---
name: finding-to-insight-development
description: >
  Converts research findings into genuine insights: explanations of why a finding
  is true, tested against the whole evidence base and traceable back to it. Use
  for "what's the insight", "so what does this mean", "turn these findings into
  insights", "why is this happening", "make this more insightful", "we have the
  data but no story", "the client says this is just data".
category: 08 Insight Development
ref: 08.01
tier: 0
inherits: [K2, K3, K4, K5]
---

# Finding to Insight Development

## 1. One-line description
Takes established findings and works the step from what the data shows to why it is happening, by generating competing explanations, testing each against the entire evidence base, eliminating the ones the evidence does not support, and reporting what survives at the confidence it has earned.

## 2. What this skill is used for

**The research problem it solves.** Most reports stop at findings and label the last slide "insights". The numbers are checked, the quotes are real, and the reader still learns nothing they did not already know. What is missing is explanation: why the finding is true, what mechanism produces it, and what that means for a decision. That step is hard, so it gets faked, and the standard fake is restatement. A statistic is reworded in more confident prose, the word "insight" is put above it, and the output looks finished. This is common under deadline and close to universal in AI-generated work, because rewording a sentence is what a language model is built to do, and the result is fluent, plausible and empty. This skill separates explanation from rewrite, and supplies the tests that catch the difference.

**Where it sits.** Synthesis. It takes tested findings from quantitative and qualitative analysis and hands explained, evidenced, confidence-rated insights to prioritisation, implication and recommendation development.

**Typical use cases.**
- A completed analysis with solid findings and a client asking "so what".
- Mixed-method studies where two streams need one explanation between them.
- Tracking waves where a metric has moved and nobody can say why.
- Segmentation output that describes groups without explaining what makes them different.
- A draft where the insight section restates the findings section.
- Auditing insights someone else wrote, AI-generated ones included, before they reach a client.

**Who uses it.** Research directors and senior analysts responsible for what a report concludes; client-side insight managers expected to turn findings into meaning; strategists working from someone else's data; anyone reviewing AI-drafted synthesis, where this failure is the likeliest to survive review because it looks correct.

## 3. When to use it

- Analysis is complete, findings are established with sources and bases, and the question is now what they mean.
- The draft reads as accurate and tells the reader nothing they could act on.
- Several findings look related and you suspect one explanation sits underneath them.
- Two evidence streams point in different directions and the report needs to say something honest about both.
- A client or stakeholder has said the work is "descriptive", "just data", or "a data dump".
- A number moved between waves and the report currently states the movement without accounting for it.
- You are checking work, human or AI, for insights that are actually findings in better clothing.
- The findings will feed a decision, and someone needs to be able to explain to the decision-maker why the world works the way the data says it does.

## 4. When NOT to use it

- **No findings have been supplied.** This is the hard stop. Insight development without findings is invention, and it is precisely the route K4 §1 closes: the output format has a slot, so the system fills it. If you have raw data, run the analysis first (**05.01** to **05.03** for quantitative, **07.01** for qualitative). If you have nothing but a topic and a deadline, this skill cannot help you and neither can any other; say so.
- **The findings do not support any insight.** This happens and it is not a failure. A study can produce accurate, well-based findings that are fully explained by the obvious, or that are too thin, too mixed or too method-bound to carry an explanation. The honest output is to report the findings, state plainly that the evidence does not support an explanatory claim, name what would be needed to get one (a different question, a behavioural measure, a longitudinal design, a subgroup with an adequate base), and stop. A report with six findings and no insights, clearly labelled, is more useful and more professional than a report with six findings and six restatements. Refusing to manufacture the missing rung is the single most valuable thing this skill does.
- **The findings are not yet stable.** Preliminary counts, uncleaned data, partial fieldwork or open coding are not findings. Explaining a finding that later changes wastes the explanation and, worse, tends to survive the correction, because the story is more memorable than the number underneath it.
- **The question is really "which of these matters most".** Prioritisation, sizing and sequencing are **08.05 Insight Prioritisation and Sizing**. Do not build an insight to justify a priority already decided.
- **The question is really "what should we do".** Recommendation development is **08.04**, and it depends on organisational knowledge this skill does not have. An insight that arrives pre-attached to an action has usually been reverse-engineered from the action.
- **The pattern itself is not yet established.** If you are still asking whether several findings form a pattern, that is **08.02 Pattern Identification**. Explaining a pattern that does not exist is the most expensive error available here, because everything downstream inherits it.
- **A conclusion has already been chosen.** Where a stakeholder has supplied the insight and wants evidence attached, this becomes a search for confirming evidence, which K4 §4.2 prohibits. Say so, and offer instead an explicit, labelled assessment of the stated hypothesis against all the evidence, including the evidence that cuts against it.
- **The explanation requires cultural or contextual reading you cannot do.** Why a finding is true in a market whose idiom, norms and institutions you cannot read is not answerable from a translated transcript. Per K5 §2.2, code and report the finding, and route the explanation to a human with the context.
- **The insight would identify individuals or expose a vulnerable group.** Small bases, distinctive subgroups and sensitive topics can make an explanatory claim about a group that the group would not consent to and could be harmed by. That is **13.05 Research Ethics and Consent Design**, before it is an analysis question.

