The Refusal to Compute

I. The number that would have bound the decision

In October 2021 the UK government published its Net Zero Strategy, a long policy document setting out how the country would reach net-zero emissions by 2050. The strategy named the target, the sector pathways, and the quantified policy mix. The Secretary of State was told that the quantified policies added up to about 95% of the emissions reductions needed for the legally binding sixth carbon budget, with a shortfall remaining. He was not given the policy-by-policy breakdown or an explanation of how the remaining shortfall would be made up.

Friends of the Earth, ClientEarth, and the Good Law Project brought judicial review. In July 2022 the High Court found a section 13 breach of the Climate Change Act 2008 because the Secretary of State had legally insufficient information before him when he approved the strategy, and a section 14 breach because the published strategy lacked the information needed for Parliament and the public to scrutinize delivery risk. The calculation existed inside the department. It had not been made admissible on the record the minister was held to.

The relation existed. It could be bounded. It was bounded inside the department. Bounding it as an admissible procedural object would have created an obligation — to fund, to redesign, to explain the gap. The strategy was published without that object. The Court held that the calculation had to become part of the record the decision could be judged against.

This is the structural pattern. The method existed; the data existed; the calculation existed. The refusal sat at the point where the calculation would have become the kind of object that, once on the record, would bind the decision.

II. What refusal to compute is

A computation is a bounded relation between a decision and a consequence. The bound need not be precise. It needs to be tight enough that the decision-maker can no longer act as if the consequence were zero.

Four states are easy to confuse:

Ignorance and uncertainty are honest epistemic states. Value disagreement is a legitimate domain of democratic choice. Refusal to compute occupies a different position: the relation could be approximately bounded, the bound would constrain the decision, and the institution declines to produce or admit the bound. The declination is then redescribed as one of the other three states — complex, uncertain, a matter of judgment, a matter of values, insufficient evidence, case-by-case discretion, future review.

Two operational variants share the same primitive. Production refusal: the bound is never created — the cost factor stays un-revised, the cohort schedule is never built, the correlation model is not applied. Admission refusal: the bound exists internally but is not made into an admissible object on the decision record — the calculation sits in the department; the decision is signed off without it. The repair differs. Production refusal requires the bound to be produced. Admission refusal requires the existing bound to become a procedural object on the record, where ownership and traceable consequence can attach. The diagnostic in §III applies to both; the variant determines which repair the institution owes.

The redescription is the move. Linsey McGoey’s work on strategic ignorance names the broader pattern — power includes shaping the boundary between knowledge and ignorance, and ignorance can be a resource. Refusal to compute is the policy-process subtype: a calculative relation that could be made admissible is kept inadmissible, because admissibility would attach obligation. Bachrach and Baratz called the more general version nondecision-making: agenda control as the second face of power. The refusal-to-compute version keeps not a policy option but a calculation off the agenda.

III. The obligation-bound test

Four gates distinguish a refusal-to-compute from honest non-computation.

1. Relation. A decision-relevant relation exists. Reform X changes downstream cost Y. Policy X shifts risk to group Y. Project X displaces alternative use Y. The relation may be probabilistic, lagged, or distributed across systems, but it exists in the territory the decision purports to act on.
2. Bound. The relation can be bounded well enough to constrain the decision. Not precisely measured. Not settled forever. Bounded enough that zero — or, more precisely, the operational-zero status created by leaving the relation unowned and unrecorded — is no longer defensible. The decision-sensitivity test: a bound is tight enough when all plausible values within it would require a materially different justification, mitigation, funding decision, or delay than the implicit-zero baseline.
3. Obligation. The bound would assign someone a duty. The duty may be to justify the decision against the bound, to fund the implied liability, to mitigate the implied harm, to compensate the implied loss, to redesign the policy, to delay the closure, or to stop. The duty need not be triggered automatically; it is enough that the bound, once on the record, makes the decision answerable to it.
4. Unbounded substitute. The institution substitutes unbounded language for the bound. Too complex. A matter of values. Insufficient evidence. Stakeholder disagreement. Future review. Case-by-case discretion. The substitute is the diagnostic surface. It is also the place where refusal looks identical to legitimate non-computation, and where §V matters.

