How Mechanism Analyses Are Made

I. Why a method essay exists

The companion essay, The Mechanism Analysis, defines an artifact: a document that tests whether a proposed law's causal machinery can produce its stated outcome before the law is passed. The artifact has five components — a mechanism claim, an actor and response map, a capital-stock and absorption ledger, a failure readout, and a repair specification. The artifact's template is simple enough to put on a page.

The artifact's template is not the artifact's quality. A mechanism analysis can be filled out from a checklist and produce something that looks structural while reading like punditry. The five fields can be populated with the analyst's prior policy preferences wearing the artifact's vocabulary. The lawmaker reading it cannot tell from the artifact alone whether the mechanism claim came from disciplined enumeration of the bill's incentive structure or from the analyst's sense of what the bill "really" intends. This essay exists because the artifact's authority depends on the production discipline behind it, and the production discipline is not visible from the artifact's surface.

Four drifts undermine mechanism analyses produced without discipline. Narrative substitution — the analyst tells a story about the bill rather than tracing its causal claims through actors' incentives. Partisan diagnosis — the analysis is structurally coherent but its diagnosis happens to align with the political faction the analyst was always going to favor. First-order-only analysis — the targets of the law are modelled; the targets of the targets, the system-level equilibrium, and the institutional exits are not. Overconfident behavioral prediction — the mechanism analysis claims magnitudes ("20% of clinics will close") with the same authority it claims directions ("clinics will face reduced incentive to remain open").

The method exists to reduce these drifts. It is not a checklist. Checklists prevent omission of items in a list; they do not prevent the four drifts above, because the four drifts can occur while every checklist item is dutifully addressed. The method reduces drift by routing the analysis through perspectives the analyst would not adopt unprompted, decomposing labels the analyst would otherwise treat as atomic, and demanding confidence stratification the analyst would otherwise blur. The method is a set of moves applied with discipline, not a procedure executed mechanically.

This essay describes those moves. It documents the prerequisites the analyst brings to the work, the upstream pipeline that builds the evidence base, the analytical lenses that surface mechanism failures, the typology of findings, the iterative dialectic that turns first drafts into structural diagnoses, the reflexive moves that catch the framework's own failures, the failure modes of the method itself, the epistemic discipline that distinguishes mechanism logic from speculation, and the apprenticeship pattern that builds judgment. It is a zero-to-one document for a category of artifact that does not yet exist as a recognized form. The method described here is what works at this point in the corpus; it will be revised as the form develops and as practitioners other than the original analyst produce mechanism analyses under it.

II. Prerequisites: the taste you bring

A method is necessary but not sufficient. The transferable parts of mechanism analysis — the lenses, the heuristics, the typology, the confidence stratification — can be documented in prose. The non-transferable part is judgment. Without judgment, the documented moves produce output that has the form of mechanism analysis without its substance. With judgment, the documented moves accelerate work the analyst could in principle perform without them. The method is real, and the judgment is real, and they are not the same thing.

Five capacities the analyst brings to the work.

Game-theoretic intuition. What does the rational actor do under this payoff surface? What is the cheapest available response? What strategic substitutes does the mechanism create? Where are the gaming opportunities? Which exit routes are open? How do organized winners coordinate to defend the mechanism while diffuse losers fail to organize against it? The analyst without this capacity accepts the law's intended behavior as the likely behavior. Reading: Schelling, Olson, Ostrom, Tullock.

Causal-chain reasoning. What happens next, and then what happens because of that? Direct effects are visible to anyone; cascades, feedback loops, equilibrium shifts, and recovery horizons are not. The analyst without this capacity stops at first-order impacts and misses the system response that produces the mechanism failure. The capacity is built by watching causal chains play out in many domains — economics, biology, software systems, military operations — until the analyst's default response to any proposed intervention is to ask what the system does after it.

Institutional literacy. What does this institution actually do, not what does its mandate say? Which actors hold which budgets? Which procedural routes lead where? What capacity constraints bind which actors? Which legal choke points permit which moves? The analyst without this capacity produces repair specifications that are fantasy governance — mechanisms that require capabilities the system lacks, owners the system has not assigned, or coordination the system cannot perform.

Mechanism-design sensibility. Is this incentive-compatible for the actors whose behavior the law needs? The Hurwicz/Maskin/Myerson tradition asks not "what do we want actors to do" but "what payoff surface makes the actors we have want to do it." Bills that fail this test propose rules that require actors to act against their own constraints. Reading: Hurwicz, Maskin, Myerson, Roth, Shapley.

Source discipline. Which claims in this analysis are sourced, which are inferred, and which are speculative? Every claim should be traceable to either the bill's own text, a documented institutional fact, a recognized empirical regularity, or a labeled inference from prior claims. The analyst without source discipline produces artifacts whose credibility can be destroyed by one wrong number.

The reading canon for these capacities is short and dense. Schelling on strategic behavior. Olson on collective action and concentration. Ostrom on commons and institutional design. Kornai on soft budget constraints and shortage economies. Coase on transaction costs. Williamson on institutional comparison. Hurwicz, Maskin, Myerson on mechanism design. Roth and Shapley on matching and market design. Goodhart and Strathern on what happens when measures become targets. Hayek on knowledge problems. Tullock on rent-seeking. Adjacent: Tetlock on forecasting calibration, Taleb on hidden variance and fragility, the behavioral economics literature on capacity-constrained agents, the AI safety literature on specification gaming.

None of this substitutes for production experience. Reading the canon produces a vocabulary; producing analyses produces judgment. The method essay cannot manufacture taste in its readers — it can name the capacities that taste develops within, point at the readings that accelerate the development, and document the practice patterns that produce calibration. A reader without the underlying capacities can still follow the essay and produce something with the structure of a mechanism analysis; whether the structure ends up filled with disciplined causal reasoning or with the analyst's prior preferences wearing the structural vocabulary depends on practice, external dialectic, and calibration against outcomes — none of which the method itself can supply.

III. Before drafting: the upstream pipeline

What the analyst does before writing the first sentence of analysis. Skipping any of this produces drafts that fail under dialectical attack, because the gaps appear as soon as the draft is challenged.

Telos check. What does the analysis serve? What capital stocks does the proposal obviously affect? Which stocks have the longest recovery horizons? Which is closest to a fatal threshold? The analyst who skips this step ends up auditing whichever mechanism is most visible from the proposal's own framing, rather than the mechanism whose failure would matter most. A bill nominally about funding allocation may bind hardest on institutional capacity. A bill nominally about workforce activation may bind hardest on demographic structure. The telos check is the move that asks if this bill failed in the worst plausible way, what stock would have been depleted, and would the depletion be reversible?