## 5. Required inputs

**Required.** Without these the skill cannot run. If they are absent, ask. If the first is absent, stop.

- **At least one finding, stated as a finding.** A statement of fact about the data, with its source reference and base per K2 §4. Not a topic, not a research question, not a hypothesis, not a client's belief. **If no finding is supplied, do not proceed.** Say what is missing, say why the skill cannot substitute for it, and ask for the analysis output. This refusal is not a formality: generating insights from an absent evidence base is fabrication with a professional vocabulary, and it is indistinguishable from real work at the point of reading.
- **The evidence base behind the findings**, not only the findings themselves. Explanation is tested against everything available, including the parts that were not used to build the finding. Where only the findings exist and the underlying material is gone, say so: candidate explanations can be generated but not properly tested, and confidence is capped accordingly.
- **The research objectives and the decision the study informs.** Relevance is not assessable without them, and an insight that is true and irrelevant is a cost, not a contribution.

**Optional, and what each one adds.**

- **The full dataset, transcripts or code frame:** allows candidate explanations to be tested against evidence nobody has looked at yet, which is where explanations usually die. Without it, testing is limited to the findings already written up.
- **Findings from the other method stream:** makes triangulation possible, and makes divergence visible early enough to investigate rather than smooth.
- **Behavioural, transactional or operational data:** the strongest single test of an attitudinal explanation, because it shows what people did rather than what they said they would do.
- **Previous waves or earlier studies:** establishes whether the finding is new, and whether a proposed explanation held before. An explanation that only works for this wave is usually a description of this wave.
- **What the audience already believes**, from the brief, the debrief, or the stakeholder interviews: sets the baseline against which non-obviousness is judged. Non-obvious is relative to a reader, and without the reader it is guesswork.
- **Known constraints and history** (what has been tried, what is underway, what failed): prevents insights whose only implication is something the organisation already knows it cannot do.
- **Method documentation:** lets you distinguish an explanation about the world from an explanation about the instrument, which is one of the four candidate families in step 5.

## 6. Questions to ask before starting

1. **What decision does this inform, who owns it, and when is it made?** Determines which explanations are worth pursuing and what "strategically useful" means here. *Default if unanswered:* develop insights against the stated research objectives, and flag that materiality is a researcher judgement per K5 §2.1.
2. **What does the audience already believe about this?** Sets the non-obviousness baseline and identifies which insights will need the most evidence, because the ones that contradict a held belief are the ones that get challenged. *Default:* assume a well-informed reader who knows their own category, and mark anything they would plausibly already know.
3. **What is the complete evidence base, including material not used in the findings?** Determines whether explanations can be tested or only generated. *Default:* work with what is supplied, and state explicitly that testing was limited to the findings themselves.
4. **Do any evidence streams disagree?** Divergence is investigated, not averaged, and finding it early changes the whole approach. *Default:* check for it during step 6 and report any that appears.
5. **Has this question been researched before, here or elsewhere?** Determines whether a candidate explanation has already been tested and whether an insight is genuinely new to this organisation. *Default:* treat the finding as new to the organisation and flag the assumption.
6. **What would have to be true for the business to act on this?** Distinguishes an insight that changes a decision from one that is merely interesting. *Default:* develop it anyway and mark materiality for researcher review.
7. **Are the findings final?** Determines whether explanation is premature. *Default:* ask, because explaining a provisional number is wasted work that outlives the number.

## 7. Step-by-step methodology

**The position this method takes.** An insight is a claim about a mechanism: it says why the finding is true, in a way that could be wrong. That last clause is the whole discipline. A finding is not arguable, because it is a statement about the data. An explanation is arguable, because it is a statement about the world that the data constrains but does not determine. If nobody competent could disagree with your insight, you have not explained anything, you have rephrased the finding. Everything below exists to generate real candidate explanations, kill the ones the evidence refuses, and report the survivors at the confidence they earned.

**1. Gate on inputs.** Confirm that at least one genuine finding is present, with a source reference and a base. If not, stop and ask, per Section 5. Do not proceed on a topic, a brief, a client hypothesis or a set of raw responses. *Correct result:* a numbered list of findings, each one a statement of fact about the data that could be checked against source.