When all four gates are met, the non-computation is a governance act. The institution is not failing to know. It is preserving discretion by preventing the relation from becoming the kind of object that, once on the record, would constrain the discretion.

IV. Four forms of refusal

The same primitive shows up across domains as four recurring forms, each carrying a different question. The four are not exhaustive — latency refusal (time-lagged effects), cumulative refusal (aggregation of individually-defensible omissions), and boundary refusal (relations crossing institutional borders) are real adjacent types — but the four below cover the cases that recur most often in deployed governance.

Burden refusal. What does this really require? The institution counts adoption, headline cost, or short-horizon appropriation, but does not bound the full operating burden — maintenance, replacement, enforcement, training, exit. The US Department of Defense facility-sustainment backlog reached at least $137 billion in fiscal year 2020. The Government Accountability Office found that the department’s sustainment funding model did not fully account for facilities exceeding expected lifespans, and that the department had not revised the sustainment cost factors that translate facility characteristics into required funding. GAO recommended a revision. The department non-concurred. The cost-factor calculation was the gate, not the data. Bounding the cost factor would have required either funding the implied liability or accepting a documented underfunding trajectory on the record. Both consequences were institutionally costly. The model stayed un-revised.

Incidence / sustainability refusal. Who carries it, and who finances it? Aggregate cost or aggregate benefit is computed; the distributional and fiscal-sustainability bounds that would name the bearer and the financing constraint are left unbounded. The UK Treasury’s September 2022 mini-budget is a recent specimen: the package announced tens of billions of pounds of unfunded tax cuts and energy support without the Office for Budget Responsibility’s customary economic and fiscal-sustainability forecast. The institutional machinery for the bound existed and was excluded from the announcement; the substitute language was discretion over publication timing. Outside analysts (Resolution Foundation, IFS) produced incidence and sustainability bounds within days, and the gilt market did its own pricing within hours. The test cuts both ways: a welfare expansion or universal-benefit reform that announces aggregate cost without bounding cohort-incidence or long-run intergenerational distribution is the same refusal in the opposite political direction.

Displacement refusal. What moves elsewhere? The chosen object is bounded; what it pushes out, exports, crowds out, or relocates is not. The clearest documented financial-markets specimen is in the US Senate Permanent Subcommittee on Investigations record: in July 2006 Standard & Poor’s had changed assumptions for new residential mortgage-backed securities ratings, but did not apply the revised model to reevaluate existing RMBS securities then under surveillance. Applying the revised model would likely have forced downgrades that hurt revenue, embarrassed already-issued ratings, and propagated through the institutions holding the paper. The calculation existed. The displacement of correlation risk across the system was the object the recomputation would have made visible. S&P declined to make it visible until the market did the job. The broader Senate investigation treats the rating agencies’ delayed downgrades as part of the financial-crisis mechanism.

Path refusal. What future does this lock in, compared with what? Actions are treated as reversible, neutral, or beneficial; the counterfactual world and the irreversibility of the chosen path are not bounded. The UK Net Zero Strategy specimen above is path refusal in its admission variant: the policy-level quantification existed inside the department, but was not made admissible on the record the minister was held to. The Court did not invent the calculation. It forced the calculation into the kind of object the decision could be measured against. The S&P and DOD specimens are the production variant: the bound was either never built or, in the rating-agency case, built but not applied to the population it would have constrained. The diagnostic is identical; the repair differs.

The forms overlap. A single decision can refuse on multiple gates and in either variant. The taxonomy helps the diagnostic; it does not define disjoint species.

V. When not computing is legitimate

Not every uncomputed relation is a refusal. The §III gates require the bound to exist and to constrain the decision and to create an obligation and to be replaced by evasive language. Five categories of non-computation pass the gates legitimately.