Source gathering. The evidence bundle — bill text, explanatory memorandum, fiscal tables, the proposal's own impact assessment, oversight-body statements, stakeholder responses, prior evaluations of comparable laws, academic literature on the mechanism class, international analogues. Every claim the analysis will later make has to be sourced from this bundle or labeled as inference. The bundle is the artifact's accountability layer. A mechanism analysis whose factual claims cannot be traced back to specific documents, and whose inferences are not labeled as inferences, is rhetoric.

Mechanism inventory. Before identifying what is broken, map what the proposal does. Every mechanism the bill creates, modifies, or destroys. For each: a name, the agents acting and acted upon, the intended causal chain (if X then Y then Z), the capital-stock effects, and a status — working, error candidate, or unknown. The inventory is the denominator. Four errors out of forty mechanisms is a different finding than four errors out of six. The inventory also reveals which mechanisms should be preserved — a bill is not entirely broken just because some of its mechanisms are, and the repair specification must distinguish what to change from what to leave alone.

Capital-stock scan. A full pass across the seven stocks Full Accounting names as the working baseline — financial, physical, human, social, demographic, institutional, cognitive — extended at the granularity legislation requires. Capital stocks are an analytical abstraction over what a civilization runs on, not a closed taxonomy; the baseline list is a starting point that splits and recombines under specific legislation. Common operational extensions: fiscal and administrative as components of financial and institutional capital, regional and infrastructural as components of physical and institutional, trust as the load-bearing component of social, future optionality as a derivative property of the ledger as a whole. Other extensions surface when specific bills bind on stocks the baseline does not name granularly. For each obviously affected stock: direction of effect, recovery horizon, hysteresis class. The hysteresis class matters most. A stock that is fast to destroy and slow to rebuild (institutional knowledge, demographic structure, deep social trust) is categorically different from one that recovers on policy timescales (incentive responses, parameter calibration). The capital-stock scan is the discipline that prevents the analysis from booking a flow improvement at zero stock cost when the bill is in fact liquidating a non-fiscal stock to produce the flow improvement.

Baseline outcome context. For interventions targeted at populations — integration support, workforce activation, education, healthcare access — establish what the baseline produces. The tenth, fiftieth, and ninetieth percentile outcomes under current conditions. Where the bill makes fiscal, employment, activation, or long-term capacity claims, estimate the long-run fiscal or capacity trajectory of the target cohort under pessimistic, central, and optimistic scenarios. What the bill's own impact assessment projects, and whether that projection is realistic given the baseline. This is data, not advocacy. The mechanism analysis that omits it — auditing the intervention without naming what successful intervention actually produces — is incomplete. Absence of data is itself a finding about the state's analytical capacity.

The three questions for context discipline. Throughout the upstream work, the analyst forces three substitutions. Name the specific entities: the three regions where this mechanism binds hardest, not the abstract category "regions." Name the actual role-holders whose responses determine the outcome: the municipal officials with the relevant decision rights, not the abstract category "municipalities." Name the specific failure mode in this context: "in this region, there are zero alternative providers for this service," not "competition might not increase." The questions fight the drift toward abstraction that makes the analysis decoratively coherent and operationally empty.

IV. The lens kit

The analytical perspectives that consistently surface mechanism failures. The lenses are not a checklist. The analyst scans them and applies the ones that bite on the proposal; the rest are skipped. The most productive lens on any given bill is often one that did not appear central from the proposal's own framing.

The lenses cluster into four groups: how pressure routes through the system, how the system measures itself, how costs distribute across time and capital, and how the system's structure constrains its responses.

A. Response-routing lenses

Cheapest-channel routing. Pressure does not land where the designer intends. It routes through the cheapest available channel. The Response Vector decomposes any binding intervention's response into four channels: target response (the substantive change the intervention claimed to produce), base loss (collateral capacity loss around the intervention), formal gaming (the measured surface changes while the underlying activity does not), and incidence shifting (the activity continues but the burden moves elsewhere — to other actors, to staff, to consumers, to later periods). A real intervention typically produces some response in each channel; the question is which channel became cheapest after the intervention was applied. A funding squeeze on an entity with lobbying access produces incidence shifting onto a bailout-bearing principal. A funding squeeze on an entity with service-discretion produces base loss as service cuts. A funding squeeze on an entity with neither produces some mixture of base loss and formal gaming. The mechanism analysis asks: given this bill's pressure, which of the four channels is cheapest for each actor, and where does the cost ultimately land? The lens generalizes a single insight that recurs across most mechanism failures: a policy gets the response that its architecture makes cheapest, and the response's distribution across the four channels is allocated by relative cost, not by designer intent.

Institutional exit (secondary agency). Actors do not just respond within the law's domain. They escape to neighboring institutional systems whose rules the law did not anticipate. A welfare sanction routes its cost to the medical system through disability diagnosis. A tax rule routes its cost to the corporate tax base through transfer pricing. A regulatory burden routes its cost to the judicial system through appeals. A funding cut routes its cost to a future political process through implicit bailout pressure. The mechanism analysis asks: if this mechanism makes the actor's current position untenable, what is the cheapest adjacent institutional exit, and where does the cost land when they take it? The exit is usually predictable; the analysis should trace at least one hop.

Apathy response. The first-pass actor map often begins with rational optimizers who actively respond to incentives. Severely constrained agents may not optimize at all. They collapse into passivity. A welfare recipient facing complex sanctions, conditions, and forms may not run the rational calculation that says "work is irrational"; they may simply withdraw, with the cost surfacing as psychiatric and somatic load rather than as gray-economy activity. The mechanism error identification is the same; the cost channel is different. The analyst asks: does this mechanism assume the actor has the cognitive and emotional capacity to optimize, and what happens if the actor is incapacitated rather than calculating?

API mismatch. Some bills assume their target population has the capacity to respond to incentives. If the population cannot respond — lacks language, literacy, credentials, cultural knowledge, cognitive capacity — the mechanism produces a capacity gap, not a behavioral response. The mechanism works in a model where the actor can comply; it fails in reality because the actor cannot. The deeper version of this lens questions whether the bill is treating symptoms (improve the translation layer) when the root cause is the system's structural requirements being incompatible with the population's starting position. The lens does not absolve the analyst from naming when the capability gap is itself fixable by an unaddressed design choice; sometimes the real mechanism error is the absence of an intervention that would close the gap.

B. Measurement and classification lenses

Boundary and classification gaming. Any mechanism that creates a category, threshold, or definition creates incentives to restructure across the boundary. A 10% ownership threshold for a regulatory category produces entities that hold 9.9% or 10.1%, whichever is favorable. A revenue bracket for tax progression produces revenue managed to stay in the lower bracket. The analyst asks: can the actor reclassify themselves? At what cost? Is the cost lower than the benefit of being on the other side?