**2. Re-level everything you have been given.** Findings handed over in draft reports are routinely contaminated: an interpretation sits inside the sentence, or a causal verb has crept in. Strip each input back to what was observed and park the rest as a candidate explanation to be tested later rather than a fact to build on. "Customers left because onboarding was slow" is not a finding; the finding is the association, and the "because" is an untested candidate. *Correct result:* a clean finding set, plus a separate list of the explanations that were smuggled into it. That second list is useful: it usually contains the client's prior beliefs.

**3. Inventory the whole evidence base before explaining anything.** List every stream available: which questions, which bases, which transcripts, which behavioural or operational data, which prior waves, which desk sources. Do this before generating explanations, because the order matters. Inventory first and you test against everything; explain first and you test against whatever you happen to remember, which is the evidence that produced the explanation. *Correct result:* a one-page evidence inventory with source codes per K2 §6, including material that has not yet been analysed.

**4. State the finding precisely, with its boundary.** Write it with its base, its population, its period, and the conditions under which it holds. "Self-serve abandonment is 31%" is weaker than "31% of customers who started a self-serve task abandoned it, among those who attempted one in the last three months, n=340, highest in document-upload tasks". The second version has already told you where to look for an explanation. Precision at this step does more analytical work than anything later. *Correct result:* a finding statement that names what varies, because an explanation must account for the variation, not just the average.

**5. Generate multiple candidate explanations, deliberately.** Ask "why is this true?" and answer it at least three times, with genuinely different answers. Then check that you have drawn from all four families, because the missing family is usually where the true explanation is:

- **Mechanism:** a structural or process feature of the product, journey, channel or market produces this. Something in the world makes the behaviour easy, hard, rational or unavoidable.
- **Motivation and meaning:** people are pursuing a goal, avoiding a cost, managing an identity or holding a belief that makes this behaviour sensible from where they stand.
- **Context:** something outside the study produced it. A price change, a competitor launch, a seasonal pattern, a policy shift, a news event during fieldwork.
- **Artefact:** the finding is about the instrument, not the world. Question wording, order effects, routing, sample composition, the definition of the metric, who was reachable, who dropped out.

Generating one explanation and proceeding is how confirmation bias enters this work, and it is the default behaviour of both a tired analyst and a language model. The first explanation is generated by fluency, not by evidence, and once it exists everything read afterwards is read as support for it. Write the candidates down before testing any of them, so that the discipline is visible. *Correct result:* three to six named candidates, including at least one artefact candidate, each stated as a mechanism you could argue for.

**6. Test every candidate against the entire evidence base.** For each candidate, in this order, answer three questions. **What evidence is consistent with it?** **What evidence should exist if it were true, and does it exist?** This is the question that does the work: a real explanation predicts things beyond the finding that inspired it, and those predictions can be checked against evidence nobody has looked at yet. **What evidence contradicts it?** Testing a candidate only against the evidence that suggested it guarantees survival and proves nothing. Go back to the inventory from step 3, including the streams and subgroups that were never part of the original finding. *Correct result:* a completed test grid (Section 8) in which at least one candidate has failed, and in which the surviving candidate has passed a prediction it could have failed.

**7. Eliminate, and keep the record.** Discard the candidates the evidence refuses, and write down why each was discarded. The eliminations are not waste: they are the strongest available demonstration that the surviving explanation was tested rather than chosen, and they are the first thing a sceptical stakeholder will ask about. Where an artefact candidate survives, that outranks everything: an explanation about the instrument must be resolved before any explanation about the world is offered. *Correct result:* an elimination log, one line per rejected candidate, naming the evidence that killed it.

**8. Triangulate, and investigate divergence rather than resolving it.** Where more than one stream bears on the finding, state whether they converge or diverge, per K2 §4.4. Convergence across independent methods is the strongest support an insight can have, and it is what licenses high confidence under K3 §3.1. Divergence is not a problem to be averaged away: it is usually the most informative thing in the study, and averaging destroys it. When streams disagree, ask what would have to be true for both to be right. The usual answers are that the two streams measured different things, asked different people, covered different periods, or that one captured what people say and the other what people do. Any of those is itself an insight. *Correct result:* an explicit convergence or divergence statement per finding, and where there is divergence, an account of what produced it rather than a compromise between the numbers.