Side constraints. Some relations are protected from cost-benefit override by design: rights, constitutional limits, statutory prohibitions on optimization. The point is not that no facts matter; the point is that the calculation cannot decide the permissibility question. A computation of how many violations of habeas corpus would be efficient is not a calculation the institution declines to make for evasive reasons. The constraint is the answer. The US statutory restriction on dollars-per-QALY thresholds (42 U.S.C. §1320e-1) is a hybrid case: a declared legal refusal to let a particular cost-effectiveness measure govern Medicare coverage or reimbursement, enacted through democratic value choice on disability-discrimination grounds. The refusal is statutory, explicit, and contestable. That is different from hidden evasive non-computation: the side constraint is named on the record, and the obligation it forecloses is named with it.

Corruptive measurement. Some computations damage the domain they enter. Measuring teacher quality by student test scores produces score inflation and curriculum narrowing; measuring researcher quality by citation counts produces citation gaming; measuring care quality by waiting times produces threshold gaming under target pressure. The Measurement Anchor names the maintenance failures. The refusal-to-compute analogue: where the act of measurement under consequence reliably corrupts the relation it bounds, declining the bound is not evasion. It is recognition that the bound under that institutional design would degrade the territory.

False precision under radical uncertainty. Some relations sit at the limit of model competence. The legitimate response is target shift: compute the incentive architecture rather than the outcome horizon, the exposure mechanism rather than the long-run prediction, the scenario class rather than the headline number. The legitimate version of “we cannot bound this” is paired with the explicit statement of what can be bounded — exposure, mechanism, incentive structure — and the commitment to bound that instead.

Residual democratic choice after bounding. When the relations are bounded, the weights remain a political question. Democracy may decide how to trade liberty against security, equality against growth, present against future. What democracy may not honestly do is call a boundable empirical relation unknowable because the bound is inconvenient. Refusal at this layer redescribes the four-state table: a value disagreement is invoked where a refusal to compute is what is happening.

Premature procedural fixation. Making a bound into an admissible procedural object can corrupt the institution’s capacity to think on paper. Once a number is on the record, it becomes litigable, citable, and weaponizable. Institutions facing aggressive admissibility regimes may learn to compute less rather than more — to avoid creating discoverable intermediate analysis, to launder candid estimates through lawyered uncertainty language, or to fix a provisional number prematurely because revising it would look like bad faith. A bound may legitimately remain unofficial or provisional when admission would freeze a revisable judgment, destroy candid internal modeling, or convert a live uncertainty into a litigable pseudo-fact. The category carries conditions. The institution must preserve internal learning and revision, state explicitly why admission would corrupt the process, identify what weaker disclosure protocol is available (red-team review, periodic publication, post-decision audit), and avoid using “premature” as a permanent shield. The legitimate version of this refusal is a deferred-admission protocol with named triggers; the illegitimate version is indefinite informality.

The line: legitimate non-computation says this relation cannot or should not be reduced to the available calculation, or cannot yet be made admissible without damaging the decision process. Illegitimate refusal says we will avoid the calculation because it would make the trade-off visible. The four gates separate them.

VI. The diagnostic

A field test for any non-computation in a live policy process. Six questions, subordinated to the four gates.

1. Decision relevance. Would the calculation, if produced, affect a decision now on the table? If no, the non-computation is not refusal — it is irrelevance.
2. Approximation availability. Can the relation be bounded approximately, even roughly? If no usable bound exists, the non-computation is ignorance or radical uncertainty, not refusal.
3. Owner visibility. If the bound were on the record, who would have to own it? An owner with no duty makes the bound a footnote; an owner with a duty makes the bound a constraint.
4. Obligation creation. Would the bound, once on the record, create a duty — to explain, fund, repair, stop, or compensate? If yes, the non-computation is doing institutional work.
5. Substitute rhetoric. Is the absence of the bound being explained in language that does work no actual epistemic state requires — too complex, a matter of values, insufficient evidence, case-by-case? Evasion shows in the language before it shows in the gap.
6. Movement test. What later outcome — a court ruling, a budget surprise, a demographic data release, a market repricing — would reveal that the omitted relation mattered? If the answer is concrete, the relation was decision-relevant; if no plausible later observation can be named, the refusal diagnosis is weaker, and the gap is more likely innocent.

