bias in who shared stories, and any metric the model cannot reconcile with the accompanying text.
5. Stop conditions. If required metrics for a section are missing, leave that section empty. If consented stories for a planned illustration are missing, omit the illustration and do not invent a composite. If both sides are missing for the whole brief, return empty output.

The draft should read as one voice, not a metrics appendix glued to a testimonials page. Transitions should explain why a story sits next to a number, without implying the story proves the number.

Treat the model output as a first pass for the impact-communications lead and program staff who own the data.

• Story selection stays human. Staff choose which consented stories enter the pack. The model may suggest which selected stories best illustrate a given metric; it does not scrape or promote unselected cases.
• Consent and dignity check. Re-verify consent scope (internal vs. funder vs. public), names and identifiers, photos, and any detail that could re-identify someone in a small community.
• Claim approval. Program and evaluation owners confirm that wording matches what the data support. Soften or cut language that overstates attribution, generalizes from one story, or treats a proxy as the outcome itself.
• Tone and audience fit. Adjust for funder vs. community readers; remove jargon; keep participant voice respectful and accurate to the approved quote.
• Consistency with sibling sections. Align figures and phrasing with funder section drafts and any agentic assembly of the full report so the same outcome is not described three different ways.

Only after these gates should the narrative move into layout, PDF, or CMS. Adversarial accuracy review remains a useful follow-on for high-stakes submissions.

Use a short checklist rather than a vibes pass:

• Every number in the narrative appears in the metrics input with matching period and definition.
• Every story fragment maps to a consented record still in scope for this channel.
• No section was completed when its required inputs were empty.
• Disaggregations that could identify small groups were suppressed or aggregated per policy.
• Limitations are stated where evidence is thin, not buried in a footnote the model forgot to write.
• Related report sections that reuse the same indicators use the same figures and labels.

When a check fails, fix the inputs or the human-edited claims. Do not re-prompt the model to “sound more confident.”

Use it when you already have a verified metrics cut and a small set of consented stories, and you need a unified narrative for a funder report, impact brief, or board update without spending cycles stitching the two by hand.

Skip or narrow it when evaluation results are still in flux, story consent is incomplete, legal or communications review has not cleared participant voice for the channel, or the ask is pure statistical commentary with no qualitative layer. In those cases, metrics tables, a metrics-only summary, or empty output for the narrative layer are the correct products.

Pair this use case with funder section drafting when the narrative must land inside a specific RFP or report template, with adversarial review when the stakes of a misstated claim are high, and with agentic report assembly when the approved narrative becomes one chapter in a larger packet.