**9. Handle the survivors honestly, including when there are none.** Three outcomes are legitimate. **One candidate survives:** you have an insight, at a confidence set by the strength of the test. **Several survive:** report them as competing explanations with the evidence for each, and name the analysis or the study that would separate them. Two honestly reported explanations are worth more than one arbitrarily chosen. **None survives:** report the finding without an insight, say the evidence does not support an explanation, and name what would be required. Per Section 4, this is a legitimate output and it must not be filled in with the least-bad candidate. *Correct result:* one of these three, chosen by the evidence rather than by the shape of the report template.

**10. Climb the "so what" ladder one rung at a time.** From the surviving explanation, ask "so what" repeatedly: insight, then implication for the organisation, then possible action. Climb one rung per question, and at each rung name the evidence that supports the new claim, not the evidence that supported the rung below. You have gone one rung too far when any of four things happens: the new claim rests on an assumption about the organisation that the research did not test; the only evidence you can cite is for the previous rung; the claim would still be stated if the finding had come out differently; or a quantity has appeared (a size, a value, a timeframe) that was never measured. When you hit one of those, step back down and stop. The rung above is not necessarily wrong, but it is not yours to assert. *Correct result:* a ladder written out explicitly, with the rung where the evidence stopped marked, and anything above it labelled as a question for the business rather than a conclusion from the research.

**11. Run the four restatement tests.** Every candidate insight passes all four, or it is a finding in better grammar (Section 8). Run them in writing, not in your head, because restatement is invisible to a reader and nearly invisible to the writer.

**12. Run the eight-criterion quality screen** in Section 8. Screen each surviving insight and record the answer to each test question. An insight that fails "specific" or "human" is usually recoverable by rewriting. One that fails "evidence-backed" or "explanatory" is not.

**13. Assign confidence per K3, and cap it.** Work the six factors in K3 §3. Then apply the rule that governs this skill more than any other: **an insight cannot be more confident than the findings beneath it, and an insight resting on one finding from one stream is at most moderate.** Inferential distance costs confidence, and an insight is two steps beyond the data. Where the finding is high confidence and the explanation is one of two the evidence permits, the insight is moderate, and the moderate-confidence language in K3 §4.2 is required, including the alternative explanation in the same passage. *Correct result:* a confidence level per insight, with the specific factor that set the ceiling named.

**14. Write the chain, and mark the judgement points.** Write out the full chain for each insight in the K2 notation, with the interpretation boundary marked, so a reader can see where reporting stopped and arguing began. Then mark the human judgement points per K5: whether this matters to the business (§2.1), whether the cultural reading is right (§2.2), and whether the strategic reading survives contact with what the organisation can actually do (§2.3). These are marked, not resolved. *Correct result:* an insight record a research director can audit end to end without asking a single "where did this come from" question.

## 8. Analytical framework

**The chain, per K2 §2, with the boundary marked.**

```
Evidence → Analysis → Finding | Interpretation → Insight → Implication → Recommendation
                              ↑
              This skill works across this line.
              Left of it: observed, checkable, not arguable.
              Right of it: argued, contestable, must carry its confidence level.
```

**Observation, finding, insight: worked contrasts.** *Fictional illustration, from an insurance renewal study.*

| Statement | Level | Why |
|---|---|---|
| "One participant said she could not tell why her renewal price had gone up." | **Observation** | True of one person. No analysis behind it, no claim about the set. Useful as a lead, not as evidence for a claim. |
| "19 of 26 interviewees could not say what their renewal price was based on [P01 to P26]." | **Finding** | Established across the corpus, with a base. Says what, not why. Not arguable. |
| "Renewal pricing is opaque to customers." | **Finding, reworded** | Restatement. "Opaque" is a synonym for "could not say what it was based on". Nothing has been added but confidence. |
| "Customers read an unexplained renewal price as evidence that the price is negotiable, so the letter intended to close the decision reopens it." | **Insight** | Names a mechanism the finding does not contain (opacity leads to an inference of negotiability, which triggers shopping). Contestable, and testable against quote-comparison and switching data. |
| "Renewal revenue is lost when the letter arrives, not when the competitor is contacted." | **Implication** | What it means for the organisation, and where intervention would have to sit. |

**The four restatement tests.** A candidate insight must pass all four.

| Test | Question | Fails if |
|---|---|---|
| **Reversal** | Can I recover the finding by deleting words from the insight? | Yes. Nothing was added. |
| **Mechanism** | Write it as "X happens because Y". Does Y contain something not present in the finding? | Y is a synonym, a category label, or the finding restated ("they fall behind because they get behind"). |
| **Prediction** | If this explanation is true, what else should be observable? Did I check? | Nothing else follows from it, or nothing was checked. |
| **Contestability** | Could a competent colleague, seeing the same finding, hold a different explanation? | No. Then it is not an explanation, it is the finding. |

