The Model Thinker: What You Need to Know to Make Data Work for You

1. 1State your assumptions explicitly before reasoning: name the objectives, actions, and choice mechanism for each actor. Write them as formal descriptions (a probability distribution, a ranking, a mathematical function) — not loose terms like 'beliefs' or 'preferences' left undefined.
2. 2Choose an actor model deliberately by stakes and purpose: fixed-rule for low-stakes behavior, adaptive-rule when people learn and respond to others, rational-actor for high-stakes decisions by sophisticated people. Match the assumption to whether you are predicting, explaining, evaluating, or exploring.
3. 3Trace each conclusion back to a specific assumption. You are 'tied to the mast of logical coherence' — you cannot manufacture implications. Every implication must be derivable from what you assumed, not smuggled in.
4. 4Test tractability: can you write the reasoning in a formal language (math or code) and prove each step? If a step cannot be justified, it is a gap. Apply Ockham's Razor — as simple as possible, but not simpler.
5. 5Run the derivation to its forced consequences and accept them. If you assume strong social influence in consumer choice, you must accept a few products with large market shares; if you assume information flows through networks, you must accept that people filling structural holes hold power.
6. 6Check for gaps, tautologies, and inconsistencies by having a second reader hunt specifically for logical flaws and muddled thinking (Jenna Bednar spent two weeks doing exactly this on the manuscript). Isolate the formal argument so it can be inspected on its own.
7. 7When one model's chain feels narrow, add a second logically consistent model rather than patching the first with ad hoc claims. Keep actor assumptions coherent across models so implications can be combined without contradiction.

Take the communicable-disease model of infected, susceptible, and recovered people. You begin from one explicit assumption: a fixed rate of contagion. From that single assumption, logic forces a contagion threshold — a tipping point above which the disease spreads. Nothing is added by intuition; the threshold is deduced. The same assumptions then let you derive the proportion of people who must be vaccinated to halt spread. Because each step follows from the stated rate, the chain is auditable: change the contagion rate and the threshold moves in a determinate way. This is logical coherence in action — you cannot claim a lower vaccination target than the assumptions produce.

• If a term like 'preferences' or 'beliefs' appears without a formal representation, then stop and represent it (ranking, function, or probability distribution) before reasoning further.
• If the environment is low-stakes, then assume fixed-rule actors; if people learn and watch others, assume adaptive actors; if stakes are high and actors sophisticated, assume rational actors.
• If an implication cannot be traced to an explicit assumption, then discard it — you cannot manufacture implications.
• If a step in the argument cannot be written in math or code and proven, then treat it as a reasoning gap, not a conclusion.
• If a single model's chain ignores a feature (income disparity, interdependency, diversity), then add a separate coherent model rather than bolt an unjustified clause onto the first.

• Making ad hoc assumptions per model so each requires fresh, incompatible reasoning about how people act — the tell is an 'idiosyncratic set of constructions' you cannot combine or reason across.
• Smuggling in a desired conclusion the assumptions do not support — the tell is an implication you can state but cannot derive step by step.
• Treating the rational-actor model as universally descriptive — the tell is prediction failures where the model was actually a straw man or a design tool, not a forecasting tool.
• Over-simplifying past Ockham's line ('simpler than possible') so the model drops the causal force that mattered — the tell is a clean derivation that contradicts the observed data.
• Confusing intuition for coherence and skipping the formal check — the tell is that a second reader hunting for logical flaws finds gaps, tautologies, and muddled inferences you missed.

• ▸Claim to defend: state the single conclusion you want to reach.
• ▸Explicit assumptions: list each one in formal terms (function, ranking, rate, distribution).
• ▸Actor model + justification: fixed / adaptive / rational, and why given stakes and purpose.
• ▸Derivation steps: number each inference and cite which assumption it follows from.
• ▸Unsupported implications flagged: mark any conclusion not traceable to an assumption.
• ▸Tractability check: can each step be written in math/code and proven? (Y/N per step).
• ▸Ockham check: is any assumption removable without changing the conclusion?
• ▸Second-reader flaws: gaps, tautologies, inconsistencies a reviewer found.
• ▸Cross-model consistency: do actor assumptions match those in models you'll combine with?

1. 1State the claim or proverb in its unconditional form as it was handed to you (e.g., 'lowering income taxes increases government revenue' or 'big rocks first').
2. 2Build the smallest model that captures the mechanism the claim asserts. For revenue, write revenue = tax rate × income; for task ordering, name the assumptions about task size and bin capacity from bin packing.
3. 3Derive the precise inequality or relationship the model forces. Do not stop at direction; get the threshold (revenue rises only if % income growth > % tax cut).
4. 4Read off the boundary explicitly: name the condition under which the claim flips from true to false, and state both sides of that boundary.
5. 5Contrast this conditional result against the narrative version, even when the narrative has empirical support, to expose that the proverb is unconditional while the derived rule is bounded.
6. 6Run a coarser and a finer version of the model and compare their predictions; the differences show which assumptions the conclusion actually rides on.
7. 7Check whether the calibrating conditions still hold over time — as with WORLD3, results decay when reality stops matching assumptions, so restate the claim as valid only within its current parameter range.

Take the politician's claim that cutting income taxes raises government revenue. Model it minimally: revenue equals the tax rate times the income level. Derive the implication and you find revenue increases only if the percentage growth in income exceeds the percentage cut in taxes. So a 10% cut in income taxes raises revenue only if it causes income to grow by more than 10%. The claim is not false — it is conditional, and the model names the exact boundary (growth > cut) inside which it holds. Compare this to the narrative 'first things first / big rocks first,' which is inferred from expert observation but stated unconditionally; a bin-packing model instead makes assumptions about object sizes and bin capacity and derives when big-rocks-first is actually optimal.