Proxy-target divergence. When a measure becomes a target, the activity in the measured domain changes nature. The seminal observation is Goodhart's, sharpened by Strathern: when a measure becomes a target, it ceases to be a good measure. The mechanism analysis asks: what metric does this proposal use for allocation or regulation? Can actors influence the metric without changing the underlying reality? What happens to the nature of activity in the measured domain when the metric becomes the target? In healthcare funding tied to coded morbidity, the activity of clinical care drifts toward documentation optimization. In educational outcome metrics tied to standardized tests, teaching drifts toward the test. The mechanism error is structural when the metric-target substitution is built into the funding architecture.

Information asymmetry exploitation. Who knows what, and what does the mechanism require them to report? The regulator assumes certain information flows. The regulated party has information the regulator does not and uses it strategically. The mechanism analysis asks: what information does this mechanism require to function? Who produces that information? What is their incentive regarding its accuracy? Does the mechanism create new information asymmetries — for example, moving from in-house production to procurement, which creates a buyer-seller asymmetry that did not exist before? Can actors strategically withhold, delay, or distort information, and at what cost?

C. Time and capital lenses

Temporal displacement and hysteresis. Costs pushed forward in time, and the asymmetry between destruction speed and rebuild speed. The proposal's own impact assessment typically covers an electoral cycle. The mechanism analysis asks what happens in years 5 through 10, 10 through 20, and at deeper time horizons. Many institutional losses are not literally irreversible on century timescales — trust can rebuild, competence can re-concentrate, demographic pipelines can sometimes be reopened — but a recovery time exceeding 50 years is operationally irreversible for governance purposes. Stocks divide into the fast-destroy-fast-rebuild class (incentive responses, parameter calibration) and the fast-destroy-slow-rebuild class (institutional knowledge, demographic structure, deep social trust, training pipelines). The latter class is where hysteresis bites. Three sub-patterns recur: extinction debt, where the law severs the recruitment or training pipeline while leaving the current stock intact, so the system looks healthy but is committed to collapse when the current cohort retires; niche construction, where beneficiaries of the new law terraform the legal and economic environment to lock in the change, so reverting the law no longer reverts the ecosystem; structural-versus-parametric asymmetry, where laws that change incentives or parameters for existing entities recover fast and laws that force creation or destruction of structural entities or knowledge repositories recover slow.

Balance-sheet theft. The state does not hold a balance sheet for non-financial capital stocks. The proposal's impact assessment books flow improvements (fiscal savings, throughput gains) and ignores the stock liquidations that produce them. A budget cut that produces a fiscal saving at the cost of staff burnout, regional service erosion, or loss of institutional capacity has not produced a saving in any honest accounting — it has converted a non-fiscal stock into a fiscal flow. The mechanism analysis asks: where is the saving coming from? Is it a genuine efficiency gain, or is it drawing down institutional knowledge, human capital, regional infrastructure, or rule credibility? The pattern recurs often enough that the analyst should default to checking whether the apparent saving is a stock liquidation. The phrase that compresses the move: selling the factory to pay the quarterly dividend.

Deep time net present value. When a mechanism severs a developmental pipeline — education to skills to employment to tax base — the relevant cost is the lifetime net present value of the person or cohort who never enters productive capacity, not the immediate fiscal saving the bill claims. A young person who falls out of an education, rehabilitation, or integration pipeline and never reaches stable productive capacity has a negative net present value to the fiscal system across decades — social transfers, healthcare, criminal justice, lost tax revenue. The lens is distinct from temporal displacement: it asks for the total cost of a permanently severed pipeline rather than the timing of when that cost arrives. The answer is often orders of magnitude larger than the mechanism's claimed saving. Deep-time net present value calculations are inherently speculative; the analyst states direction with confidence and magnitude with explicit uncertainty. Lifecycle net effect is multiples of the electoral-cycle saving is defensible; the cost is five hundred million over twenty years is not, without extensive modeling.

D. System-structure lenses

Thin-market assumption. Laws that assume competitive market dynamics in geographies or sectors where no functioning market exists. The mechanism's logic depends on competition driving efficiency. The empirical market structure is monopoly or oligopoly. The mechanism works in the model; it fails in the territory. The analyst asks: does this mechanism assume a market? How many actual providers exist in the relevant geography? What happens to the mechanism's logic if the number is zero or one? In rural service provision, in specialized procurement, in housing supply-constrained markets, in pharmaceutical retail, the assumption recurs and is repeatedly wrong. The mechanism subsidizes monopolist pricing where the assumption fails.

Concentration and diffusion asymmetry. Olson's observation: benefits concentrated among few visible actors who organize and lobby; costs diffused across many invisible actors who bear them without organizing. The political economy of who defends and who opposes is not symmetric. A mechanism can be net-negative in total and politically stable because the beneficiaries are organized and the cost-bearers are not. The lens does not change the mechanism analysis but predicts implementation dynamics, political feasibility, and the likely deformation of the mechanism over time as the organized winners exert pressure to preserve and extend their advantage. A mechanism whose costs fall on diffuse losers is unlikely to be repaired by the same political process that produced it unless another organized actor makes the cost legible.

Correlated risk. Laws are not independent. When multiple laws independently deplete the same stock, the individual impacts may be moderate and the correlated drawdown catastrophic — individually survivable, collectively fatal. Two sub-cases recur. The uncapitalized reinsurer of last resort: when multiple ministries independently shift structural risk onto a single entity, that entity becomes a balance sheet absorbing correlated risks the original arrangement assumed were independent. Welfare regions absorbing healthcare cost overruns, pharmacy retrenchment, procurement de-skilling, and poverty-driven service demand at once is the canonical case. Correlated catastrophic failure stops being a tail risk and becomes an expected failure mode once the shared backstop's capacity is exceeded. The geographic copula: when multiple laws simultaneously hit the same geographic segment, the compound effect exceeds the sum, and the region breaches the critical mass threshold for viability. The analyst asks: what is the existing load on the stocks this law depletes? Is there systemic margin left, or is this the straw that breaks the camel's back? The lens requires checking other recent mechanism analyses, not just the current law in isolation.

Counterfactual design. Delta-analysis (the bill changes X from three to two; is two worse than three?) anchors on the status quo. If the status quo is itself wrong, the analysis detects the delta's harm but misses the frame. Counterfactual design corrects this by depriving the analyst of the status quo reference point. Three modes. Counterfactual domain question: should the state be in this business at all? The bill's entire policy domain may be an intervention in the wrong place. Counterfactual mechanism design: given state action in this domain, what would a mechanism designer build from zero, with no knowledge of existing law and no anchoring on current parameters? Counterfactual parameter derivation: given this specific policy structure, what parameters does the mechanism's own logic require, blinded to both prior and proposed parameter values? The counterfactual is hypothesis, not prescription: clean-sheet designs ignore path dependencies, smuggle non-interventionist priors, and treat load-bearing constraints as bugs. The counterfactual output must be tested against the dialectical pressure that asks "what if the constraints the designer proposed removing exist for reasons the blank-slate designer cannot see?" Existing constraints may be features.