**The candidate test grid.** One row per candidate, completed before any is chosen.

| Candidate explanation | Family | Evidence consistent | Should exist if true, and does it? | Evidence contradicting | Survives |
|---|---|---|---|---|---|

**The eight-criterion quality screen.** Each criterion has a test question with a checkable answer.

| Criterion | Test question |
|---|---|
| **Evidence-backed** | Can I name at least two independent pieces of evidence, with sources and bases, and point to each in the material? |
| **Explanatory** | Does it answer why, and does the "because" clause contain something the finding does not? |
| **Non-obvious** | Would a well-informed person in this organisation have written this sentence down before the study? If yes, is it at least now established rather than assumed? |
| **Relevant** | Does it bear on the decision this research exists to inform, and can I name that decision? |
| **Specific** | Could this sentence be pasted into a report for a different organisation in a different category unchanged? If yes, it is a platitude. |
| **Human** | Is there a recognisable person doing a recognisable thing for a reason a reader could recognise in themselves? |
| **Strategically useful** | Does it change what someone would do, or what they would stop doing? |
| **Decision-connected** | Can I name the decision, and the person who owns it, that this bears on? |

**On non-obviousness.** The criterion is a filter, not a target. An insight that confirms what everyone believed, and is now supported by evidence for the first time, is valuable and is reported: converting an assumption into a tested finding removes a risk the organisation was carrying without knowing it. Say plainly that it confirms the prior belief, and say what the evidence adds. Equally, a surprising claim is not an insight by virtue of being surprising. Surprise is a property of the reader, not of the evidence, and the surprising reading is the one most likely to have skipped a test. Where an insight is both surprising and true, it needs more evidence than a dull one, not less, because it will be challenged harder and because the most common cause of surprise is an error.

## 9. Output format

**A. Insight record** (one per insight)

```
INSIGHT           One sentence. The explanation, not the finding.
CONFIDENCE        High / Moderate / Low-hypothesis, with the factor that set the ceiling
RESTS ON          Findings by reference, with bases
EVIDENCE STREAMS  Named, with convergence or divergence stated
THE CHAIN         [EVIDENCE] → [FINDING] → [INTERPRETATION] → [INSIGHT]
                  with the interpretation boundary marked, per K2 §3.1
MECHANISM         Why this produces the finding, in two or three sentences
CANDIDATES TESTED The alternative explanations considered
ELIMINATED        Each rejected candidate and the evidence that rejected it
SURVIVING RIVALS  Any competing explanation still standing, and what would separate them
NON-OBVIOUSNESS   New / confirms a prior belief with evidence for the first time / already known
COUNTER-EVIDENCE  What cuts against this, or "none found in this evidence base"
WHAT WOULD CHANGE IT   The observation or study that would refute or strengthen it
SO-WHAT LADDER    The rungs climbed, with the evidence limit marked
REVIEW POINTS     Per K5, at the point of the judgement
```

**B. Insight summary table**

| # | Insight | Confidence | Rests on | Streams | Converge / diverge | What would change it |
|---|---|---|---|---|---|---|

**C. Divergence register**
Every place two streams disagreed, what differs, what was investigated, and what was concluded. Never a reconciled midpoint.

**D. Findings without insights**
Findings for which no explanation survived testing, with what would be needed to explain each. This section is required, and an empty one is unusual enough to be worth checking.

**E. Rejected candidate explanations**
The elimination log, including any client or stakeholder hypothesis that the evidence did not support. This is frequently the most valuable page in the document.

**When the evidence is thin.** The format does not get filled to look complete, per K4 §1. Where one finding supports one moderate-confidence insight, the output is one insight, and the deck has one. Where nothing survives, section D carries the findings and sections A and B say so. Do not manufacture a second insight to balance a layout, do not split one insight into three to fill a template, and do not promote a low-confidence candidate because a section would otherwise be empty. An insight count is not a quality measure, and the number of insights in a study is a property of the evidence, not of the report.

## 10. Quality checks

Run before anything is presented. Sits on top of K4 §8.

1. Was at least one genuine finding supplied, and is every insight traceable to one by reference?
2. Has every candidate insight passed all four restatement tests, in writing?
3. Were at least three candidate explanations generated for each finding, including one artefact candidate?
4. Was each candidate tested against evidence beyond the evidence that suggested it?
5. Is the elimination log present, with the evidence that rejected each candidate named?
6. Does every insight beyond low confidence rest on at least two independent pieces of evidence?
7. Is the interpretation boundary visible in every chain, so a reader can find where argument begins?
8. Does any insight carry more confidence than the weakest finding beneath it?
9. Where streams diverge, is the divergence reported and investigated rather than averaged?
10. Does every moderate-confidence insight carry its alternative explanation in the same passage, per K3 §4.2?
11. Is any causal verb used where the design does not license it, per K4 §3.2?
12. Could any insight sentence be pasted into a different study in a different category unchanged?
13. Has any insight been discarded for being unsurprising, or promoted for being surprising?
14. Does every insight name the decision it bears on?
15. Is the "findings without insights" section present, and honest?
16. Are the K5 judgement points marked at the point of the judgement rather than collected at the end?

