The decision that cannot be averaged

A forecast can tell us which action wins across imagined repetitions. Wisdom begins when we ask what one failure would remove from the only life in which the action actually occurs.

Expected value is indispensable when outcomes can be represented with credible probabilities and a defensible value function. If a treatment succeeds seventy times in one hundred, a bridge design survives all specified loads, or a portfolio gains across repeated bets, averaging is informative. The trouble begins when the same grammar is applied to choices whose losses are not repeatable, fungible or cleanly compensable.

Suppose two decisions each have an expected loss of ten. In the first, a warehouse may lose replaceable stock. In the second, a community may lose the only wetland that protects its water source. The arithmetic can be identical while the decision structure is not. Stock can be repurchased. A destroyed ecological relation may be slow, uncertain or impossible to restore. The expected value calculation does not become false. It becomes incomplete because it leaves the change in future possibilities outside the frame.

From outcome value to option loss

Let the credible future options available at time t be a set A_t. An action a changes that set. A simple heuristic defines option loss as the difference between the weighted value of the options available before and after the action:

Here Ω is not a count of doors. A thousand trivial options do not outweigh one essential relationship or a community’s ability to recover. Each weight w(o) represents the defensible value of retaining that future. The equation is useful only when the options are sufficiently comparable and their bearers are represented. When they are not, Ω should remain a vector: ecological recovery, legal rights, livelihood, identity, trust and safety should stay visible rather than being forced into a counterfeit common unit.

This gives us a first distinction. Risk concerns uncertain outcomes within a choice. Irreversibility concerns how the choice changes what can be chosen later. A reversible mistake may be costly but informative. An irreversible mistake can destroy the very conditions needed to learn, compensate or try again.

flowchart LR
  A["The gate that can be opened once"] --> B["The decision that cannot be<br/>averaged"]
  B --> C["Finitude changes the problem,<br/>not just the budget"]
  C --> D["Restraint has option value"]
  D -. "re-test" .-> A
  classDef first fill:#0a1424,stroke:#007D68,stroke-width:2px,color:#e6f3ff
  classDef step fill:#0a1424,stroke:#3f7da0,stroke-width:1px,color:#e6f3ff
  class A first
  class B,C,D step

Finitude changes the problem, not just the budget

A common machine comparison treats human finitude as a resource limit. People have less memory, slower search and fewer years than a sufficiently scaled system. This is true but shallow. A token budget and a mortal life are both bounded, yet their bounds do different causal work.

Mortality is more than a small context window. It makes time non-renewable for a particular bearer. A machine process can be terminated, but termination becomes existential only if the process has a persisting subject, welfare or identity whose loss is not replaced by launching an equivalent copy. This paper does not assume that current systems meet that condition. It also does not declare that no artificial system ever could.

Four limits that should not be collapsed

Consider a decision meeting that is running out of time. The model has 30 seconds of compute left, the organisation’s licence expires at midnight, the chair must leave for hospital, and a riverside family will lose its home if the gate opens. Each fact is a limit, but the limits are not interchangeable. More hardware may remove the compute limit. A legal extension may remove the licence deadline. Another authorised chair may replace the first. No comparable substitution gives the family back the history of a destroyed home.

The difference is not sentiment added after calculation. It changes which remedies are coherent. Computational scarcity calls for compression, prioritisation or more resources. Institutional scarcity calls for mandate, sequencing and escalation. Existential finitude calls for attention to non-fungible bearers, relationships and time. Responsibility adds a fourth limit: the decision must remain attached to an agent or institution capable of explaining, repairing and learning after the action. A system that optimises all four as one “cost” can trade away the very distinction the decision requires.

This also explains why an artificial agent could, in principle, become relevant to the same analysis without merely imitating human vulnerability. The question would not be whether it says that deletion frightens it. We would need evidence of persisting, non-substitutable continuity; projects or commitments whose destruction is a loss to that bearer; and a governance relation capable of recognising the claim. Until then, machine shutdown is primarily an operational event, while human mortality remains an established existential constraint. Similarity of limits does not establish similarity of stakes.

Scarcity creates a tunnel

Herbert Simon’s bounded rationality replaced the fantasy of global optimisation with agents that must decide using limited information and computation. Scarcity research adds a sharper mechanism. When a resource becomes tight, attention is pulled towards the pressing demand. That can improve focus inside the tunnel while causing neglect outside it.

