--- name: insight-prioritisation-and-sizing description: > Decides which of a study's insights matter, using eight named criteria applied as structured argument rather than averaged into a score, and sizes an opportunity or risk only as far as the data honestly allows, always as a range with every assumption stated. Keeps the full list so nothing disappears silently. Use for "which insights matter most", "we have too many insights", "what should lead the report", "size the opportunity", "how big is this", "prioritise the findings", "what goes in the executive summary". category: 08 Insight Development ref: 08.05 tier: 1 inherits: [K2, K3, K4, K5] --- # Insight Prioritisation and Sizing ## 1. One-line description Ranks a study's insights against eight explicit criteria without collapsing them into a single score, sizes opportunities and risks only to the limit the data supports and always as a range with stated assumptions, and retains the full list with reasoning visible so that nothing important is lost without a record. ## 2. What this skill is used for **The research problem it solves.** A study produces twelve insights and three matter. The gap between those numbers is where most research value is lost, in two opposite directions. Under-selection: everything is reported at equal weight, the report is forty slides, and the decision-maker picks whichever three they already believed. Over-selection: three are chosen by whoever was in the room on the day, the other nine vanish without a record, and two years later nobody can tell whether an important finding was rejected or simply never noticed. The usual attempted fix makes it worse. Someone builds a scoring model: eight criteria, a one-to-five scale each, weights, a total. The total is a number, so it looks like a decision procedure, and it quietly averages a low evidence score against a high magnitude score to produce a middling rank that hides the exact trade-off the decision-maker needed to see. Sizing goes the same way: a measured incidence gets multiplied by an unmeasured value, and a point estimate with no error bars enters a business case and is quoted for years. This skill supplies criteria that are applied as argument rather than arithmetic, a sizing discipline that stops where the data stops, and a retention mechanism so that deprioritised insights are parked rather than deleted. **Where it sits.** Late synthesis, after insights are explained and before the report is architected. It determines what leads, what supports, what is deferred, and what can honestly be quantified. **Typical use cases.** - A study with more insights than any audience will absorb, needing a defensible short list. - Deciding what belongs in an executive summary, where the selection does most of the work. - Sizing an opportunity or a risk for a business case, without inventing the number. - A study that answered the client's question and also found something more important that nobody asked about. - Insights that matter but not this quarter, which need somewhere to go other than a deleted slide. - Reviewing someone else's prioritisation, particularly where a scoring model has produced a rank order nobody can explain. **Who uses it.** Research directors deciding what a report leads with; insight managers preparing a board or executive read; evaluators with many outcome findings and one decision window; anyone building a business case from research data, where the sizing discipline matters more than the ranking. ## 3. When to use it - The insight set is complete and larger than the audience or the decision can absorb. - An executive summary or a board slide has to carry three things and there are twelve candidates. - Someone has asked how big an opportunity or a risk is, and the honest answer needs bounding. - The study found something outside the brief that may matter more than what was asked for. - Insights differ sharply in evidence strength and a rank order must not quietly promote the weakest. - A decision is imminent and the report must be organised around what bears on it. - A previous prioritisation produced a rank order that nobody can now reconstruct or defend. - Insights are being dropped in editing and there is no record of what was dropped or why. ## 4. When NOT to use it - **The insight set is incomplete.** Prioritising a partial set produces a rank order built on the accident of which analysis finished first, and the items added later never recover their place because the shape of the report is already set. Wait for the set, or prioritise explicitly as provisional and say so. - **The insights have not been developed.** Ranking findings is not the same as ranking insights, because a finding's importance cannot be judged before you know what it means. Run **08.01 Finding to Insight Development** first. Where the question is whether the findings even form a structure, that is **08.02 Pattern Identification**. - **The question is what follows for the organisation, or what to do.** Consequence is **08.03 Implication Development** and action is **08.04 Recommendation Development**. Prioritisation ranks insights; it does not turn them into anything. - **A sizing is wanted and the inputs do not exist.** Where there is no measured incidence on an adequate base, or no defensible population count, a size cannot be produced. Per K4 §2.1, the answer is to state what is measured, name what would be needed for the rest, and produce no number. A range built on two invented assumptions is not more honest than a point estimate; it is a point estimate with decoration. - **A single score is required by the process.