## 11. Common failure modes

| Failure | How to recognise it | How to prevent it |
|---|---|---|
| **Restatement as insight** (the signature failure) | The insight and the finding say the same thing at different grammatical temperatures. Deleting words from the insight recovers the finding | The four restatement tests, run in writing. Require a mechanism clause that contains something absent from the finding |
| **Synonym substitution** | The finding's key term has been swapped for an abstract noun: "could not say why" becomes "opacity", "did not return" becomes "disengagement" | Naming a behaviour is not explaining it. Ask what produces the thing the noun names |
| **Single-explanation lock-in** | Only one explanation appears anywhere in the working documents, and all evidence cited supports it | Generate candidates before testing any. Write them all down first. Require an artefact candidate every time |
| **Testing against the inspiring evidence only** | The insight is supported by exactly the finding that produced it and nothing else | Step 6, question two: what should exist if this were true, and does it? Return to the full inventory |
| **The unfalsifiable human truth** | A confident psychological flourish ("people crave control") that no evidence could contradict and that would fit any category | Apply the specificity test. If it survives being pasted into a different study, it is a platitude |
| **Causal smuggling** | "Drives", "leads to", "because of", attached to an association from a cross-sectional design | K4 §3.2. State the association, state the proposed mechanism as an interpretation, name what design would establish it |
| **Confidence inflation across the boundary** | A high-confidence finding produces a high-confidence insight, with no acknowledgement of the inferential step | K3 §3.6. The insight inherits the weakest link, and the interpretive step is itself a link |
| **Averaging divergent streams** | "The survey says 64%, the interviews suggest lower, so it is probably around half" | Never average. Ask what would have to be true for both to be right, per step 8 |
| **Novelty-seeking** | The most counter-intuitive reading was chosen, and its evidence is thinner than the dull reading it displaced | Surprise raises the evidential bar, it does not lower it. Test the boring explanation too |
| **Obviousness-phobia** | A well-evidenced insight was cut because someone said "we knew that" | Report it, labelled as confirming a prior belief with evidence for the first time, and say what the evidence adds |
| **Insight without findings** | The output is fluent, structured, and no reference points to a base | The step 1 gate. Refuse, name what is missing, do not proceed |
| **Template-filled insight count** | Exactly five insights, one per section, several of them thin | The number of insights is set by the evidence. Sections D and E absorb the rest |
| **The insight that is really a recommendation** | It contains a verb the client is meant to perform | Recommendations are 08.04. An insight explains the world; it does not instruct the organisation |
| **Chain amputation** | The insight travels into the deck without its findings, bases or confidence | K2 §7. The chain travels with the claim into every downstream document |

## 12. AI guardrails

Skill-specific only. Universal prohibitions are inherited from **K4** and are not repeated here. Evidence chain and level boundaries follow **K2 §2 and §3**; confidence language follows **K3 §4**.

1. **Never generate an insight without at least one supplied finding.** If no finding is present, stop, state what is missing, and ask for the analysis. Producing insights from a topic, a brief or raw responses is the fabrication route K4 §1 closes, and it is the specific failure this skill exists to prevent.
2. **Never present a restatement as an insight.** Run the four tests in Section 8 in writing and keep the record. A reworded statistic with a confident verb is a finding with better grammar, whatever it is labelled.
3. **Never generate one explanation and proceed.** A minimum of three candidates, spanning the four families, including one artefact candidate, before any is tested.
4. **Never test a candidate only against the evidence that suggested it.** Testing is against the full inventory, including material not used to build the finding.
5. **Never advance an explanation about the world while an artefact explanation is unresolved.** Question wording, routing, sample composition and metric definition are checked first.
6. **Never let an insight carry more confidence than the findings beneath it**, and never treat a single finding from a single stream as support for a high-confidence insight, per K3 §3.6.
7. **Never resolve divergence between streams by averaging, splitting the difference, or reporting only the stream that fits.** Report both, investigate the gap, and per K2 §4.4 say what differs.
8. **Never discard an insight because it is unsurprising, and never promote one because it is surprising.** Both are judgements about the reader rather than about the evidence.
9. **Never assert a psychological or emotional mechanism the evidence cannot reach.** "People felt disrespected" requires evidence that people said, showed or behaved as though they felt disrespected. Inferred inner states are interpretations and are labelled as such, with their confidence.
10. **Never let an insight leave the working document without its chain**: the findings it rests on, their bases, the streams, the confidence level, and what would change it.
11. **Never fill an insight slot to complete a template.** Where no explanation survived, the honest output is the finding plus a statement of what would be needed, per Section 9D.