In laboratory studies, Shah, Mullainathan and Shafir found that scarce budgets changed attention and encouraged excessive borrowing of future resources. Their later high-powered self-replication supported the focus and over-borrowing effects more strongly than the proposed cognitive-fatigue effect. That distinction matters. The evidence does not justify saying that scarcity simply makes people less intelligent. It suggests that scarcity reallocates cognition, often towards the immediate problem and away from downstream costs.

A limited horizon reorders goals

Socioemotional selectivity theory offers a second mechanism. Laura Carstensen and colleagues argue that perceived time horizon, not chronological age alone, changes which goals receive priority. When the future feels open, people tend to invest more in exploration, knowledge and new relationships. When time feels bounded, emotionally meaningful goals and established relationships become more salient. Experiments and lifespan observations support the claim that time perspective is malleable and motivationally important.

This is not evidence that a shorter horizon makes every decision wiser. A preference for meaning can improve selectivity while reducing information seeking. Nor does age guarantee wisdom. Research on wise reasoning shows some age-related gains in social-conflict reasoning, but other traditions find early plateaus or large individual variation. A daily-diary study by Grossmann, Gerlach and Denissen found substantial within-person variability. People did not carry a fixed quantity of wisdom from situation to situation.

Psychological wisdom research describes several practices that help with this balance: intellectual humility, recognition of uncertainty and change, consideration of multiple perspectives, and search for workable integration. These are not decorations around intelligence. They are ways of reopening a field narrowed by ego, urgency or a single metric. Aristotle’s older account of practical wisdom makes a compatible point in a different vocabulary: deliberation concerns actions that can be otherwise, and it must reach the particulars of the situation rather than remain at the level of universal knowledge.

The result is a practical design claim. Do not assume wisdom is a stable trait possessed by a special person. Build situations that make wise reasoning easier to perform. A good decision environment preserves dissent, displays the cost of delay, separates reversible probes from irreversible commitments, and records who bears each consequence.

Restraint has option value

Economics gives a precise language for one part of this mechanism. When a decision is irreversible, the future is uncertain, and waiting can produce useful information, delay may have positive option value. Arrow and Fisher developed this logic for environmental preservation. Later work on irreversible investment showed why a project can have positive conventional expected value and still be worth staging or delaying.

The intuition is simple. Acting now gives the best action under current beliefs. Waiting allows the action to depend on new information. The value of waiting is the expected improvement in the chosen action, minus the cost of information and the harm caused by delay.

I is information that may arrive, a is an available action, U(a) is the value of its outcome, and the two costs cover gathering information and postponing action. The formula assumes meaningful probabilities, comparable values and a known action set. When those assumptions fail, the same logic can be used with scenarios and separate value dimensions instead of a single utility score.

The minimum commitment principle

The worked example points to a broader strategy: take the smallest action that creates useful evidence without surrendering the futures that evidence might recommend. A prototype, trial cohort, temporary permit, shadow deployment, escrowed decision, reversible delegation or limited-duration contract can convert passive waiting into active learning.

This is not “move slowly”. It is commit only the portion of the future needed to learn what the next commitment should be. The principle fails when small trials do not transfer, when delay imposes unacceptable harm, when a staged action itself creates lock-in, or when the affected people are being used as experimental material without authority or protection.

When preserving options becomes evasion

Optionality sounds universally attractive because the person retaining options usually experiences it as freedom. The cost may be carried elsewhere. A firm that postpones a factory closure preserves strategic flexibility while workers postpone mortgages, schooling and care. A regulator that requests another study preserves its freedom to decide while exposed communities continue to bear the hazard. A leader who keeps three successors “in consideration” may preserve bargaining power by making all three organise their lives around uncertainty.

So option value must be indexed to a bearer. Ask whose options increase, whose shrink, and whether one party’s reversibility is purchased with another party’s irreversible waiting. The same temporary contract may be a reversible probe for the employer and six months of unrecoverable career time for the worker. The same medical watch-and-wait plan may preserve clinical choices but consume the patient’s remaining time. A journal that records only the decision maker’s option set will systematically favour those already holding authority.

There is a second trap. Reversible actions can create path dependence before any formal commitment occurs. A pilot establishes infrastructure, trains users, changes expectations and creates a constituency for expansion. Cancelling it may remain legally possible while becoming politically expensive. The correct test is therefore not “Can we undo the transaction?” but “Can we restore the relevant relationships, capabilities and expectations at acceptable cost?” Reversibility is graded, multidimensional and time-dependent.