** Where a client or a template demands one number per insight, this skill's method conflicts with it. Say why: incommensurable criteria averaged into a total produce a rank that cannot be explained and that systematically buries the evidence-strength gate. Offer the tiering and the written reasoning instead, which answers the same question and can be argued with. - **The prioritisation has already been decided and evidence is being arranged behind it.** Where the top three are fixed and the task is to justify them, this is confirmation search under K4 §4.2. Offer instead an explicit assessment of the chosen three against the criteria, including the insights that outrank them. - **The ranking would suppress a finding for political reasons.** An insight dropped because it is unwelcome is a K4 §4.1 failure, and dropping it silently is worse than arguing about it. Route to a human per K5 §2.1, with the insight and its evidence attached. - **Materiality depends on organisational knowledge nobody available holds.** Whether a segment is worth the company's attention is a K5 §2.1 judgement. Where nobody can make it, produce the tiering on the criteria that can be assessed, mark the rest as unassessed, and do not substitute a guess. ## 5. Required inputs **Required.** Without these the skill cannot run. If the first is absent, stop. - **The complete insight set**, each stated as an insight with its confidence level, the findings beneath it with their bases, and its evidence streams, per **08.01**. A partial set produces an unreliable rank order. - **The decisions the study informs**: what is open, who owns each, when it is taken, and how reversible it is. Decision relevance is the highest-weighted criterion and it cannot be assessed without this. Where no decision can be identified, say so and rank on the research objectives, marking the whole exercise for researcher review. - **The population and base structure of the study**: who each insight applies to, how many they are within the sample, and what population the sample represents. Both the reach criterion and any sizing depend on it. **Optional, and what each one adds.** - **A defensible population count for the group each insight concerns**, from client records, official statistics or an audited source: the difference between being able to size and not. Without it, reach is expressed in sample proportions only, which is honest and much less useful for a business case. - **Operational or transactional data held by the client:** allows a measured incidence to be checked against actual behaviour, which is the single strongest input a sizing can have, and the one that most often shrinks a number. - **What the organisation already knows and has acted on:** sets the novelty criterion, which is otherwise guesswork, and prevents an insight leading a report because it was new to the researcher. - **Cost, effort or capacity information for the actions each insight points to:** allows actionability to be assessed rather than assumed, and turns a rank order into something a delivery team can use. - **The decision timetable and funding cycle:** distinguishes urgent from important, which is what the deferred register at step 12 exists to preserve. - **The brief and its scope boundary:** identifies which insights were asked for and which were not, which matters at step 9 rather than as a reason to drop anything. ## 6. Questions to ask before starting 1. **Which decisions are open, who owns them, when are they taken, and how reversible are they?** Sets the primary criterion and the whole frame. Reversibility in particular changes how much evidence strength is required for an insight to lead. *Default if unanswered:* rank against the research objectives, state that no decision frame was supplied, and mark for review per K5 §2.1. 2. **How many insights can the audience actually absorb, and in what format?** Determines the length of the short list. A board reads three, a working team can use eight, and a repository holds all of them. *Default:* produce a ranked short list of three to five plus the retained full register. 3. **Is a size actually needed, and for what?** Determines whether sizing runs at all. Sizing for a business case has a different standard from sizing to convey rough scale in a sentence. *Default:* do not size unless asked; express reach in measured proportions with bases. 4. **What population counts are available, and how defensible are they?** Determines whether any sizing is possible. *Default:* express everything in sample terms and state that no population projection can be made. 5. **What does this organisation already know?** Sets the novelty criterion. Novel to the researcher is not novel to the client. *Default:* assume the well-informed reader, and mark novelty as unassessed. 6. **Is anything in this set unwelcome?** Identifies where prioritisation pressure will come from, and it is better identified before the ranking than during the edit. *Default:* proceed, and check at step 8 whether any low-ranked insight is low-ranked on the criteria or only on comfort. 7. **What happens to the insights that do not make the list?** Determines whether a deferred register is needed and who owns it. *Default:* build one, because without it the answer is that they disappear. ## 7. Step-by-step methodology **The position this method takes.