• If a claim asserts a direction of effect (X raises Y), then derive the threshold at which it reverses before accepting it — direction without a boundary is incomplete.
• If a rule comes from expert observation or narrative rather than derivation, then treat it as unconditional and suspect until you can restate its conditions from a model.
• If two models of different granularity give different predictions, then the difference tells you which assumption the conclusion depends on — investigate that assumption, not the average.
• If a model was calibrated to past data and its predictions now drift, then its conditions no longer hold; re-scope the claim rather than discard the mechanism.
• If empirical support exists but the mechanism is unmodeled, then keep the claim provisional — evidence confirms a pattern, not the conditions of its validity.

• Accepting the direction of an effect without its threshold. The tell: you can state 'cutting taxes raises revenue' but cannot say by how much income must grow (>10% for a 10% cut) for it to hold.
• Treating a proverb as a law. The tell: you apply 'big rocks first' to every task set without asking about task sizes or capacity, the very assumptions a bin-packing model would fix.
• Letting empirical support substitute for conditionality. The tell: you cite that the claim 'has data behind it' but cannot name the boundary within which the data were generated.
• Building only one granularity. The tell: you never construct a coarser comparison model, so you cannot see which assumption drives your result — you mistake a fragile conclusion for a robust one.
• Ignoring drift in calibration conditions. The tell: like WORLD3 fertility predictions, your model keeps being applied after population/parameters shifted and reality stopped matching the assumptions.

• ▸Unconditional claim: write the proverb/result exactly as received, no hedging
• ▸Asserted mechanism: one sentence on why the claim is supposed to hold
• ▸Minimal model: the variables and the equation/relationship (e.g., revenue = rate × income)
• ▸Derived threshold: the precise inequality that must hold (e.g., %income growth > %tax cut)
• ▸Boundary statement: name the condition where the claim flips from true to false
• ▸Coarse-vs-fine check: what differs between a simpler and more detailed version, and which assumption that isolates
• ▸Calibration status: are the conditions/parameters still matching current reality? (yes/no + note)
• ▸Re-scoped claim: rewrite the original claim as valid only within its stated conditions

1. 1State the model you are using and name what it deliberately leaves out. If your model of a person assumes full rationality, write down that it ignores diversity, social influence, cognitive bias, purpose, and adaptation — these are the wrongness you accept for tractability.
2. 2Set expectations proportional to system complexity before you predict. For a complex social system, declare up front that even an ensemble may explain only a small share of variation, so you commit to identifying the set of possibilities rather than a point forecast.
3. 3Build a second model that includes a feature the first omitted (social influence, positive feedback, a cognitive bias). Run both and compare, so you can discern whether that feature matters and by how much.
4. 4Take the model to data to test, refine, and improve it — do not stop at logical coherence. When data is thin, use an austere model; when data is rich, use a richer behavioral model.
5. 5Check robustness across your ensemble: if independent models with different assumptions reach the same classification, treat that as a robust theorem (Levins: 'the intersection of independent lies'). If they disagree, keep the disagreement visible rather than averaging it away.
6. 6When a wrong prediction arrives, treat it as fuel: locate which omitted feature broke the model, add a model that carries that feature, and re-run. Do not let 'all models are wrong' stop the work — treat it as the reason to build a crowd.
7. 7Distinguish which issues get superficial treatment and which get deep engagement, and say so. Page treats opioids qualitatively as a template and inequality with full data engagement — declaring the depth level keeps you from overselling.

Page applies many-model thinking to the opioid epidemic explicitly as a 'superficial' template: he gathers no data and calibrates no models, applying them only qualitatively to gain insight into why the crisis was hard to solve and how experts might have anticipated it. The humility move is the disclaimer itself — he states outright that he does 'not want to oversell the potential for models to avoid disaster.' Contrast this with his treatment of income inequality, where he engages a variety of models in detail and connects tightly to the academic literature — the opposite extreme of engagement. The same disposition shows in the music-lab download study: one strong result (social influence lengthens the tail and the big winners are not the best) still earns the caution 'we must be careful not to draw too strong an inference from a single study.' In both cases the model produces insight and the humility fixes the reach of that insight to the evidence actually held. That pairing — use the model, bound the claim — is the discipline.

• If you can only name possibilities, not a point prediction, then still ship the model — knowing what could happen is a benefit.
• If data is sparse, then use an austere model and focus attention elsewhere; if data is rich, then move to a richer behavioral model.
• If independent models with different assumptions agree, then raise confidence sharply; if they disagree, then keep both and report the spread.
• If a result rests on a single study or an uncalibrated model, then attach the caution explicitly and label the analysis superficial.
• If a model omits a feature you suspect matters, then build a second model containing it rather than arguing about the first.

• Overselling insight as prediction. Tell: you present a qualitative, uncalibrated analysis (like the opioid template) as if it forecasts outcomes.
• Treating 'all models are wrong' as an excuse to stop. Tell: you abandon modeling in the face of complexity instead of building more models to probe interdependencies.
• Drawing strong inference from one result. Tell: a single music-lab-style study becomes a general law with no caveat about sample or replication.
• Averaging away disagreement to look decisive. Tell: you collapse a diverse ensemble into one number and hide the diversity that carried the information.
• Assuming identical, fully rational agents by default. Tell: your human model has no room for diversity, adaptation, bias, or social influence, and you are surprised when behavior deviates.

• ▸Model in use: name the single model and its core assumption
• ▸Known omissions: features it deliberately ignores (bias, feedback, social influence, purpose)
• ▸Depth level: superficial template or deep data-engaged — state which
• ▸Data status: none / sparse / rich, and how it constrains claims
• ▸Claim boundary: point prediction, set of possibilities, or classification only
• ▸Ensemble check: do independent models agree or disagree, and how far
• ▸Companion model: which omitted feature the next model will add
• ▸Refinement trigger: the observation or mistake that will force a rebuild