A note on lens prioritization

The lens kit is drawn from a corpus of welfare-state legislation. The lenses themselves — Kornai's soft budget constraint, Olson's concentration-diffusion asymmetry, Goodhart's measure-target divergence, Schelling's strategic substitutes — are general. The prioritization is local. In an adversarial-legalism state where judicial veto and rulemaking capture dominate, the lens kit should foreground capture analysis and procedural-route mapping. In a state where the dominant pathology is elite extraction rather than complex execution-traps, rent-seeking and patronage-network lenses dominate. The general lenses transfer; the priority weights depend on which institutional pathologies bind hardest in the local environment. Practitioners should treat the kit as a starting point and develop their own priority weights from the corpus of analyses they produce in their own context.

V. Three kinds of mechanism findings

Not every analytical finding is a mechanism error in the same sense. Three types stabilize the distinction.

Type 1: the mechanism fails by the proposal's own standard. The law does not do what it says. The mechanism contradicts the bill's stated objective. This is the politically safest finding to land — your bill does not do what you say it does — because the political faction defending the bill has committed to its stated goal as the bill's purpose, and the analysis demonstrates that the mechanism does not produce that purpose. The acid test for a Type 1 finding: would a reviewer who accepts the bill's policy goal still identify this behavioral distortion? If yes, Type 1. An example: a bill stated to incentivize employment that produces a marginal effective tax rate of 100%. The mechanism contradicts the stated objective at the level of arithmetic.

Type 2: the mechanism works but burns hidden capital. The law achieves its stated objective while depleting capital stocks the impact assessment did not count. The bill works as designed; the design is corrosive at the systemic level. The acid test: does this mechanism deplete institutional, human, cognitive, demographic, social, fiscal, regional, or moral capital in ways that undermine the system's long-term capacity, even if it achieves the bill's own stated goal? Many findings are both Type 1 and Type 2 — convenient, because the Type 1 framing is politically harder to dismiss while the Type 2 identifies the deeper harm. When a finding is only Type 2 — the bill achieves its stated objective but corrodes the system's vitality — the artifact flags it explicitly: not a mechanism error by the proposal's own criteria, but a mechanism risk to long-term capacity.

Type 3: the stated objective itself is a mechanism error. The goal presupposes a causal model that is wrong. Achieving the objective perfectly would still produce net harm. This is the deepest finding and the hardest to land politically, because it does not accept the bill's framing as the analytical floor. The acid test: does the bill's stated objective treat a capital stock as a resource to extract from, when the stock is in fact infrastructure to maintain? Does the objective assume a causal relationship that does not hold? Would achieving the objective perfectly still corrode what the objective itself presupposes? An example: a bill that frames a maintained service infrastructure as a tax base. Even a perfectly designed extraction mechanism corrodes the infrastructure it extracts from. The objective is the error.

The distinction compresses to a single sentence. Type 1 says the machine does not do what its builders claimed. Type 2 says it does, but by burning what the builders did not count. Type 3 says the builders asked the wrong machine to exist.

Political commentary is none of the three. A statement that the bill is "wrong" without identifying a specific behavioral distortion, capital depletion, or causal model failure is opinion, not analysis. The discipline is to upgrade opinions to one of the three types or to drop them. A finding that cannot be classified into Type 1, 2, or 3 is not yet a mechanism finding.

The Type 1/2/3 axis specializes the broader taxonomy of stuck-equilibrium primitives that Bad Equilibria Are Not One Thing develops as five diagnostic gates — object, destination, payoff, joint-move, execution. The two axes cross rather than nest. Most Type 1 findings localize to payoff or execution gate failures: the bill's incentive structure or implementation capacity does not produce its stated outcome. Most Type 3 findings localize to object or destination gate failures: the proposal misrepresents the problem as a stable object or routes responsibility to a destination that cannot receive it. Type 2 findings cut across the gates, because hidden-capital depletion is the failure mode that surfaces when the gates appear individually closed but the absorption audit reveals what is being silently consumed. The Type axis classifies the bill's relationship to its own stated objective; the five gates classify the structural primitive missing from the equilibrium. Both classifications matter, and a complete failure readout names both.

VI. The first draft, and why it is wrong

The first draft is raw material for the dialectic, not a product. Treating it as a product is what makes most analyses fail.

Steelman first. The strongest argument for the bill is in the bill's own admission of tradeoffs. The proposal's drafters knew their preferred outcome required accepting costs elsewhere. Where the proposal acknowledges those costs, the strongest pro-bill argument is implicit: the costs are worth it because the benefits are larger. The analyst who attacks the weak version of the bill's argument — the rhetoric used to sell it rather than the substantive case for it — wastes the first dialectical round resurrecting the steelman that should have been built into the draft. The draft attacks the strongest version of the proposal's causal model, not the weakest.

Expand and contract, separately. Two phases, deliberately not blended. Expand: cast wide across the lenses and the orders of consequence; surface every candidate mechanism failure without filtering; push past the natural stopping point and explicitly ask what was missed — what cross-sector interaction, what temporal displacement, what second- or third-order effect was not checked. Target two to three times more candidate findings than will survive. Contract: prune ruthlessly. Marginal findings drop. Findings that are sub-components of larger findings merge. Findings kept to fill an implicit quota are dropped. The two phases must remain separate because the natural mode of analytical work is to converge to a comfortable middle; the discipline that pushes past the convergence point only operates when discovery and filtering do not contaminate each other.

Production-rhythm convergence. Legislation varies enormously in complexity, quality, and mechanism density; finding counts should vary correspondingly, including zero. A bill whose mechanisms work is a finding, not a failure of the method. The discipline failure to watch for is the analyst's natural tendency to converge on a comfortable output volume per analysis, independent of what the underlying material warrants — the production rhythm settles into a middle and outputs cluster there regardless of input variance. The expand and contract phases, applied separately, are the response: expand past the comfortable stopping point during discovery, then contract ruthlessly during pruning, so the count is set by what survives attack rather than by what fits the rhythm.

Confidence tagging is mandatory. Every prediction in the draft is labeled. Mechanism logic — the prediction follows from the incentive structure alone, near-tautological once the structure is specified ("confiscating surplus creates incentive not to produce surplus"). State plainly. Structural claim — the prediction follows from institutional or market analysis but could be wrong if the analysis is wrong ("thin markets produce monopoly pricing"). Flag the reasoning chain that the prediction depends on. Behavioral prediction — the prediction is sociological or psychological, dependent on actors' context, alternatives, inertia ("the sanctioned population responds to the sanction by withdrawing rather than complying"). State the direction; state that the magnitude depends on empirical work the analysis has not performed. Speculative cascade — the prediction depends on prior predictions in a chain ("if the previous behavioral response materializes, the next stage is..."). Present explicitly as conditional speculation. Mismatched confidence destroys credibility. A draft that asserts behavioral magnitudes with mechanism-logic confidence is worse than a draft that admits the magnitude is unknown.