** Prioritisation is a judgement, and the professional response to a judgement is not to disguise it as arithmetic. The eight criteria below are genuinely incommensurable: evidence strength and population reach are not the same kind of thing and there is no exchange rate between them. A scoring model supplies one anyway, and its output is a number that cannot be explained, that averages away the trade-off a decision-maker most needs to see, and that systematically advantages the insight with a big population and thin evidence over the one with a modest population and strong evidence. This method therefore uses the criteria as an argument structure: assess each one explicitly, use two of them as gates rather than as scores, tier the set, order within tiers, and write the reasoning for each placement so it can be contested. Sizing follows the same logic: go as far as the measurement carries you, state every assumption beyond it, and present a range. **1. Gate on inputs, and compile the full register first.** Confirm the insight set is complete and each item carries confidence, findings and bases. Then build the register of every insight before ranking anything, and fix it: this register is the permanent record, and nothing is deleted from it at any later stage. Insights that fall out of the short list are marked, not removed. This single discipline is what prevents the commonest institutional failure in research, which is a finding disappearing between the analysis and the deck with no trace that it ever existed. *Correct result:* a numbered register of every insight with confidence and base, which will still be complete at the end of the process. **2. Establish the decision frame.** For each open decision, record the owner, the date, the options, and how reversible it is (cheap and reversible, expensive and reversible, effectively irreversible). Reversibility does real work here: an insight bearing on an irreversible decision needs stronger evidence before it leads, because the cost of being wrong is asymmetric. *Correct result:* a decision map that each insight can be assessed against. **3. Assess criterion by criterion, across all insights, not insight by insight.** Take one criterion and judge every insight on it before moving to the next. Assessing insight by insight produces halo: an insight that scores well on the first criterion gets read generously on the rest. Column-wise assessment breaks that, and it also forces genuine comparison, which is what a rank order is. The eight criteria, each with the question it asks, are in Section 8. *Correct result:* eight completed columns, each judged in one pass. **4. Record each assessment as a short verbal judgement with its evidence, not as a number.** "High: bears directly on the timetable decision taken in eleven weeks" is a usable assessment. "4" is not, because it cannot be argued with and its meaning is gone within a week. Use three levels (high, moderate, low) plus "unassessed" where the input was unavailable, and require a clause of reasoning with each. Unassessed is a legitimate and important value: it says the criterion could not be judged rather than that it scored low, and those are frequently confused to the disadvantage of insights nobody has organisational knowledge about. *Correct result:* a matrix of verbal judgements, each with a reason, and explicit "unassessed" cells. **5. Apply the two gates before ranking.** Two criteria are gates rather than contributors. **Evidence strength:** a low-confidence insight cannot lead a report on the strength of its magnitude, however large that magnitude appears. It is routed to validation instead, per K3 §4.3, and it can be reported as a labelled hypothesis without occupying the top of the list. **Decision relevance:** an insight bearing on no open decision does not lead, however elegant it is; it goes to the deferred register with a trigger. Treating these two as scores to be traded off against the others is the specific failure a scoring model produces, and it is how an under-evidenced, high-magnitude claim reaches an executive summary. *Correct result:* the set divided into those that pass both gates and those routed elsewhere, with the routing recorded. **6. Tier the survivors.** Sort into three tiers by decision relevance against evidence strength: **lead** (directly bears on an open decision, strong or well-caveated moderate evidence), **support** (bears on a decision but with weaker evidence, or strong evidence on a decision that is less immediate), and **context** (true, evidenced, and not decision-bearing this cycle). Tiering rather than a continuous rank is deliberate: the difference between rank three and rank four is usually noise, while the difference between tiers is real and is the distinction a reader needs. *Correct result:* three named tiers with membership, not a list of twelve positions. **7. Order within tiers on the remaining criteria.** Inside the lead tier, order by magnitude of effect and reach (how many people, how much changes), using actionability and urgency as tiebreakers. Write the placement reason for each item as a sentence. If a placement cannot be justified in a sentence, it is not a judgement, it is a preference. *Correct result:* an ordered short list where every position has a stated reason. **8. Stress-test the order with swap tests.