1. 1List every plausible explanation for the phenomenon before committing to one. For the 2008 collapse, that meant enumerating foreign investment, over-leverage, weak mortgage oversight, home-flipper optimism, instrument complexity, misunderstood risk, and expected bailouts — not stopping at the first that fits.
2. 2For each account, name the causal force it accentuates and the model that formalizes it (supply and demand, path dependence, contagion), so each explanation becomes a testable frame rather than a slogan.
3. 3Run an internal-consistency check on each account: does it make logical sense given the actors' incentives? Lo rejected the 'knowing bubble' story because investors would not fund a collapse they foresaw.
4. 4Calibrate and test magnitudes against data. Lo checked leverage ratios and found 2008 levels were not much higher than 1998 — so leverage alone cannot carry the explanation.
5. 5Look for the disconfirming fact each single model ignores. Toxic mortgage bundles found buyers (so collapse was not foreseen); Lehman failed with no bailout (so 'expected bailout' fails). Use these to prune, not to pick.
6. 6Correct for statistical artifacts before inferring cause. Ask whether an outlier reflects a real effect or just small-sample variance — divide the standard deviation by the square root of sample size before concluding.
7. 7Evaluate models to choose the best one; do not shop for the model that ratifies a decision you have already made. Ball tested two models and picked supply and demand — he did not reverse-engineer a justification.
8. 8Weave the surviving models into a lattice: state which conditions each holds under, where they overlap, and where they interweave, then build the nuanced explanation from the ensemble.

Andrew Lo applied many-model thinking to twenty-one accounts of the 2008 collapse and found each individually lacking. The 'greedy bankers knew about the bubble and expected a bailout' account fails internal consistency: it does not make sense that investors would fund a bubble they knew would trigger a global crisis, so the bubble's extent must have surprised many. The 'toxic mortgages' account is undercut by the fact that clearly low-quality bundles found buyers — buyers who would not have existed had collapse been a foregone conclusion. The 'over-leverage' account fails a magnitude test: leverage had risen since 2002 but was not much higher than in 1998. The 'expected bailout' account is refuted by data: Lehman Brothers, with over $600 billion in holdings, collapsed on September 15, 2008, in the largest US bankruptcy, and the government did not intervene. Because the data privileges no single explanation, Lo's conclusion is to entertain as many interpretations of the same objective facts as possible and let a more internally consistent understanding emerge — no single model suffices.

• If an explanation requires actors to knowingly act against their own interest (fund a bubble they foresee collapsing), then treat it as internally inconsistent and demand a mechanism that resolves the gap.
• If a claimed cause shows a magnitude no larger than in periods without the outcome (2008 leverage ≈ 1998 leverage), then it cannot be the sufficient cause and must be downgraded.
• If a striking result comes from the smallest units in the data (best schools are small), then compute the standard deviation of the mean before inferring causality — small samples produce both extremes.
• If you already favor an action, then evaluate models to test it rather than search for one that endorses it; the tell of good practice is rejecting your preferred frame when a rival fits the facts better.
• If the data privileges no single account, then hold multiple mutually contradictory narratives as a lattice rather than forcing a premature winner.

• One-to-one reasoning: applying a single model to a complex phenomenon ('Trump voters were left behind economically,' 'inequality results from concentrations of capital'). Tell: the explanation is a book-title slogan with no stated conditions under which it fails.
• Motivated model shopping: selecting the frame that justifies a decision already made. Tell: no rival model was seriously evaluated and no disconfirming fact was sought.
• Inferring causality from outliers: concluding small schools perform better because the best are small, when the very best will be small even if size has no effect. Tell: the analysis never divided the standard deviation by the square root of sample size.
• Skipping the internal-consistency check: accepting an account because superficial evidence aligns (money did flow from China, banks were leveraged) without asking whether it makes logical sense given incentives. Tell: correlation is cited but no causal mechanism is defended.
• Ignoring magnitude: naming a cause without calibrating its size against baseline periods. Tell: the claim is directional ('leverage increased') but never quantified against a comparison year.

• ▸Phenomenon to explain: state the specific event or pattern in one line.
• ▸Candidate accounts: list every plausible explanation before analyzing any.
• ▸Causal force / model per account: name the mechanism and the formal model for each.
• ▸Internal-consistency check: does each account make logical sense given actor incentives? Note the gap.
• ▸Magnitude / calibration test: what data or number would confirm or refute the effect's size?
• ▸Disconfirming fact: the single observation each account cannot accommodate.
• ▸Statistical-artifact check: could sample size, base rates, or recency bias explain the pattern instead?
• ▸Surviving lattice: which models remain, under what conditions each holds, and how they interweave.
• ▸Decision test: am I evaluating models to find truth, or shopping to justify a chosen action?

1. 1Frame the decision as a state of the world: write down the full description of relevant information (team, structure, plan, situation) before you choose, so you know what each model is trying to classify.
2. 2Assemble an ensemble of distinct models rather than one—apply several logical frames whose insights overlap and interweave, because each accentuates different causal forces and different blind spots.
3. 3When you can, partially separate the major causal threads (as analysts of the 2008 crisis split foreign asset purchases, asset bundling, and leverage ratios), then explicitly trace how the isolated threads interweave.
4. 4Decide granularity by use: go big and aggregate when predicting, go small and specific when explaining why. Don't reach for one size.
5. 5Run each candidate action against a network view: check whether your choice depends on high-clustering, on long-leap connections, or on nodes that could fail—these determine whether the outcome holds when one link breaks.
6. 6Aggregate the models by majority classification (buy/sell, run/wait, guilty/innocent) so that no single frame's error decides for you—the Condorcet logic only pays off if the models are diverse and independently derived.
7. 7Test choices in an unrealistic environment: ask 'what if a key assumption were reversed?' to reveal the limits of the process before committing, the way abandoning constraints reveals boundaries.
8. 8Name the causal thread each model omits, since all models are wrong by simplifying; robustness comes from crisscrossing the omissions of many frames, not from perfecting one.

On August 14, 2003, sagging trees leaning on power lines near Toledo, Ohio caused a localized outage that spread when a software failure stopped an alarm from telling technicians to redistribute power; within a day over 50 million people across the northeastern U.S. and Canada lost power, and the same year a downed line between Italy and Switzerland darkened 60 million Europeans. Engineers modeled the grid as a network of nodes and edges and used that single model three ways: to explain how a small failure cascaded, to predict where future failures were likely, and to guide action by siting new lines, transformers, and supplies to raise robustness. The decisions held up because the model exposed the interdependencies—one unaddressed alarm feeding back into a continent-wide collapse—rather than treating the outage as a local event. Note the design tradeoff the passages flag: power companies avoid high clustering because it is inefficient, but that same lack of clustering reduces robustness, so a robust choice must weigh efficiency against tolerance to node failure.