## 13. Best-practice principles

1. **An insight is a claim that could be wrong.** If no competent colleague could disagree with it, it is not an explanation. Contestability is the property that separates insight from summary, and it is uncomfortable by design.
2. **The explanation must account for the variation, not just the average.** The strongest insights explain why the finding is stronger here and weaker there. An explanation that would predict the same result everywhere has not engaged with the data's shape.
3. **The first explanation is the one to distrust most.** It arrives by fluency and familiarity, before any evidence has been consulted, and everything read afterwards will be read as supporting it.
4. **Ask what else should be true.** A real mechanism has consequences elsewhere in the data. Going to look for them, and finding them, is what turns a plausible story into an insight. Going to look and not finding them is more valuable still.
5. **Include the boring explanation on purpose.** Seasonality, a question change, a sample shift, a competitor's promotion. It is often right, it is cheap to check, and skipping it is how a methodological artefact becomes a strategy.
6. **Divergence between streams is a gift.** People saying one thing and doing another is not a data quality problem, it is usually the most useful finding in the study, and it dies the moment someone averages it.
7. **Explain the finding, not the topic.** An insight about "customer loyalty in general" attached to a finding about renewal letters has floated free. The test is whether the insight would change if this specific finding changed.
8. **Prevalence is not importance and neither is novelty.** A mechanism affecting a small group at a decisive moment can outrank one affecting everyone at an irrelevant one.
9. **Write the mechanism in the language of a person, not a category.** "Segment B exhibits low engagement propensity" explains nothing. "People who missed the first week could not find a way back in" explains, and can be checked.
10. **An insight with one piece of evidence is a hypothesis.** Say so, label it, and name the study that would test it. It may still be the most valuable thing in the report, provided nobody mistakes it for established.
11. **Confirming an assumption is a result.** Organisations act on untested beliefs constantly. Turning one into an evidenced finding removes a risk, and reporting it as such is more honest than dressing it as a discovery.
12. **Keep the elimination log.** It is the proof that the surviving explanation was tested rather than chosen, and it is the first thing a serious stakeholder asks for.
13. **Stop climbing when the evidence stops.** The rung above what the evidence supports is where the research ends and the opinion begins. Both can appear in a report; only one of them can be labelled as a finding of the study.

## 14. Worked example

*Fictional scenario, used to demonstrate method. All organisations, participants, figures and findings below are invented.*

**INPUT.** A national vocational training body commissions research into why learners abandon its online certificate courses. Findings supplied: **F1**, 58% of withdrawals occurred between week 3 and week 5 of a 12-week course (course records, n=1,410 enrolments across four cohorts). **F2**, learners attending at least one live tutorial withdrew less often, 21% versus 39%, untested. **F3**, 19 of 26 interviewed leavers described falling behind rather than losing interest [P01 to P26]. **F4**, 64% of leavers agreed "I intended to come back to it" (exit survey, n=380). Commissioning belief, recorded beforehand: learners lose motivation.

**PROCESS.**

*Steps 1 to 4.* The gate passes: the findings are genuine and referenced. F2 arrives as "tutorials reduce drop-out", an interpretation inside a finding, so it is stripped back to the association and the causal claim parked as a candidate. The finding is restated with its boundary: withdrawal concentrates on course weeks 3 to 5, across cohorts that began in different calendar months.

*Step 5, candidates.* **A (motivation):** novelty fades and learners disengage. **B (mechanism):** material is released weekly with no catch-up route, so one missed week becomes a permanent deficit. **C (context):** weeks 3 to 5 coincide with a seasonal work peak. **D (artefact):** withdrawal is defined as 21 days without login, so learners who paused are recorded as leavers.

*Step 6, testing against everything.* A is contradicted by F3 and F4: leavers describe falling behind, not boredom, and most meant to return. C is killed by evidence never used to build the finding, since withdrawal tracks course week rather than calendar week across cohorts. D partly survives, so it is resolved first: re-enrolment records show 6% of recorded leavers ever returned, so the definition overstates permanence but does not manufacture the week 3 to 5 concentration. B predicts checkable things, and they hold: withdrawal follows course position (F1), leavers describe deficit rather than disinterest (F3), intention to return is high while realisation is not (F4 against the 6%), and any re-entry point is associated with lower withdrawal (F2, association only).