** For each of the top three, ask what would have to be true for it to be fourth. For the highest-placed item that did not make the list, ask what would promote it. These two questions expose rank orders driven by one criterion, and they expose the ranking that was really about comfort: if the only thing keeping an insight low is that it is unwelcome, the swap test makes that visible, and per K4 §4.1 it does not stay low for that reason. *Correct result:* a written answer for each of the top three and for the highest near-miss, with any change made and recorded. **9. Handle the unasked-for finding on its merits.** An insight outside the brief's scope is assessed on the same eight criteria as everything else. If it clears them, it is reported, in a labelled section that says plainly it was not commissioned, why it is being raised, and what it would take to establish it further. Dropping an important finding because it was out of scope is one of the quietest failures in commissioned research: the scope was set before anyone knew what the study would find, and a study that only answers the questions it was given cannot tell an organisation anything it did not already suspect. Where the finding is uncomfortable as well as unasked-for, route it to a human per K5 §2.1 with the evidence attached, but route it, do not delete it. *Correct result:* out-of-scope insights ranked on merit, with the ones that clear the gates present and labelled. **10. Size only what the data can carry.** A sizing is a chain, and it is only as strong as its weakest link. The chain runs: **measured incidence** (from the study, with its base and its coverage limits) → **defined population** (from a defensible external or client source, named and dated) → **stated assumptions for every step between** → **range**. Three rules govern it. First, every number is either measured, with a source and a base, or an explicitly labelled assumption; nothing in between, and no arithmetic silently joining the two. Second, present a range, produced by varying the two most sensitive assumptions across a defensible interval, never a point estimate: a point estimate on a modelled quantity claims a precision the method cannot deliver and is the form most likely to be quoted back years later, per K3 §4.4. Third, name the unit honestly: appointments released is not appointments filled, intention is not behaviour, and neither is money. Where a required input does not exist, stop the chain there and say what is missing. *Correct result:* a sizing with every assumption listed, a range rather than a figure, and the unit stated precisely. **11. Separate the size of a population from the size of an opportunity.** These are routinely conflated and the gap between them is usually large. Four things sit in between. **Incidence against behaviour:** what people report is not what they do, and the discount between them is an assumption, not a measurement. **Total against addressable:** the organisation can reach some of the affected population, not all of it, and the reachable share is frequently the binding constraint. **Intention against conversion:** stated willingness converts at a rate this study did not measure. **Gross against net:** some of the outcome would have happened anyway, and some of the recovered value is not capturable. State each of the four explicitly for any opportunity claim, with what is measured and what is assumed. An opportunity stated as the population multiplied by an incidence has skipped all four. *Correct result:* an opportunity statement that names the four gaps and shows where the number lost its measurement. **12. Route important-but-not-urgent insights to a deferred register with a trigger.** An insight can be well-evidenced, genuinely important, and irrelevant to any decision in this cycle. Deleting it wastes the study; leaving it in the lead tier crowds out what is actionable now. The mechanism is a deferred register holding the insight, its evidence, its confidence, an owner, and a **trigger**: a date, an event, or a decision that would make it live (a system replacement, a contract renewal, a planning round, a regulatory change). A deferred insight with no trigger and no owner has been deleted with extra steps. *Correct result:* a deferred register in which every entry names what would bring it back and who holds it. **13. Produce both outputs: the ranked short list and the retained full register.** The short list carries the reasoning for each placement. The full register carries every insight, its criterion assessments, its tier or routing, and the reason it is where it is. Both go into the deliverable set, even if only the short list goes into the deck. The register is what allows a future reader, or the same client eighteen months later, to see that an insight was considered and placed rather than never noticed. *Correct result:* two documents, consistent with each other, with nothing present in one that is missing from the other. **14. Mark the judgement points and hand over.** Per K5 §2.1, materiality is a human judgement: whether a segment is worth this organisation's attention, whether a difference is large relative to what the business can act on, whether this is the third study to say the same thing. Mark it where it bites. Per K5 §2.5, any sizing that will enter a business case carries sign-off, because a number in a business case is a high-stakes claim regardless of how carefully it was bounded. Hand the ranked list to **12.01 Research Report Architecture** and **11.02 Executive Summary Development**, with confidence levels and ranges intact. *Correct result:* a prioritisation another researcher could audit, and a sizing nobody can quote without its assumptions. ## 8. Analytical framework **The eight criteria, with the question each asks.