• If the decision spans career, community, and personal life at once, then apply multiple models and take the majority classification—don't let one frame stand in for the whole state of the world.
• If you are predicting an outcome, then aggregate and go big; if you are explaining why something happened, then use smaller, more specific models.
• If your chosen action relies on a single node, hub, or long-leap connection, then treat it as fragile and plan for that link's failure before committing.
• If two models divide the system into 'independent' parts, then verify the parts are actually separable—if a feedback crosses the boundary (an alarm not firing), reconnect the threads and re-decide.
• If you assume the world is self-similar ('turtles all the way down'), then stop—use a different model at each level, because the same frame rarely holds across scales.

• Applying one model correctly and calling it done (the traditional one-to-one habit taught in school). Tell: you can name the model you used but not the one it omits.
• Drawing conclusions from one frame while ignoring another—Allison drew implications from the game-theoretic model without consulting the organizational model. Tell: your rationale never references an alternative lens.
• Assuming self-similarity across levels. Tell: you use your top-level model to reason about sub-levels without checking whether structure changes.
• Optimizing for efficiency and treating robustness as free—like a grid built with low clustering. Tell: your plan has no slack, redundancy, or fallback for a failed node.
• Building an ensemble of near-identical models and expecting the Condorcet benefit. Tell: your 'many models' share the same assumptions, so their errors correlate and majority vote gains nothing.

• ▸State of the world: write the full set of relevant facts the decision depends on.
• ▸Model 1 + its classification: which frame, and what action/prediction it yields.
• ▸Model 2 (different assumptions) + its classification: name the causal force it emphasizes that Model 1 ignores.
• ▸Model 3 + its classification: add a third independent frame to break ties.
• ▸Majority verdict: the action favored across the models.
• ▸Interdependencies & feedbacks: which links, if broken or triggered, would cascade (the 'unfired alarm').
• ▸Single points of failure: nodes/hubs/long-leaps the choice relies on, and the fallback for each.
• ▸Efficiency-vs-robustness note: where you traded slack for efficiency and what it costs under stress.
• ▸Reversed-assumption test: pick one key assumption, flip it, and record whether the decision still holds.

1. 1Assemble a set of predictive models (or people acting as models) that each produce a numerical estimate or categorical classification of the unknown value.
2. 2Compute the many-model prediction by averaging the individual predictions; for categorical judgments, take the majority classification instead.
3. 3Diagnose whether the set is genuinely diverse: check that models categorize differently, predict differently, or both — two models built on correlated attributes (education vs. income) will be accurate but not diverse.
4. 4Apply the diversity prediction theorem explicitly: Many-Model Error = Average-Model Error − Diversity of Model Predictions. Higher diversity subtracts more error from the average.
5. 5Verify no shared bias: if every model leans the same direction (all predict too high), averaging cannot cancel the error — the collective inherits the bias.
6. 6Add models in order of categorical difference, not accuracy. A less accurate model that partitions the world differently improves the average; a redundant accurate one does not.
7. 7Stop adding once marginal contribution flattens — in practice 'many' lands near five, not fifty; track each added model's incremental accuracy gain and quit when it collapses.
8. 8For categorical prediction, use majority vote across the models and raise confidence toward certainty only when independent models converge on the same classification (Levins: 'the intersection of independent lies').

Two models predict how many Oscars a film wins. One predicts two, the other predicts eight, so the many-model prediction is their average, five. The film actually wins four. Model one's squared error is 4 (2 squared), model two's is 16 (4 squared), giving an average-model error of 10. The diversity equals 9, because each prediction sits 3 away from the mean of five (3 squared each). The many-model error is therefore 10 − 9 = 1, which is (5 − 4) squared. The average of the two models is more accurate than either single model, because one ran high and one ran low and the opposite-signed errors partly cancel.

• If two candidate models categorize the data the same way (e.g., education and income sort cities identically), then treat them as one — adding the second buys accuracy without diversity.
• If a second or third model is slightly less accurate but categorically different, then add it; diversity, not individual accuracy, is what lowers many-model error.
• If all your models share a common bias, then averaging will not help — fix the shared assumption before adding more models.
• If independent models converge on the same categorical answer, then raise confidence sharply; if they diverge, treat the question as genuinely uncertain.
• If marginal accuracy from each added model has flattened (as around a handful), then stop — Google's interviewer data shows the jump from 1→4 evaluators does most of the work, and 20 barely reaches 90%.

• Chasing the single 'best' model: single-model thinkers miss large events like market collapses or the Arab Spring; the tell is confident prediction with no second framework to catch blind spots.
• Mistaking correlated models for diverse ones: you stack five models that all sort by essentially the same attribute; the tell is that they rarely disagree, so their average never improves.
• Manufacturing diversity without accuracy: categorizing cities by the first letter of their name creates 26 diverse buckets but a useless model; the tell is high variance across models with poor per-category prediction.
• Assuming averaging cures bias: if every model overpredicts, the average still overpredicts; the tell is a persistent one-directional error that more models never shrink.
• Believing the theorems promise near-zero error: the diversity theorem implies error can approach zero only IF you can build many diverse accurate models — in practice you can build maybe three to five, so over-adding wastes effort.

• ▸Target quantity/category — state the exact value or class being predicted and the true value once known.
• ▸Model/predictor list — name each model or person supplying a prediction.
• ▸Individual prediction — record each model's numerical estimate or classification.
• ▸Many-model prediction — average the estimates (or take majority vote for categories).
• ▸Categorization basis — note the attribute each model sorts on, to flag correlated (non-diverse) models.
• ▸Average-model error & diversity — compute both to confirm Many-Model Error = Average-Model Error − Diversity.
• ▸Shared-bias check — mark whether all predictions lean the same direction (yes = averaging won't help).
• ▸Marginal gain per added model — log accuracy improvement from each addition; stop when it flattens.