The first version will be wrong in important ways. The errors will surface under attack. The value of the first draft is in having something concrete to attack, not in being right.

VII. Dialectical iteration

The pattern: round one finds errors inside the proposal's frame. Rounds two through four find the frame.

A round-one analysis applied with the lens kit and the typology produces competent within-frame work — the proposal's stated mechanisms tested against rational actors' likely responses, the obvious capital-stock impacts catalogued, the first-order failure modes identified. Round one is necessary and rarely sufficient. The structural findings — the moves that question what round one treated as fixed, the reversals that question round two's premise, the exhaustive enumerations that hold regardless of which diagnosis is correct — are downstream of treating the first round as raw material rather than as product.

The discipline that makes the iteration work is anti-defensive. When an attack reveals an error, the move is to accept the correction, not to defend the original. Defense is sophistry. Following the logic where it leads is the method. Practitioners who cannot distinguish their attachment to a draft from the draft's correctness produce work that calcifies under attack rather than improving. The dialectic only operates if the analyst treats their own drafts the way they treat the bill's drafters — as fallible producers of structured claims that may be wrong in their framing as well as their details.

The substrate that performs the dialectic varies. It may be a disciplined solo analyst rotating through perspectives, analysts on a team, expert reviewers brought in to attack a draft, adversarial peers in a structured review process, AI systems prompted to argue against the draft from divergent priors, or any mixture. The implementation substrate is not the method. The method is the sequence of attacks. The heuristics divide unevenly across substrates. Some — constraint inventory, frame-level bypass, meta-reflexivity, recursive decomposition — are mostly internal and a disciplined solo analyst working carefully performs them well. Others — convergence suspicion, feature-bug inversion, the discomfort-signal check — gain disproportionately from external attackers because they require framings the analyst would not have adopted unprompted, and cognitive lock-in resists self-administered versions of the move. A skilled solo practitioner can do substantial mechanism analysis. External dialectic, whether human or AI-assisted, strengthens the work and is especially valuable for the moves that depend on genuinely orthogonal perspectives or on parallel attack from divergent priors at scales no single analyst can sustain. The substrate will change as tools change. The moves themselves do not depend on the technology that performs them.

The nine heuristics

Each heuristic is a move applied to a structured output — a draft analysis, an enumeration, a causal chain, a proposed alternative. The heuristics are not sequential; they are tools applied where they bite.

The Unpopulated Meta makes the stronger claim that this kind of named-heuristic codification approximates a substantial fraction of the meta-level function mechanically — that anyone with a language model can execute the heuristics and the limiting factor shifts from capability to awareness. That claim is about the operations themselves. The judgment claim of Section II is about the surrounding work: knowing which heuristic to apply to which output, recognizing when an answer is structurally wrong rather than merely incomplete, integrating the heuristics' outputs into a coherent revision. The operations are codifiable. The application discipline is trained, not transferred in prose. Both can be true at once: the heuristics are mechanically executable, and the method's effectiveness still depends on the practitioner's calibration.

H1. Constraint inventory. List everything the analysis treats as given. The bill's domain, the policy goal, the institutional architecture, the parameter range, the actor mandates. Each item on the list is a candidate for questioning. The heuristic is cheap — one enumeration pass — and it surfaces the assumptions the analysis is making by treating them as background.

H2. Feature versus bug inversion. For every constraint the analysis flagged as a problem, ask: what if this is a correct design choice? The inversion sometimes reverses the finding. A constraint identified as a barrier may be a load-bearing filter. A friction identified as inefficiency may be the mechanism that prevents a worse failure mode. The discipline of running the inversion as a hypothesis — not a conclusion — surfaces cases where the blank-slate critique would have removed a feature the system needed.

H3. Convergence suspicion. When multiple analyses or attackers converge on the same finding, treat the convergence as a signal to test harder, not as confirmation. Convergence among perspectives that share the same training, priors, or analytical tradition is ambiguous: it may be shared truth, or it may be shared bias. The mechanical version of the heuristic: when three or more independent attackers agree, launch a fourth with the explicit instruction to assume the others are wrong. The fourth perspective is the test of whether the convergence is signal or shared blind spot.

H4. Exhaustive enumeration. When conflicting diagnoses exist for the same observed failure, ask whether there is a logical structure that holds regardless of which diagnosis is correct. The structure becomes the finding, not either diagnosis. The mechanical version: given conflicting hypotheses A and B, enumerate all options under A, all options under B, identify what is in both sets, and identify what is in neither. Options that are not viable under either hypothesis are not viable; the political consensus that assumes the existence of an option outside this enumeration is taking a position the analysis does not support. Cohort-dependent option collapse is the deepest version of this move: an option viable for one subpopulation collapses for another because the cohort's baseline characteristics make the option's prerequisites unreachable. The enumeration tightens further when this collapse is recognized.

H5. Frame-level bypass. Before resolving an empirical sub-question, ask whether the structural conclusion changes depending on the answer. If it does not, flag the sub-question as unresolved and proceed. Analytical effort spent resolving questions that do not change the finding is effort not spent on questions that do.

H6. Meta-reflexivity. Apply the framework's tools to the framework itself. If the framework says "name the mechanism," name the mechanism by which the framework's own lenses produce their findings. If the framework says "question assumptions," question the framework's assumptions. The heuristic catches cases where the framework's own vocabulary is doing work it should not — labels that compress what should be decomposed, terms with smuggled priors, distinctions that paper over collapses.

H7. Discomfort signal. When a finding is uncomfortable from every political direction simultaneously — the left dislikes one option, the right another, technocrats a third, populists a fourth — the analyst treats it as a useful anti-contamination signal. The structure is producing a finding that is not obviously aligned with any pre-existing narrative. That does not prove the finding correct; it means the finding deserves harder structural testing rather than immediate dismissal as partisan. When a finding aligns neatly with one faction's prior commitments, the analyst checks for frame contamination. The faction-aligned finding may still be correct, but it is suspect in a way that the faction-orthogonal finding is not.

H8. Iteration budget. The deepest findings emerge in rounds two through four. Never stop after round one. The minimum discipline: after within-frame analysis is complete, always run at least the constraint inventory (H1), the feature-bug inversion (H2), and the recursive decomposition (H9) before declaring the analysis done. The other heuristics apply where they bite.