** Two are gates. Six are ordered against each other by argument, never averaged. | Criterion | The question it asks | Role | |---|---|---| | **Decision relevance** | Does this bear on a decision that is open, owned and dated? | **Gate** | | **Evidence strength** | What is its confidence, on what bases, from how many independent streams? | **Gate** | | **Population affected** | How many people does this concern, as a share of the relevant population? | Ordering | | **Magnitude of effect** | How much changes for those it affects? | Ordering | | **Actionability** | Is there something the organisation could do about it, within its control? | Ordering | | **Urgency** | Does the value of acting decay with time, and is there a window? | Ordering | | **Novelty to this organisation** | Is this new here, or a confirmation of something already acted on? | Ordering | | **Reversibility of the decision it informs** | How costly is it to be wrong on the decision this bears on? | Modifier on evidence strength | **Why these are not averaged.** Evidence strength and population reach have no exchange rate. Averaging them lets a weakly-evidenced insight about a large group outrank a well-evidenced insight about a smaller one, and produces a rank a decision-maker cannot interrogate. The reversibility criterion makes this concrete: for an irreversible decision, evidence strength should dominate; for a cheap reversible one, magnitude can. A single weighted total cannot express that, because the weights would have to change per decision. State this refusal explicitly in the output, because someone will ask for the score. **The tiering structure.** ``` Evidence strong Evidence moderate Evidence low Bears on open decision LEAD LEAD, caveated VALIDATE (not lead) Decision less immediate SUPPORT SUPPORT DEFER with trigger No open decision DEFER with trigger DEFER with trigger REGISTER only ``` **The sizing chain.** Each arrow is a place where measurement can stop. ``` Measured incidence (base, coverage) → Defined population (source, date) → Stated assumptions, one per step → Sensitivity on the two weakest → RANGE, with the unit named ``` **The four gaps between population size and opportunity size.** | Gap | What it is | What it costs the number | |---|---|---| | Incidence against behaviour | Reported is not done | An unmeasured discount | | Total against addressable | Not all of the population is reachable | Usually the largest single reduction | | Intention against conversion | Willingness converts at an unmeasured rate | An assumption, always | | Gross against net | Some would have happened anyway | The difference between a size and a gain | **The deferred register entry.** ``` INSIGHT | EVIDENCE | CONFIDENCE | WHY DEFERRED | TRIGGER | OWNER | REVIEW DATE ``` ## 9. Output format **A. Ranked short list** Three to five insights, in order, each with: the insight, its confidence, its tier, the criteria that placed it there in a sentence, and its decision. This is the page that drives the report. **B. Full insight register** Every insight, with its eight criterion assessments (verbal, with reasons, and "unassessed" where applicable), its tier or routing, and its placement reason. Required, and it goes into the deliverable set whether or not it goes into the deck. **C. Sizing statements** (where sizing was requested and possible) ``` WHAT IS BEING SIZED The unit, precisely. Appointments released, not appointments filled MEASURED INPUTS Each with source, base and coverage limitation POPULATION Source, date, and how defensible ASSUMPTIONS Every one, numbered, with its basis and its plausible interval RANGE Low to high, from varying the two most sensitive assumptions WHAT THIS IS NOT The gaps between this and a financial figure CONFIDENCE Per K3, with the limiting factor SIGN-OFF Per K5 section 2.5 where it enters a business case ``` **D. Deferred register** Important, not urgent, with trigger, owner and review date for each entry. **E. Insights outside the brief** Labelled, ranked on the same criteria, with a note that they were not commissioned. **F. The refusal note** A short statement of why no composite score was produced, so that the absence is a stated methodological position rather than an omission. **When the evidence is thin.** Where most insights are low confidence, the lead tier is small or empty and the output says so: the honest result is a short list of one, plus a validation proposal for the rest. Where sizing inputs are missing, section C states what is measured and what would be needed, and contains no number, per K4 §2.1. Do not produce a five-item short list because five is the expected number, do not promote a low-confidence insight to fill the lead tier, and do not convert a missing population count into an estimate to complete a sizing. A ranked list of two, well-evidenced and decision-connected, does more for a decision-maker than five where they cannot tell which two to trust. ## 10. Quality checks Run before anything is presented. Sits on top of K4 §8. 1. Is the full insight register complete, and does every insight in it appear in exactly one tier or routing? 2. Was the assessment done column-wise, one criterion at a time across all insights? 