1. 1Place the phenomenon on the wisdom hierarchy: separate raw data (uncoded events) from information (named, partitioned categories) from knowledge (correlative, causal, logical relationships, often in model form). You cannot apply wisdom to a data hairball—first structure it.
2. 2Assemble a lattice, not a single lens: pull two or more models that accentuate different causal forces (e.g., a gravity model and a terminal velocity model; nations-as-self-interested-actors and organizations-as-networks). Aim for a few dozen models in your head over a career, not hundreds.
3. 3Choose your combination method by task. When one model is clearly correct for the conditions, select it (draw the right arrow from the quiver). When predicting, average models. When acting, run all models like a doctor's diagnostic panel to rule out and privilege actions.
4. 4Test separability: partially isolate the major causal threads (as in the 2008 crisis—foreign asset purchases, asset bundling, leverage ratios) and then examine how they interweave. You cannot carve the world at its joints, so map overlaps rather than assuming independence.
5. 5Construct a dialogue across the models: state where they agree, where they diverge, and why. Divergence is the signal—the cheetah problem is only solved because two models disagree and one fits the conditions.
6. 6Ground the resolved logic in evidence: clarify assumptions, then fit, calibrate, and test the causal and correlative claims against data before carrying the choice into the world.
7. 7Check for the single-model tells before committing: if your explanation reduces to one cause (grit, sugar, capital concentration), add models until blind spots—income disparity, identity diversity, system interdependencies—surface.

A small stuffed cheetah falls from an airplane's hold at 20,000 feet; how much damage on landing? A student who knows only one model applies it blindly. The gravity model predicts the toy tears through a car's roof. The terminal velocity model predicts its speed tops out around 10 mph—so a person could stand on the ground and catch it in her hands. Wisdom here is not knowing more models but knowing which to apply: the terminal velocity model fits a light, high-drag object, so it governs. The gravity model isn't wrong in general—Newton's laws hold only at certain scales—it is wrong for this condition. The dialogue between the two, and the selection of the correct one, is the wise choice.

• If the task is prediction across noisy conditions, then average multiple models rather than betting on one.
• If one model's assumptions clearly match the situation's conditions (light object, high drag), then select that single model and discard the others for this case.
• If you are choosing an action, then run all relevant models as diagnostics to rule out some actions and privilege others—do not act on the first model that fits.
• If two models diverge sharply on the same phenomenon, then treat the divergence as the decision point and interrogate which assumptions the real situation satisfies.
• If your account of a complex phenomenon rests on a single cause, then it is incomplete—add models until the blind spots (interdependencies, diversity, disparity) appear.

• One-to-one thinking: applying exactly one model per problem, as taught in ninth grade. Tell: your explanation names a single cause for a high-dimensional phenomenon (Trump voters were left behind economically; success depends on grit).
• Charisma of a clean form: trusting a single spare equation to predict complex systems. Tell: you expect exact numerical predictions of sea levels in 10,000 years or unemployment in 10 months from one model.
• Applying wisdom to unstructured data: reasoning directly off raw events with no categories or knowledge. Tell: you are staring at incomprehensible hairballs and calling it analysis.
• Ignoring separability and scale: using one level's model everywhere ('turtles all the way down'). Tell: you apply molecular-biology logic to neuronal-network robustness or to cognitive bias.
• Selecting the wrong model despite knowing the right one: having both gravity and terminal-velocity models but applying the ill-fitted one. Tell: your answer contradicts a sanity check (a stuffed toy tearing through a roof).

• ▸Phenomenon: state the specific complex situation you must understand or act on.
• ▸Hierarchy check: list the raw data, then the information categories, then the knowledge/models you hold.
• ▸Model ensemble: name 2-5 models and the distinct causal force each accentuates.
• ▸Combination method: mark whether you will SELECT, AVERAGE, or run a DIAGNOSTIC panel—and why.
• ▸Separability map: list the major causal threads you can partially isolate, and where they interweave.
• ▸Divergence log: note where models disagree and which assumptions the real situation actually satisfies.
• ▸Wise choice: the action or prediction you privilege, and the models you ruled out.
• ▸Evidence test: the data you will use to fit, calibrate, or test the chosen claim.
• ▸Blind-spot audit: features (interdependencies, diversity, disparity) no single model captured.

1. 1Adopt the one-to-many stance: rather than learning hundreds of narrow models, master a modest set of flexible ones and practice applying each across unrelated domains. Page's whole method is that most models travel.
2. 2Build the repertoire across the structural families in Page's index — distributions (normal, power-law), functional forms (linear, concave/convex), networks, dynamics (random walks, Markov, systems dynamics), and strategic models (game theory, cooperation, signaling, mechanism design). Coverage of families matters more than count.
3. 3For each model, record its core equation and its parameters so you can port it. Example: exponential growth is V_t = V_0(1+R)^t; the epidemiology diffusion model reduces to the basic reproduction number R_0 built from contact rate, spreading rate, and abandonment rate.
4. 4Deliberately stress-test each model outside its home discipline: take the epidemiology contagion model and apply it to seed corn, Facebook, crime, and pop stars; take signaling and apply it to advertising, marriage, peacock feathers, and insurance premiums.
5. 5When two disciplines give opposing proverbs (e.g. 'culture trumps strategy' vs. 'incentives drive behavior'), map each onto a formal model and derive the conditions under which each wins, turning slogans into conditional logic.
6. 6Treat failed applications as information: attempts at creative uses reveal a model's limits, so keep notes on where a model breaks and why — that boundary knowledge is part of the mastery.
7. 7Check that expanding your category/model set actually helps by remembering the bias-variance trade-off: more categories reduce categorization (model) bias but raise valuation (estimation) variance. More is not automatically better.