*A candidate insight, rejected.* The first draft reads: "Learners drop out mid-course because they fall behind." It fails reversal, since deleting five words recovers F3; it fails mechanism, since "fall behind" is the finding's own language rather than an account of what produces it; and it fails contestability, since no colleague could hold a different view of it. It is F3 with a causal conjunction inserted, and exactly the sentence that passes unchallenged in a deck.

*Step 8, divergence, and the judgement call.* 64% said they intended to return; 6% did. The tempting move is to call the survey figure inflated and move on. Instead: what would make both true? Intention was real and there was no route back. That turns a data quality irritation into the core of the explanation. The survey figure is reported as evidence about how leavers understood their exit rather than as a forecast of return, and the gap between the two numbers becomes a finding in its own right.

*Steps 10 to 13.* The ladder: insight, then the implication that retention spend aimed at motivation is aimed at the wrong mechanism, then a possible action on re-entry design. The third rung assumes things about what this body can deliver under its funding rules, so it is marked as a question for the business, not a conclusion. Confidence is **moderate**: three streams converge and a competing candidate was eliminated by evidence, but the design is observational and F2 is an untested association.

**OUTPUT.** Insight: *learners are not leaving because they stop wanting the qualification; they are leaving because the course has no way back in. A missed week becomes a permanent deficit, so learners who still intend to return have nowhere to return to.* Recorded with its chain, its moderate confidence, the eliminated candidates and the evidence that killed each, the artefact caveat on the withdrawal definition, the unresolved intention-to-return divergence, and a K5 §2.1 review point on whether a re-entry route is materially available here.

## 15. Advanced usage

**Working across a set of findings rather than one.** Where several findings may share one explanation, test candidates against all of them simultaneously. An explanation that accounts for four findings is far stronger than four separate explanations, and it is also far more dangerous, because a single elegant story is the thing least likely to be tested. Require it to fail on at least one finding, or check whether it has been stretched to fit.

**Insight from the null result.** A finding that something did not differ, did not move or did not matter is explainable and routinely discarded. Segments that were expected to diverge and did not are frequently more useful than the ones that did, because they tell an organisation that a distinction it is spending money on does not exist for customers.

**Adversarial testing.** Take each surviving insight and assign the strongest available case against it, using the evidence base. An insight that survives a deliberate attempt to break it earns a higher confidence level under K3 §3.5 and arrives at the client already stress-tested. Where a stakeholder hypothesis exists, run this on the hypothesis too, so that agreement is demonstrably not deference.

**Repository and multi-study work.** Insights held in a repository decay, because the mechanism was true of a market that has changed. Store each with its date, its findings, its bases and its confidence, and re-test rather than re-quote. An insight quoted three years later without its chain has become folklore, per K2 §7.

**When the standard approach does not fit.** Where findings arrive without their underlying data, generate candidates, mark them explicitly as untested, cap confidence at low, and name the tests that would be run if the data were recovered. Where the evidence base is a single stream, say so in the insight record: a single-stream insight is a hypothesis with good manners, and the language in K3 §4.3 applies.

## 16. Skill chain

**Recommended previous skills:**
- **05.01 Descriptive Statistical Analysis**, **05.02 Cross-Tabulation and Segmentation Analysis**, **05.03 Statistical Significance Testing.** Hand over quantitative findings with bases, subgroup structure and tested differences, so that explanation starts from established variation rather than from an average.
- **07.01 Thematic Analysis.** Hands over tested themes with prevalence, counter-evidence and preserved contradictions, which are the raw material of mechanism.
- **07.06 Qualitative and Quantitative Integration.** Hands over convergence and divergence already identified across streams, so triangulation at step 8 starts from a mapped position.
- **08.02 Pattern Identification.** Establishes that a set of findings forms a pattern before this skill attempts to explain it.

**Recommended next skills:**
- **08.05 Insight Prioritisation and Sizing.** Takes explained, confidence-rated insights and decides which are material enough to lead the report.
- **08.03 Implication Development.** Takes the insight and works the rung from why this is true to what it means for the organisation.
- **08.04 Recommendation Development.** Takes implications and works the final rung to action, with the organisational knowledge this skill deliberately does not assume.
- **12.01 Research Report Architecture.** Takes insights with their chains intact and builds the narrative around them without breaking the traceability.

**Runs well alongside:**
- **13.03 AI Output Verification.** Run against any insight set, including this one's output. The restatement tests in Section 8 are the specific check it should apply here.
- **13.04 Bias Detection**, particularly where stakeholder hypotheses were recorded before analysis.
- **K5**, at the three judgement points this skill mandates: business materiality, cultural interpretation, and strategic reading.

---
A Yazi Supplied Skill and resource.