3. Is every criterion assessment a verbal judgement with a reason, rather than a number? 4. Are "unassessed" cells marked as unassessed rather than treated as low? 5. Were evidence strength and decision relevance applied as gates rather than traded off against magnitude? 6. Does any low-confidence insight lead the report on the strength of its size? 7. Is the placement reason for every short-list item written as a sentence that could be contested? 8. Were the swap tests run on the top three and the highest near-miss, and recorded? 9. Is any insight ranked low because it is unwelcome rather than because of the criteria? 10. Were out-of-scope insights assessed on the same criteria, and are those that cleared them present and labelled? 11. In every sizing, is each number either measured with a base or explicitly labelled as an assumption? 12. Is every sizing presented as a range, with the two most sensitive assumptions named? 13. Does every opportunity statement address the four gaps between population size and opportunity size? 14. Is the unit of every size stated precisely, and distinguished from money? 15. Does every deferred insight have a trigger and an owner? 16. Is the refusal to produce a composite score stated, so the absence is deliberate? 17. Are K5 §2.1 materiality markers placed where they bite, and does any business-case sizing carry sign-off? ## 11. Common failure modes | Failure | How to recognise it | How to prevent it | |---|---|---| | **The composite score** (the signature failure) | Eight criteria averaged into one number that drives the order and cannot be explained | Gates plus tiers plus written reasoning. State the refusal in the output | | **Silent disappearance** | An insight is in the analysis and not in the deck, with no record of a decision | The full register, fixed at step 1 and never deleted from | | **Halo in assessment** | An insight rated highly on every criterion, including ones nobody had evidence for | Column-wise assessment, one criterion across all insights | | **Unassessed read as low** | Insights about which the organisation supplied nothing sink quietly | An explicit "unassessed" value, distinguished from low in the output | | **Magnitude beating evidence** | A large, weakly evidenced claim leads the summary | Evidence strength is a gate, not a score | | **Scope obedience** | An important finding dropped because it was not in the brief | Step 9. Scope was set before anyone knew what the study would find | | **The comfortable ranking** | The unwelcome insight is fourth, and the reason cannot be stated | The swap test at step 8, and K4 §4.1 | | **Population as opportunity** | The size is the affected population multiplied by an incidence | The four gaps at step 11, each stated | | **The point estimate** | One number, no range, quoted for years | Always a range, from varying the two most sensitive assumptions, per K3 §4.4 | | **Assumption laundering** | A measured figure and an assumed multiplier combined into one number with a single source | Every number labelled measured or assumed. No arithmetic across the boundary | | **Unit drift** | "Appointments released" becomes "appointments filled" becomes "savings" between the analysis and the summary | Name the unit in every restatement, per K2 §7 | | **Deferral as deletion** | A deferred register with no triggers, owners or review dates | Every entry names what brings it back and who holds it | | **Template-length short list** | Exactly five, regardless of the evidence | The number is set by the tiering, and a lead tier of one is a legitimate result | | **Novelty measured on the researcher** (the signature AI failure) | Insights ranked as new because they were new to the analyst, with no reference to what the organisation knows | Novelty is assessed against the organisation's existing knowledge, or marked unassessed | ## 12. AI guardrails Skill-specific only. Universal prohibitions are inherited from **K4**; traceability follows **K2**; confidence language follows **K3 §4**; materiality judgement follows **K5 §2.1**. 1. **Never average the criteria into a single score.** They are incommensurable, the total cannot be explained, and it buries the evidence-strength gate. Where a score is requested, explain why one is not produced and supply the tiering instead. 2. **Never delete an insight from the register.** Deprioritised insights are marked and retained. The register is the record that a judgement was made. 3. **Never let a low-confidence insight lead on the strength of its magnitude.** It is routed to validation, per K3 §4.3. 4. **Never produce a size without a measured incidence on an adequate base and a defensible population count.** Where either is missing, state what is measured, name what would be needed, and produce no number. 5. **Never present a modelled size as a point estimate.** A range, from varying the two most sensitive assumptions, with the interval and its basis stated. 6. **Never combine a measured figure and an assumed multiplier into a single number without labelling which is which.** Per K4 §2.1, this is the specific route by which invented quantities acquire the authority of measurement. 7. **Never equate population size with opportunity size.** State the four gaps and where the number stops being measured. 8. **Never change the unit between statements.** Released is not filled, incidence is not behaviour, and none of them is money. 9. **Never drop an insight because it was outside the brief.