Page confronts the organizational-theory claim that 'culture trumps strategy.' Instead of picking a side, he reaches into a broad repertoire and applies two different models. First he uses a network contagion model: a CEO announces a new strategy, some initial proportion adopt it, and enthusiasm spreads through work-network contacts against a countervailing rate of abandonment. The three drivers — contact rate, spreading rate, abandonment rate — map directly onto the parameters of the basic reproduction number R_0. This yields conditional logic: culture trumps strategy if people don't believe in it, if they abandon it quickly, or if advocates are poorly connected; otherwise strategy trumps culture. He then brings in a second model, replicator dynamics, to view the same question through another lens. Breadth is what lets him convert an unfalsifiable proverb into testable conditions.

• If a model's core structure (its equation and parameters) matches a new phenomenon's mechanism, then port it — don't wait for a same-discipline model; the epidemiology model fits crime and pop-star spread because the diffusion mechanism is shared.
• If you face competing proverbs or expert intuitions, then instantiate each as a separate formal model and derive the boundary conditions rather than debating in words.
• If a mechanism depends on actor sophistication, then pick the model to match: repetition works for almost any behavior, but reputations and norms require forward-looking actors and information sharing.
• If you're deciding how many models/categories to layer on, then apply the bias-variance trade-off — add models until the reduction in categorization error is offset by rising estimation variance.
• If a model fails in a new context, then keep it and record the failure — the discovered limit is itself worth the attempt.

• Disciplinary siloing: keeping models locked to their origin field ('you wouldn't apply physics to the economy'). Tell — you can only cite models from your own training and never reuse a structure across domains.
• Collecting count over structure: memorizing many models that are all the same family (all distributions, or all game-theory variants). Tell — your 'many models' all answer the same kind of question.
• Applying a model whose behavioral assumptions don't fit: using reputation/norm mechanisms on actors who aren't forward-looking or don't share information. Tell — the mechanism 'should' work but predictions keep missing.
• Assuming more categories or models always improves accuracy. Tell — you keep subdividing and estimates get noisier, not sharper (rising variance of the within-category mean).
• Treating learning as optimization: assuming that because actors learn they reach the optimum, ignoring the Lucas-critique feedback where the model changes behavior it describes. Tell — your model works until people adapt to it, then breaks.

• ▸Model name & home discipline: where it originated (e.g. epidemiology, economics)
• ▸Core structure/equation: write the minimal formal form (e.g. V_t = V_0(1+R)^t)
• ▸Parameters: list each tunable input and what it means
• ▸Behavioral assumption: what actors must be (rule-based, forward-looking, adaptive)
• ▸Home application: the phenomenon it was built for
• ▸Ported applications: 3+ unrelated domains you've mapped it onto
• ▸Boundary/limit: context where the model failed and why
• ▸Bias-variance check: does adding this model reduce categorization error more than it adds estimation noise? (Y/N + note)

1. 1Master a modest set of flexible models first (random walk, signaling, diffusion/contagion, rugged landscape, geometric scaling, linear/success equation) rather than trying to learn fifty or a hundred. The one-to-many property means each covers many cases.
2. 2Take a model out of its disciplinary silo. Ask the generative question the passages model: 'How many uses can I think of for a random walk?' Do this for each mastered model to build a mental inventory of candidate applications.
3. 3Reapply a model to a new domain by reassigning names and identifiers. The epidemiology diffusion model becomes 'seed corn adopters,' 'Facebook users,' 'crime spread,' or 'pop star popularity' by swapping what a 'carrier' and 'infection' denote.
4. 4Modify the assumptions to fit the target. A model built for jurors making binary guilty/innocent calls becomes a model of classifiers where each 'juror' is now a model producing buy/sell actions or Democratic/Republican predictions.
5. 5Construct a novel analogy where the target has no obvious model. Samuelson reinterpreted physics equilibria for markets; Downs mapped ice-cream vendors on a beach onto electoral competition. Look for the shared abstract structure, not surface similarity.
6. 6Test the fit against the three functions a model must serve: explain, predict, or help you reason. Run the model on a known case and check whether it produces the right sign, magnitude, or insight.
7. 7Skeptically discard models that do not fit — set the model aside if it cannot explain, predict, or aid reasoning. Treat the failed attempt as informative because it reveals the model's limits.
8. 8Keep a scale check: not every model fits every task (you would not use a sewing machine to repair a leaky pipe). Match the model's mechanism to the domain's actual causal forces before committing.

Take the geometric formula for area and volume — a single model — and run one-to-many. Reassign its variables to supertankers to explain their size: volume (cargo) scales with the cube of length while the hull surface (steel cost) scales with the square, so bigger ships carry disproportionately more per unit cost. Then modify the target: apply the same square-cube relation to criticize the body mass index, which treats mass as scaling naively with height. Predict metabolic scaling by mapping surface-area heat loss against volume. Finally stretch it by analogy to explain why we see so few women CEOs. Each reuse required no new model — only renamed variables and adjusted assumptions — and where the fit strains, the strain itself exposes the model's boundaries.

• If a model can explain, predict, or help you reason in the new domain, keep it; if it does none of these, set it aside.
• If the target and source share the same abstract causal structure (not just surface features), reassign names and apply; if only surface resemblance exists, do not force it.
• If the domain's outcome depends on a variable outside actor control (like crude oil price for an oil company), model it as luck-heavy and expect regression to the mean — do not reward the actor for a good year.
• If you can find no existing model for a phenomenon, build a novel analogy from a mastered model before assuming the phenomenon is unmodelable.
• If an attempted application fails, extract the limit it revealed rather than discarding the exercise — failure still yields information about scope.

• Trying to master fifty or a hundred models instead of exploiting the one-to-many property; the tell is a reading list that never gets applied.
• Keeping a model out of loyalty after it fails to explain or predict; the tell is you defending the model rather than the model producing insight.
• Matching on surface similarity instead of shared structure — like using a sewing machine on a pipe; the tell is the analogy sounds clever but yields no correct prediction or action.
• Ignoring scale relationships when transferring (treating volume as if it scaled linearly like length); the tell is conclusions that break at the extremes, as BMI does for very tall or short people.
• Rewarding or reasoning about outcomes as skill when they are luck-driven; the tell is paying CEO or salesperson bonuses in luck-heavy contexts and being surprised by regression to the mean.