The diagnostic produces a record: the omitted relation, the available bound, the named owner, the resulting duty, the substitute language used, and the movement-test outcome that would confirm or refute the diagnosis. Without that record shape, the diagnostic is a blog checklist. With it, the diagnostic can become a procedural object: a record-shaped demand that the institutional route can be made to acknowledge.

VII. Wrong repairs and corpus relation

When non-computation surfaces as a problem, the instinct is to add: more data, more modeling, more consultation, more dashboards, more analytic capacity, more AI. Each fails on the same vector:

• More data without an owner produces a dataset, not a computation. The relation remains unbounded because no one is required to act on the bound.
• More modeling without a response duty produces optional appendices. The model sits in the impact assessment; the decision does not move.
• More consultation without a calculation trigger produces opinions. Opinions are not constraints. The institution can hear a hundred submissions naming the omitted relation and still publish the decision without bounding it.
• More dashboards without a baseline or counterfactual produces visible non-computation. The number is present; the relation it would constrain is absent.
• More AI without a legitimacy gate produces powerless intelligence: the calculation exists, the decision is not bound by it.

The structural repair operates one level up from analytic capacity. The obligation-bound test made procedural: when a non-computation passes the four gates, the institution is required to produce the bound, name the owner, and accept the duty — or publicly justify the refusal under one of the §V legitimate categories. The diagnostic gives the requirement a form that can be installed — in audit rules, consultation requirements, court standards, or legislative checklists. Without installation, the diagnostic is a blog post. With it, the test becomes part of the route a decision must travel before it closes.

The Refusal to Compute sits at Layer 7 of the Stack (Computation/adoption). Calculemus is the parent: it argues that policy disputes classified as value disagreements are usually uncomputed empirical disagreements, and proposes the architectural response. The Measurement Anchor is the Layer 4 sibling: it asks how a consequential proxy stays attached to territory; this essay asks how institutions avoid producing the constraining relation in the first place. Constructive Diagnosis supplies the wrong-repair methodology this essay specializes. Non-Compilation is the supply-side sibling: it asks why some questions never become parameterized; this essay asks why, once parameterized, they are not computed.

In the broader translation chain from intelligence to binding action, this essay covers the production and admission gates: the bound is either not produced, or produced but not admitted to the decision record. The downstream gates — accuracy, binding, execution, audit — are the territory of The Measurement Anchor, Powerless Intelligence, Implementation Ledger, and The Severed Map respectively.

VIII. Close

The UK Net Zero Strategy passed the form of approval and failed the test of admissible information: the calculation existed inside the department, but had not been made into the kind of object the minister or the public could be held to. The Court did not produce the calculation. It forced the calculation onto the record.

A measurement regime can drift under consequence. A computation regime can fail earlier, by leaving the relation that would constrain the decision unbounded, while redescribing the gap as uncertainty, complexity, values, or discretion. The institutional move is the same across burden, incidence, displacement, and path: keep the bound off the record, and the duty does not attach.

The diagnostic question before any consequential institutional decision: what relation here could be bounded approximately, who would own the bound, what duty would attach, and what language is being used in place of the bound? If the answers point to a refusal, the repair is not more data. It is the obligation-bound test, applied on the record, before the decision closes.

Related:

• Calculemus — the philosophical parent: most policy disputes classified as value disagreements are uncomputed empirical disagreements; this essay supplies the operational gate.
• The Measurement Anchor — the Layer 4 sibling: how a proxy stays attached to territory under consequence.
• Non-Compilation — supply-side sibling: when the question never becomes parameterized in the first place.
• The Procedural Object — the admissibility discipline that admission refusal violates and that repair invokes.
• Structural Residue — the orthogonal case to refusal-as-evasion: the gate correctly applies its rule and refuses the relief under its own grammar; the dependency the claim pointed at remains live and routes through other actors, scales, time horizons, or institutional channels.
• Powerless Intelligence — the downstream failure when computation exists but no authority binds.
• Constructive Diagnosis — the wrong-repair methodology this essay specializes for Layer 7.
• The Stack — Layer 7 (Computation/adoption) anchor.