H9. Recursive decomposition. After producing any structured output — an exhaustive enumeration, a causal chain, a mechanism inventory, a proposed alternative — decompose each element and verify its internal mechanistic coherence before treating the output as done. This is not dialectical attack from outside; it is decompositional trace from inside. For each element, ask: what does this actually require? What are its prerequisites? Does it remain independent under realistic conditions, or does it collapse into another element? Labels that sound atomic are often bundles. Lower the threshold, invest sufficiently, accept the cost, restructure the system — each of these labels survives sophisticated dialectical analysis intact, but a thirty-second decomposition often reveals that the label hides a structurally different move, or that the labeled option collapses into another option under realistic conditions. H9 is the cheapest heuristic and the highest yield per minute. It catches a class of error that external attack systematically misses because external attack targets the relationships between elements while decomposition targets the content within them.

A worked pattern, abstracted

The shape of an iterative analysis on a hypothetical population-intervention bill — say, a generic welfare-to-work activation scheme that modifies support durations and benefit interactions — compressed. The example demonstrates the rhythm of iterative deepening, not the rightness of any particular finding. A reader who wants to check whether the heuristics produce defensible findings on real cases should run the same shape against a historical mechanism failure where outcomes are known (Section X). The worked pattern teaches the shape; the practice cases test the validity.

Round one (within-frame). The bill modifies parameters of an existing program — shorter support durations, lower benefit levels, new fee structures. Standard analysis: four mechanism errors. Funding cut without corresponding obligation reduction. Sanction whose collection cost exceeds its revenue. Time parameter that ends support before the target threshold is reached. Erosion of a fallback mechanism the displaced population depends on.

Round two (H1 applied). The constraint inventory reveals what the analysis treated as fixed: that the state provides this category of service, that the current service model is the relevant baseline, that the labor market the program prepares people for is what it is, that the parameter range under debate is the range that matters. Each assumption is a candidate for questioning. A new round of analysis blind to the existing law's framing produces a different finding: the binding constraint is not in the service the bill modifies but in the labor market the service prepares people for. The bill is optimizing the translation layer while the receiving system rejects the input regardless.

Round three (H2 and H3). The convergence in round two — multiple independent attackers identifying the same labor-market binding constraint — invites suspicion. The feature-bug inversion runs the constraint as a design choice rather than a problem. What if the receiving system's selectivity is a load-bearing filter that prevents a worse failure mode, not a barrier to remove? The reversal produces a counter-finding: the bill's failure mode under "remove the barrier" is structurally different from its failure mode under "the barrier is a feature," but in both cases the parametric debate (two years versus three years of support) is happening in a window narrower than the actual design space.

Round four (H4 and H5). The conflicting diagnoses from rounds two and three — barrier versus feature — invite exhaustive enumeration. If the receiving system's structural property is X, exactly N coherent options exist. If the property is not X, the options are the same minus the ones that depended on X being true. The political consensus assumes an option outside this enumeration exists. The enumeration shows it does not. The structural finding holds regardless of whether the barrier-versus-feature question is resolved. Frame-level bypass applies: the analysis does not need to resolve that empirical question for the structural conclusion to hold.

Recursive decomposition pass (H9 applied to the round-four output). Each option in the enumeration is decomposed. Option 2 — "invest sufficiently to cross the threshold" — assumes the threshold is reachable for the target population. For some cohorts within the population, the gap between baseline characteristics and the threshold is wide enough that no realistic investment closes it within any reasonable timeframe. Option 2 collapses for those cohorts; the enumeration tightens further. Option 4 — "restructure the system" — turns out to bundle structurally different moves, one of which collapses into another option under realistic conditions. The decomposition pass is often cheap and catches errors that survive much more expensive dialectic.

The pattern: round one was competent delta-analysis correct within the bill's frame. Rounds two through four found the frame. The recursive decomposition pass found that elements in the analysis's own output were internally incoherent. The marginal value of each additional layer was higher than the previous, because each layer questioned a deeper assumption.

Exit conditions

The dialectic ends when the remaining gaps are factual rather than structural. When the highest-severity attack against the draft is "I am not sure whether the claim about the historical comparable case is accurate," the move is not another round of perspective-rotation; the move is to verify the historical claim. Perspectives generate plausible-sounding factual assertions in character; the cost of believing them is high, the cost of checking them is low. Exit the persona loop when the work shifts from "test the analysis's structure" to "verify the analysis's facts."

The dialectic's exit condition is about the analyst's own work. It is not a prediction that the analyzed system will adopt the analysis. Trapped Equilibria describes the structural pattern in which captured insiders cannot adopt globally better frames because adoption is not a legal move on the local board — concentric capture bonds epistemic, status, material, and coordination positions tightly enough that defection on one axis triggers cliff-edge losses on the others. A structurally correct mechanism analysis can be ignored for reasons the analysis itself names. That ignorability is a finding about the destination's gate, not a failure of the method. The discipline is to exit the analyst's iteration when the analysis is structurally sound, and to recognize the separate question of whether any actor with the standing to apply the finding can pay the local cost of acting on it.

VIII. Assumption audit and reflexive failure modes

The dialectic operates on the draft. Some moves operate on the framework itself.

Assumption audit. At least once per analysis, run an explicit audit of what the analysis treats as background rather than as testable claim. The audit asks of every assumption: is this a physical constraint, an empirical finding, or a consensus assumption inherited from convention? For each, attempt to backpropagate it to the proposal's purpose. Does the assumption derive from the goal the analysis serves, or is it floating free — accepted because of consensus, institutional inertia, or unexamined moral priors? Assumptions that align with the social or institutional consensus pass unexamined unless the audit forces them to surface, because they do not feel like claims. The deepest mechanism failures the analysis can identify are sometimes failures of the assumptions, not of the mechanisms within the assumptions' frame.

Capacity paradox check. If the proposed alternative requires a capability the system currently lacks — if the analysis just demonstrated that the state cannot do substantive oversight, and the alternative requires the state to do substantive oversight — the alternative must address the bootstrap problem. What mechanism builds the missing capacity? Without this, the analysis is doing exactly what it criticized the original bill for doing: assuming the solution into existence. The check does not require the analyst to drop the alternative; it requires honesty about the alternative's prerequisites. The reader should see what the logic requires, what the requirement depends on, and how to build what the requirement depends on.

Legibility trap awareness. The framework's analytical lens makes certain features of reality visible — incentive distortions, behavioral responses, capital-stock depletion — and others invisible — tacit knowledge, organizational culture, individual judgment, political relationships. Non-legible factors may be decisive. The dialectic partially compensates by introducing perspectives with different lenses, but the risk is systematic: all perspectives applied within the same analytical tradition share the same broad blind spots. The mitigation is not a fix; it is awareness. When proposing alternatives, ask what non-legible factors could make the alternative fail even if the mechanism design is correct. The map is not the territory.