** Assess it on the criteria and report it, labelled, if it clears them. 10. **Never rank an insight down because it is unwelcome.** Where the ranking is contested for that reason, route it to a human with the evidence attached, per K5 §2.1. 11. **Never defer an insight without a trigger and an owner.** Deferral without either is deletion with a longer paper trail. 12. **Never assess novelty against your own knowledge.** Novelty is relative to what this organisation already knows and acts on, or it is unassessed. ## 13. Best-practice principles 1. **Refuse the single number, and say why.** The refusal is a methodological position, not an omission, and stating it prevents someone reinstating the score after you leave. 2. **Two criteria are gates.** Evidence strength and decision relevance decide whether an insight can lead at all. The other six decide the order among those that can. 3. **Reversibility sets the evidence bar.** A cheap reversible decision can be informed by moderate evidence; an irreversible one cannot, and the same insight therefore ranks differently against different decisions. 4. **Assess down the columns, not across the rows.** It is the cheapest available defence against halo and it forces the comparison that ranking actually is. 5. **Unassessed is not low.** Conflating them penalises exactly the insights nobody has organisational knowledge about, which are often the ones worth attention. 6. **Keep everything, forever.** The full register is what lets someone eighteen months later see that a finding was weighed rather than missed, and it costs one page. 7. **The out-of-scope finding is often the most valuable thing in the study.** The brief was written before anyone knew what would be found. A study that answers only the questions it was given cannot surprise the organisation, which is most of what it was for. 8. **A range is not a hedge.** It is the honest shape of a modelled quantity, and it survives scrutiny that a point estimate does not. 9. **Name the unit every single time.** Unit drift is how a careful sizing becomes a financial claim between the analysis and the board pack, and nobody can point to where it happened. 10. **The addressable population is usually the binding constraint.** Of the four gaps, this is the one that most often turns a large opportunity into a modest one, and it is the easiest to check. 11. **Deferral needs a trigger.** Without one, "we will come back to this" is a polite deletion, and everybody in the room knows it. 12. **A short list of two is a result.** The number of insights that matter is a property of the study, not of the slide layout. ## 14. Worked example *Fictional scenario, used to demonstrate method. All organisations, figures and findings below are invented.* **INPUT.** A regional health provider commissions research into missed outpatient appointments, where the non-attendance rate is 13%. The study produces twelve insights. The open decision is which service changes to fund in the next improvement cycle, owned by the outpatient services director, taken in ten weeks, and reversible at moderate cost. **PROCESS.** *Steps 1 to 4.* All twelve insights are registered with their confidence and bases. The decision map records one immediate decision and one longer-term one (a booking system replacement, eighteen months away). Assessment runs column-wise. Novelty is marked **unassessed** for four insights because nothing was supplied about what the provider already knows, and this is recorded as unassessed rather than allowed to sink them. *Step 5, the gates, and the judgement call.* The obvious move was a scoring model: eight criteria, five points each, weighted, totalled. It was refused, and the reason is worth stating. Insight I9 concerns patients requiring an interpreter and rests on a base of 28, which under K3 §3.2 caps it at low confidence. It has a large apparent effect. In a weighted total it would have placed second and led a board summary on a base of 28. Under the gate, it does not lead; it is routed to validation with a specified follow-up. The composite score would have produced exactly the failure the gate exists to prevent, and the refusal is written into the output so nobody reinstates the model later. *Steps 6 to 8.* Three insights reach the lead tier. **I3**, that patients who must arrange transport treat a reminder as a prompt to reconfirm transport, and where they cannot they do not cancel because cancelling feels like forfeiting their place. **I7**, a threshold effect: non-attendance steps up for appointments booked more than nine weeks ahead. **I11**, that the referral letter's wording leads a subset of patients to read attendance as optional. Swap test on I7: it would drop to fourth only if the booking horizon were outside the provider's control, which the operations pack shows it is not. Swap test on the highest near-miss, I5, about staff communication during the future system replacement: nothing about the current cycle promotes it, which routes it to the deferred register with the system go-live as its trigger. *Step 9, the unasked-for finding.* I11 was outside the brief, which covered reminders and booking only. It was assessed on the same criteria, cleared both gates, and is reported in a labelled section noting it was not commissioned. It is also mildly uncomfortable, since the referral letter is another department's, so it carries a K5 §2.1 marker rather than a softened wording. *Steps 10 and 11, the sizing.