• ▸Source model: name the mastered model you are reusing (e.g., diffusion, signaling, geometric scaling).
• ▸Original variables/actors: what the model's names originally denoted.
• ▸Target domain: the new phenomenon you want to explain, predict, or reason about.
• ▸Renamed identifiers: map each original variable/actor onto its counterpart in the target.
• ▸Assumption tweaks: which original assumptions you must relax, drop, or add for the target.
• ▸Intended function: explain / predict / reason / design / communicate — pick and state why.
• ▸Fit test: one known case run through the model and whether it gave the right result.
• ▸Keep-or-discard: verdict plus, if discarded, the model limit the failure revealed.

1. 1Split your data before fitting: reserve part of an existing data set to fit the model, then test the fitted model against the held-out remainder. If prediction is the goal, fit on existing data and test against future data as it arrives.
2. 2Count your arbitrary parameters before celebrating fit. Ask Fermi's question of yourself: how many free parameters did I use? Treat each added parameter as a cost, not a win.
3. 3Compute R² — the percentage of variance explained — to measure how much of the observed variation your model accounts for, rather than relying on the plausibility of the narrative.
4. 4Watch for higher-order terms (squares, cubes, fourth powers) because they amplify errors: 10⁴ is 16 times as large as 5⁴, so small input differences produce large output errors.
5. 5Compare in-sample fit against out-of-sample error. If fit is high but out-of-sample error is high, you are overfitting; simplify the model.
6. 6Calibrate category granularity by trading off categorization error (too few categories clumps different means together) against overfitting (too many categories, one per unit, explains everything and predicts nothing).
7. 7Run more than one model against the same evidence. Fit an ensemble so their insights overlap and interweave, checking whether the magnitude and direction of effects hold across frames rather than trusting a single story.

Take Page's grocery-spending categorization model. You want last month's purchase data to predict this month's, but households fluctuate month to month. If you follow Borges's mapmakers and place each household in its own category, you fit the data perfectly — the model predicts each household spends exactly what it did last month — but that overfits and predicts poorly given monthly fluctuations. Instead you bin households into categories by attributes so members share similar means, then use each category's average as the predictor. Larger categories raise categorization error (you clump different means together); tiny categories overfit. The craft is choosing the granularity where one household's first-month spending genuinely informs another's second-month spending, which you verify by testing the fitted model against the held-out month rather than by admiring how well it matched the month you fit it on.

• If in-sample fit is excellent but out-of-sample error is large, then you are overfitting — cut parameters and higher-order terms.
• If a result depends on four or more arbitrary parameters, then treat the fit as suspect: with enough parameters you can fit an elephant.
• If adding a category or parameter no longer reduces out-of-sample error, then stop adding — you have crossed into overfitting.
• If one model explains the data, then fit a second and third; if the magnitude or direction of the effect flips across models, distrust the single narrative.
• If your model uses higher-order terms, then check error amplification explicitly, since higher powers magnify small differences into large ones.

• Mistaking good fit for a good model — the tell is delight at agreement between calculated and measured numbers without counting parameters (Fermi's challenge to Dyson).
• Overfitting by maximizing categories or parameters — the tell is near-perfect in-sample explanation paired with poor prediction of the next period.
• Testing the model only on the data it was fit to — the tell is you never held out a sample or waited for future data.
• Relying on a single-model narrative (grit explains success, sugar explains poor health) — the tell is a comprehensive-sounding claim that no competing model was run against.
• Ignoring the conditions under which results hold — the tell is applying an intuition universally, as if the Pythagorean relation held for non-right triangles.

• ▸Model / hypothesis: state the effect and its claimed magnitude
• ▸Fitting data: which subset was used to fit
• ▸Test data: held-out remainder or future data used to test
• ▸Number of arbitrary parameters: count them (Fermi test)
• ▸R² / variance explained: in-sample figure
• ▸Out-of-sample error: prediction error on test data
• ▸Higher-order terms present: list squares/cubes/powers and amplification risk
• ▸Category granularity: categorization error vs overfit tradeoff noted
• ▸Competing models run: list ensemble members and whether effect held

1. 1Classify the use before choosing the model. Ask which of the four uses applies — explain, predict, design, or act. Explanation tolerates low-fidelity abstraction; prediction, design, and action can absorb high-fidelity, granular models when data supports it.
2. 2Match granularity to data volume. As a rule of thumb, the more data you have, the more granular the model. With BIG data, add categories or attributes; with thin data, coarsen the partition so category averages stay stable across future draws.
3. 3Pick the behavioral assumption to fit the actors. Where individuals conform or align with neighbors (discrete choice, physical alignment), a rule-based model like local majority fits. Where actors optimize, a rational/coordination-game model fits — and note these can converge on the same behavior, so choose the simpler adequate one.
4. 4Size the dimensionality to the marginal variation each dimension explains, not to realism. Add a dimension only when it materially raises variance explained; stop when the added dimension buys almost nothing.
5. 5For prediction under stakes, assemble a diverse set of moderately accurate models rather than one. Diversity — not shared bias — drives the accuracy gain; a diverse collection beats its average member.
6. 6Guard the bias-variance trade-off. Decompose model error into categorization (bias) and valuation (variance) error, and stop adding categories at the point where overfitting starts degrading out-of-sample prediction.
7. 7Apply one-to-many deliberately: take a modest number of flexible models and test them against the new context, treating failures as information about where the model's limits lie.

Take the grocery-chain data: millions of households, total variation of 1 billion (sum of squared deviations from the $500 mean). A categorization model partitions households into groups and predicts each group's average. If the model estimates a household at $600 and it spent $600, the model explains all 10,000 that household contributed; if it says $700 and the household spent $800, the 90,000 contribution ((800−500)^2) shrinks to a 10,000 contribution ((800−700)^2). Adding categories raises R^2 — but push it to Borges's limit, one household per category, and you explain 100% of last month's variation while predicting each household spends exactly what it did last month, a poor predictor given monthly fluctuation. So for explanation you can go granular, but for prediction you place each household in a category of similar households and use their average, treating each month's spend as a draw from a distribution with a mean and variance. The right granularity is the one that minimizes error out of sample, not in sample.