Execution-complexity trap. Theoretically superior mechanisms can underperform simpler ones because implementers rationally simplify the design in practice. A mechanism that requires implementers to track fifteen variables across three time horizons produces, in the field, an implementation that tracks the variables the implementers understand and ignores the rest. The simplification is uncontrolled and destroys the formula's assumptions more thoroughly than the simpler mechanism it replaced. When proposing alternatives, the analyst asks: does this require more implementation capacity than the current mechanism? If so, is the theoretical gain worth the execution-quality loss that follows from rational implementer simplification? Sometimes the correct answer is to stop formula-governing entirely and use a coarser, more legible mechanism that survives implementation degradation.

Political-palatability filter rejected. The analysis traces logical implications of all branches. A conclusion that every political faction finds uncomfortable is a signal the analysis is working, because the finding is not aligned with any faction's pre-existing narrative. Self-censoring an analytical branch because the conclusion "sounds" wrong — too unsympathetic, too radical, too aligned with a faction the analyst dislikes — produces analysis that confirms the analyst's priors rather than revealing the constraint structure. The exhaustive enumeration technique is the tool for presenting uncomfortable findings without triggering defensive reactions: enumerate all options, let the reader see that their preferred option has costs, do not advocate for any option.

Failure modes of the method

Patterns of method failure to watch for.

• Spherical cow bias. Clean-sheet designs are internally consistent but ignore path dependencies, institutional memory, transition costs, and political economy. The counterfactual-design output sounds elegant; the actual transition path is impossible from the current state.
• Smuggled priors. "Assume zero policy" is not a neutral starting point in a welfare state where extensive intervention is the baseline. The counterfactual starting point shapes the output. Blank-slate designs in a welfare state context smuggle non-interventionist priors; blank-slate designs in a minimal-state context smuggle interventionist priors. The starting point is itself a choice.
• Oracle illusion. Three independent perspectives converging on the same finding feels robust. The convergence may reflect shared training priors, not independent evidence. The discipline of H3 (convergence suspicion) is the response, but the failure mode recurs because convergence reads as confirmation by default.
• Constraint-blindness. Blank-slate designers systematically treat existing constraints as bugs. Some constraints are load-bearing features. The discipline of H2 (feature-bug inversion) is the response, but the failure mode recurs because the analyst's framing is built around what is wrong with the current state, not what the current state's frictions might be preventing.
• Investment-ceiling blindness. Options that propose "invest sufficiently to solve" assume investment has no ceiling. For cohorts within the population whose baseline characteristics place the threshold above any realistic investment level, the option collapses. The analysis that does not check this produces policy that wastes resources on impossible transitions while underinvesting in achievable ones.
• Production-rhythm convergence. Discussed above. The analyst's tendency to converge on a comfortable output volume per analysis, independent of input variance. The expand-and-contract discipline applied with the phases separated is the response.
• Zero-shot structured outputs. Each element of an enumeration or causal chain is a compressed label that may be internally incoherent. The discipline of H9 (recursive decomposition) is the response.
• Self-censoring on political grounds. Branches of the analysis dropped because the conclusion sounds wrong to one of the analyst's political reference groups. Discussed under the political-palatability filter.
• Overclaiming on speculative cascades. Sociological predictions presented with mechanism-logic confidence. A small parameter change in a benefit structure does not produce "radicalization," "broken families," or "social collapse" with the same epistemic standing that it produces "reduced compliance" or "reduced labor-market participation." The hard-nosed adaptation check asks: what if the actor adapts? What if the short-term pain produces long-term behavioral change? What if the devastating scenario has a 10% probability rather than 90%? The check catches cases where the analysis has adopted the framing of the social sector or advocacy groups defending the status quo. Mechanism analysis serves the analysis of the bill's mechanisms, not the rhetorical interests of the actors affected by the bill.
• Confusing capacity with incentive. Actors may behave badly because the mechanism rewards bad behavior (incentive-constrained) or because the actor cannot do what the mechanism requires (capacity-constrained). Conflating the two produces fixes that assume willingness when capability is the binding constraint, or vice versa. The repair specification differs by an order of magnitude depending on which constraint is binding.
• Mistaking political disagreement for mechanism error. Not every policy disagreement is a mechanism error. The discipline of the three-type classification is the response: a finding that cannot be classified as Type 1, Type 2, or Type 3 is not yet a mechanism finding.

IX. Epistemic honesty

The temptation in any analytical method is to claim more than the method establishes. The opposite temptation, equally corrosive, is to retreat into false humility — "these things are unknowable" — when the truth is "this work has not yet been done." Both temptations produce analysis that mislabels its own confidence and thereby destroys its credibility with the audience the artifact most needs.

The honest framing is to stratify by what reduces uncertainty, organized as tiers of effort required.

Tier 1, effectively certain conditional on the specified mechanism. Mechanism logic. The direction of incentive response is derivable from the structure of the incentive alone, assuming the payoff surface is correctly specified and the actor can perceive and respond to it. Confiscate the surplus, and actors have less incentive to produce surplus. Penalize the metric, and actors optimize the metric. The claims at this tier are near-tautological once the mechanism is specified. The analyst states them plainly, without hedging, while keeping the specification's correctness on the table as the condition the certainty rests on.

Tier 2, high confidence where boundary conditions match. Structural patterns. Empirical regularities well-attested across institutional cases when the relevant architecture is present. The soft budget constraint in central-financing systems without local revenue authority. The concentration-diffusion asymmetry in regulated industries with organized incumbents. The metric-target divergence in formula-driven funding. These patterns are not theorems; they are default predictions when the architectural preconditions are present, and the analyst checks that the preconditions hold before invoking the pattern.

Tier 3, moderate, reducible with work. Behavioral magnitudes. The direction is given by mechanism logic. The magnitude depends on actors' context, alternatives, inertia. The magnitude can be estimated through comparable historical cases, behavioral economics literature, simulation, or expert elicitation. The analysis usually does not have time to perform this work within its production budget. The honest move is to name the bound the analysis can defend, name what would tighten the bound, and refuse to substitute confidence for evidence.

Tier 4, declared bounds. Deep-time and system-level trajectories. The framework requires deep-time analysis because the temporal horizon of the goals the analysis serves exceeds any single electoral cycle. The honest move is to model the trajectory with explicit assumptions, state where the chain branches on unresolved empirical questions, and apply frame-level bypass (H5) wherever the structural conclusion holds regardless of which branch is correct. Deep-time projections are not unknowable; they are uncertain in ways that disciplined modeling can characterize.

Tier 5, mitigated, never eliminated. The framework's own legibility limits. The lens makes some features of reality visible and others invisible. Mitigation comes from meta-reflexivity, cross-domain pattern transfer, and dialectical attack — but the dialectic at this tier only works to the extent the attackers operate under priors genuinely outside the framework's analytical tradition. Convergence-suspicion (H3) flags the failure mode: attackers drawn from the same canon Section II names supply divergent roles within shared priors, not divergent priors. Escaping framework blind spots at Tier 5 requires perspectives from outside the framework's tradition — sociologists of organizations on tacit knowledge and institutional culture, anthropologists on relational structures, practitioners with ground-truth experience the framework's vocabulary does not encode, historians on how comparable mechanisms failed for reasons the analytical literature did not surface until decades later. These perspectives are harder to summon than intra-canon adversaries and the method is honest that the mitigation at this tier is partial. The legibility limits are named, not hidden behind. Where the method's lens cannot see, the artifact says so.