* Asked how big the transport issue is. The chain: 13% non-attendance (provider records, 148,000 appointments a year, client-supplied); 31% of surveyed non-attenders gave transport as the reason they did not travel (survey of non-attenders, n=210); of those, 62% said they would have cancelled had cancelling not felt like forfeiting the slot (n=65, small base, flagged). The naive product is about 3,700 appointments a year, and it was not produced. Four problems were stated instead: the 62% rests on a base of 65; the survey reached only non-attenders who could be contacted, which is a coverage limitation, not a sampling one; "would have cancelled" is stated intention, not behaviour; and a cancelled appointment is only released, not filled, since the short-notice refill rate was never measured. The output varies the two most sensitive assumptions, the intention-to-behaviour discount and the refill rate, across defensible intervals, and reports **roughly 900 to 3,000 appointments a year potentially released**, with both assumptions numbered and sourced, and an explicit line that released is not filled and neither is money. The provider's own booking data would narrow it, and that is named. **OUTPUT.** A three-item ranked short list with a placement sentence each, the full twelve-insight register with column-wise assessments and four unassessed novelty cells, one sizing statement as a range with five numbered assumptions, a deferred register with two entries and their triggers, one labelled out-of-scope insight, a validation proposal for I9, and a stated refusal note explaining why no composite score was produced. ## 15. Advanced usage **Prioritising against several decisions at once.** Where a study informs three decisions with different owners and timetables, produce one tiering per decision rather than one master rank. The same insight legitimately leads for one decision and sits in context for another, and a single list forces a false compromise. This is more work by roughly a page and it is what makes a report usable by more than one function. **Sizing a risk rather than an opportunity.** The chain is the same and the asymmetry is not. For a risk, the range's upper bound does more work than its midpoint, because the decision is usually about exposure rather than expectation. Present the upper bound with its assumptions prominently, and state the conditions under which it would be reached, rather than leading with a central figure that understates what the decision-maker needs to plan against. **When the client's question and the study's answer diverge.** Occasionally the most important insight makes the commissioned question less relevant. Handle it by reporting both: answer what was asked, at the length it merits, and then present the reframing with its evidence and its criterion assessments. Do not simply substitute your question for theirs, and do not bury the reframing at the back. Route the judgement of how hard to press it to a human, per K5 §2.1. **Maintaining the register across studies.** A deferred register that spans studies becomes an organisational asset: it accumulates evidenced insights waiting for their trigger, and it makes visible the pattern of what an organisation repeatedly learns and does not act on. Review it at each new brief. Re-date and re-check entries rather than re-quoting them, per K2 §7, because an insight whose trigger has finally arrived is often three years old and may no longer hold. **Where a scoring model is contractually required.** Produce the tiering and the reasoning as the primary output, and where a score must be supplied, derive it from the tier rather than the reverse, state that it is a label for the tier and not a computation, and keep the gates outside it. This prevents the number acquiring an authority the method does not support, which is the only real risk in the exercise. ## 16. Skill chain **Recommended previous skills:** - **08.01 Finding to Insight Development.** Hands over explained, confidence-rated insights with their chains and bases. This is the required predecessor. - **08.02 Pattern Identification.** Establishes which findings form structures, which determines the reach and scope of what is being ranked. - **05.02 Cross-Tabulation and Segmentation Analysis.** Supplies the base and population structure that the reach criterion and any sizing depend on. - **01.01 Research Objective Definition.** Supplies the decision frame, the scope boundary and the owner, without which decision relevance cannot be assessed. **Recommended next skills:** - **12.01 Research Report Architecture.** Takes the ranked short list as the report's spine, with the full register retained as an appendix so nothing disappears. - **11.02 Executive Summary Development.** Takes the lead tier with confidence levels and ranges intact. Per K2 §5, every claim appearing there must be in the traceability map. - **08.03 Implication Development** and **08.04 Recommendation Development.** Take the prioritised insights, so that implication and recommendation effort goes to what will actually be reported. **Runs well alongside:** - **13.03 AI Output Verification.** The sizing chain and the register completeness are the two checks it should apply here. - **13.04 Bias Detection**, particularly where a ranking places an unwelcome insight low. - **K5**, at the materiality judgement this skill mandates per §2.1, and at sign-off for any sizing entering a business case per §2.5. --- A Yazi Supplied Skill and resource.