• If the goal is to explain and reduce detail to essentials, then favor a low-fidelity abstraction — a life-sized map is useless.
• If you have BIG data and are predicting or designing, then increase granularity; if data is thin, then coarsen categories so averages are reliable.
• If adding a dimension or category raises out-of-sample accuracy materially, then keep it; if it adds near-zero variance explained (like the second congressional dimension's 4%), then drop it.
• If actors conform or align locally, then use a rule-based model; if they optimize, then use a rational model — and if both give the same behavior, choose the simpler.
• If stakes are high and outcomes hard to anticipate (market collapses, Arab Spring), then use many diverse models, since single-model thinkers miss large events.

• Chasing fidelity for its own sake. The tell: R^2 climbs toward 1.0 in-sample while predictions of next period get worse — you've built Borges's map.
• Averaging models that share a common bias and calling it wisdom of crowds. The tell: your ensemble is confident and consistently wrong in the same direction.
• Treating the theorems as promises. The tell: you assume 'enough models' drives error to near zero, ignoring the real limit on constructing genuinely diverse, accurate models.
• Forcing a model onto a context it can't hold. The tell: you use a sewing machine to fix a leaky pipe — applying a model whose assumptions the situation violates and ignoring the failure as signal.
• Adding dimensions to look realistic rather than to explain variation. The tell: the extra dimension changes the picture but barely moves accuracy, and may alter theoretical conclusions in ways you can't defend (radial-symmetry-type strong assumptions).

• ▸Use case: mark explain / predict / design / act
• ▸Data volume: estimate rows and attributes available; note BIG or thin
• ▸Target granularity: number of categories/dimensions and why that count
• ▸Behavioral assumption: rational / rule-based / adaptive — and the actor evidence for it
• ▸Single vs. many: one adequate model, or a diverse ensemble; justify by stakes
• ▸Bias-variance check: expected categorization error vs. valuation error at chosen granularity
• ▸Marginal-value test: variance each added dimension/category explains; stop threshold
• ▸Failure limits: contexts where this model's assumptions break and what that would reveal

1. 1State the problem, then list every distinct causal force that could plausibly drive it—income disparity, identity diversity, interdependencies with other systems—rather than settling on the first frame that comes to mind.
2. 2Assign at least one model to each force whose logic accentuates that force: a contagion model for spread, a network model for structure, a Markov model for state transitions, a distributional model for spread of outcomes.
3. 3Deliberately choose models that accentuate DIFFERENT causal forces; if two candidate models tell the same story, drop one and add one that illuminates a blind spot the others share.
4. 4Apply each model separately—clarify its assumptions, run its logic to conclusions—so you can see where predictions and implications overlap and where they diverge.
5. 5Compare outputs: note where the ensemble agrees (robust conclusion) and where models disagree (the disagreement flags a real uncertainty or a missing force).
6. 6Use the one-to-many property to economize: reassign names, identifiers, and assumptions to reapply a model you already know to this domain instead of learning a wholly new one.
7. 7Ground the surviving models in data—fit, calibrate, and test their causal or correlative claims—and discard or refine any frame that cannot explain, predict, or match evidence.
8. 8Decide granularity by use: go big (fewer, larger models) when predicting; go small (many specific models) when explaining.

Consider the 2003 power grid failures. Page models the grid as a network to explain how cascading failures propagate through connected nodes—one frame accentuating structural interdependence. That single network model helps explain how failures occurred and predicts regions where future failures are likely. But it also guides action: it identifies specific locations where new lines, transformers, and power supplies would raise the network's robustness. The point is one model put to many uses (explain, predict, design), and it demonstrates diversity of engagement because the network frame captures a causal force—connectivity and cascade—that an isolated component-reliability model would miss. Add a distributional model of failure sizes and you catch the long-tailed risk the network view alone understates. The ensemble, not either frame alone, produces the robust guidance.

• If two models you are considering emphasize the same causal force, then replace one with a model that targets a force none of your current models address.
• If the use is prediction, then favor one big model or fewer large ones; if the use is explanation, then favor many small, specific models.
• If a model cannot explain, predict, or match the evidence when taken to data, then discard or refine it rather than keep it for symmetry.
• If your models all agree, then treat the conclusion as robust; if they disagree, then treat the disagreement as a signal of genuine uncertainty or an omitted force, not noise to average away.
• If a needed model exists in another domain, then reassign its names and assumptions to reuse it rather than assuming you must learn a new one.

• Applying the traditional one-to-one logic—identify the one proper model and apply it correctly—so you never build an ensemble at all; the tell is that you feel finished after a single frame explains the problem.
• Stacking multiple models that all accentuate the same causal force, producing false confidence rather than genuine diversity; the tell is that every model gives you the same answer and no blind spots surface.
• Ignoring features single models omit—income disparity, identity diversity, system interdependencies—because your chosen frames share the same omissions; the tell is a policy that works on paper but ignores who is affected and how systems interact.
• Treating models as a spectator sport—reading formalism without applying it—so you never actually engage the logic; the tell is you can name models but cannot run one on your own problem.
• Choosing granularity by habit rather than use, e.g., over-detailing when you only need prediction; the tell is 'nothing is less real than realism'—details confuse and you lose the real meaning.

• ▸Problem statement: one sentence naming what you must explain, predict, or design.
• ▸Candidate causal forces: list every distinct force that could drive the outcome.
• ▸Model per force: name the model whose logic accentuates each force.
• ▸Distinctness check: mark any two models that tell the same story—drop one.
• ▸Assumptions and output per model: what each frame concludes when run to logic.
• ▸Overlap and divergence: where the models agree (robust) and disagree (uncertainty).
• ▸Data test: which claims you will fit, calibrate, or test against evidence.
• ▸Granularity setting: big-for-prediction or small-for-explanation, with reason.
• ▸Reuse note: existing models reassigned to this domain by renaming/reassumption.