The discipline that follows: state your current confidence; state what would raise it. Do not substitute "unknowable" for "I did not have time to model this within the artifact's production budget." False humility is itself an analytical failure — it hides the gap between what the artifact claims and what disciplined work could establish. The analyst whose method consistently retreats to Tier 5 is producing analysis that has stopped doing work.

Mechanism analysis is not omniscience. It is disciplined non-omniscience. The artifact's authority comes from showing its uncertainty in the right places — high confidence where mechanism logic supports it, named bounds where empirical magnitude requires more work, explicit branches where deep-time trajectories diverge on unresolved questions — not from achieving uniform certainty across all claims.

Why the method is assertive within its scope. Mechanism analysis is not the whole epistemic standard for evaluating legislation — legal coherence, fiscal accounting, distributional analysis, constitutional review, and the qualitative judgment of practitioners with ground-truth experience all matter — but it is a minimum viable component of that standard, and one that is routinely absent. Any structurally consequential proposal that does not model behavioral responses to its own incentive structure has not done the analysis it claims to have done. The bill's own impact assessment typically assumes static actors who do not change behavior in response to the law. A dynamic model that accounts for actors optimizing against the mechanism is a priori closer to reality than a static model that does not. When the analysis is stronger than the proposal's, the artifact says so. False balance — "some say the bill works, some say it does not, judgments differ" — is analytical evasion when the analysis has been done. The artifact's authority depends on the analyst's willingness to state, plainly, what the disciplined analysis establishes.

X. Apprenticeship: how the taste is built

The method documents what can be transferred in prose. The remainder is built only by practice.

Four components of the apprenticeship pattern.

Practice on known failures. Analyze a historical mechanism failure. Predict what mechanism analysis at the time of the law's enactment would have surfaced. Check the prediction against the recorded history of how the law actually failed. The exercise is cheap, the feedback is unambiguous, and the calibration accumulates. Soviet output-by-weight quotas, Mexico City's license-plate driving restrictions, Wells Fargo's cross-selling quotas, the EU Common Agricultural Policy's land-use distortions — each is a worked case where the mechanism failure was structural, predictable from the incentive design, and produced a behavioral response the policy's drafters did not anticipate. The analyst's task is to anticipate the response from the design alone, and then to check against the recorded history.

Production with external attack. Produce mechanism analyses. Expose them to attackers operating under different priors than the analyst's own. The attackers may be peers, expert reviewers, structured adversarial AI agents, or a combination. A disciplined solo analyst can do substantial work — rotating through perspectives, applying the internal heuristics carefully — but solo work without periodic external dialectic tends to ossify, because cognitive lock-in resists self-administered attack on the analyst's own framing. The discipline of accepting attacks and following the logic where it leads — defense is sophistry, following the logic is the method — is built fastest by repeated exposure to attacks the analyst did not anticipate.

Feedback from outcomes. When a mechanism analysis predicts a behavioral response and that response can be checked against later evidence, the practitioner learns what the method's actual hit rate is. The feedback loop is slow — laws produce their downstream effects over years or decades — but it is the feedback that distinguishes practitioners whose taste is calibrated from practitioners whose taste is confident without basis. Track predictions explicitly. Compare to outcomes when outcomes become available. The taste that consistently misses in the same direction is taste that has not yet developed. Calibration is the operational definition of taste; in its absence, taste is unfalsifiable, and the method's epistemic foundation rests on assertion.

This standard binds the present essay. The method described here has been applied to a small corpus of legislative proposals; the systematic outcome-tracking that would calibrate the method's predictions against later empirical evidence is started but not yet completed. Until that calibration accumulates across years of paired predictions and outcomes, the method's authority rests partly on the structural coherence of its moves and partly on assertion. A reader who picks up the method should treat its predictions on a new proposal with the same confidence stratification the method itself demands of any analysis — mechanism logic with high confidence, behavioral magnitudes with explicit bounds, the method's overall hit rate as a Tier 3 claim awaiting evidence. The method does not exempt itself from its own discipline.

Cross-domain pattern transfer. Practitioners who have seen mechanisms fail in multiple domains — finance, military, software, biological systems, market design, organizational behavior — develop a pattern recognition that single-domain analysts lack. The same structural pattern recurs across domains under different vocabularies. The cobra-effect of induced supply, the moral hazard of socialized losses, the Goodhart distortion of metric-target substitution, the principal-agent gap of mandate-incentive mismatch — these are not domain-specific phenomena. The analyst who has seen them in three or four domains recognizes them in the fifth before the local literature has named them. The apprenticeship benefits from breadth as well as depth.

The method can be documented. The judgment must be trained. A mechanism analyst is not someone who has memorized the lenses or the heuristics. A mechanism analyst is someone who has watched enough mechanisms fail across enough domains that the next failure becomes visible before it happens.

XI. Closing

A mechanism analysis is a way to become uncertain in the right places, not a way to become certain.

The method requires the analyst to write down the evidence bundle, reconstruct the proposal's causal model, enumerate the mechanisms before judging them, scan the capital stocks the proposal affects, rotate through the analytical lenses, attack the first draft, decompose the labels that survived attack, and tag the confidence of every prediction. The result is a disciplined readout, not an oracle's verdict: what the proposal claims, what the mechanism makes rational, where the pressure routes, what stock absorbs the cost, and what would have to change for the machine to work.

The rules can be written down. The taste cannot. The taste forms through production: analyze known failures, publish, be attacked, compare to outcome, repeat. Across enough iterations, the practitioner develops the ability to see the failure mode the proposal's drafters did not see. The method does not guarantee this. The method makes it possible.

The artifact is the test bench. This essay describes how the bench is built.

Related:

• The Mechanism Analysis — the artifact this method produces.
• Mechanism Realism — take the mechanism, not the stated intent, as the load-bearing causal variable.
• Bad Equilibria Are Not One Thing — the broader taxonomy of stuck-equilibrium primitives the Failure Readout specializes for legislation.
• Constructive Diagnosis — the repair specification standard the artifact implements.
• Full Accounting — "hard to measure" is not the same as zero; approximately right beats precisely wrong.
• The Response Vector — pressure routes through the cheapest available channel.
• Calculemus — most legislative disagreement is uncomputed empirical disagreement.
• Flourishing Is Maximum Safety Margin — the objective function the analysis ultimately serves.
• The Unpopulated Meta — why the engineering meta level where this method's institutional substrate would live is structurally near-empty.