Restraint becomes wisdom only when it protects the affected field rather than the chooser’s comfort. A valid staged action names its exit conditions, limits who is exposed, prevents the pilot from silently becoming the default, and schedules a decision point. Without those constraints, “learning” can become a respectable name for transferring risk to people with less power.

What machines can calculate, and who can answer for the choice?

Modern models can help precisely where human finitude hurts. They can search a wider option space, generate counterfactuals, identify inconsistent assumptions, estimate sensitivity and maintain a decision record when attention is overloaded. They can also worsen the problem by compressing values into a proxy, presenting one ranking as inevitability, or flooding the decision maker with more analysis than the deadline permits.

The relevant contrast is not wise human versus unwise machine. Humans tunnel, rationalise and protect identity. Machines inherit objectives, data and interface choices that can make the tunnel look mathematical. A coupled system becomes wiser only when its division of cognitive work preserves the dimensions that either party would otherwise erase.

Responsibility needs continuity, not just causation

A model can be causally important without being the responsible authority. A loose bolt can cause a crash; a misleading chart can cause a vote; an optimisation system can cause a gate command. Responsibility asks further questions. Who could foresee the class of harm? Who controlled the objective, evidence and permission? Who had a duty to challenge the proposal? Who can answer to affected people and organise repair afterwards?

These questions reveal a second kind of continuity. The decision process needs an accountable thread that survives staff rotation, model replacement and changing explanations. Otherwise every participant can point to a transient component: the operator followed the model, the model followed the objective, the objective came from policy, and policy was inherited. The action is causally overdetermined and institutionally ownerless.

A decision receipt counters that drift. It binds the proposal to the evidence available at the cutoff, the named authority, the options rejected, the dissent recorded and the readback promised. It does not guarantee virtue or eliminate collective responsibility. It prevents responsibility from evaporating merely because computation was distributed. Answerability must persist at least as long as the consequences remain material.

This is where finite life contributes restraint. A person knows, however imperfectly, that time spent repairing one decision cannot be recovered for another purpose. An institution can approximate that burden only by designing durable obligations: compensation funds, review dates, named owners, preserved records and successors who inherit duties rather than only software. Machines can help maintain this continuity. They cannot create the mandate by announcing confidence.

A realistic configured-system example

For Sahana’s gate, a useful system would not display one recommendation and a confidence percentage. It would maintain a world-state record with sensor freshness, forecast versions, actuator health and the time at which each action ceases to be available. It would separate the objective “protect life” from proxies such as expected property loss. It would show which communities bear each branch, identify the eight-minute partial update, and test whether any reversible probe exists.

The model proposal might say: open now; wait eight minutes with crews pre-positioned; divert flow elsewhere; or accept controlled flooding in a different zone. Each option carries assumptions, scenario sensitivity and option loss. The authority boundary verifies that Sahana has the mandate, that consultation requirements are satisfied or validly suspended, and that the selected action is typed to the actual gate. The system records the reasons, cutoff, dissent and expected readback.

This does not guarantee wisdom. A narrow objective can still generate a beautifully auditable mistake. Yet the configuration makes several failures harder: hidden stale data, unauthorised context, unrecorded value trade-offs, silent delay and the fiction that a model “made” the decision.

Executable lab

stateDiagram-v2
  [*] --> Claim
  state "From outcome value to option loss" as Claim
  state "Four limits that should not be<br/>collapsed" as Test
  state "Scarcity creates a tunnel" as Revision
  Claim --> Test: expose mechanism
  Test --> Decision: evidence survives
  Test --> Revision: rival survives
  Revision --> Claim: narrow or rebuild
  Decision --> [*]

The wisdom decision journal

The journal below is not a moral calculator. It is a frame-expansion instrument. It turns nine hidden judgements into a visible decision receipt and chooses among four process moves: stage and learn, commit with a receipt, bounded deliberation, or stop and escalate. The logic is intentionally conservative about authority and explicitly compares learning value with delay harm.

How to apply the result

Stage and learn means a decision-changing probe appears more valuable than the present delay hazard. It permits only the smallest authorised reversible step. Commit with a receipt means delay now threatens more value than likely learning can recover. It permits action only after authority, dissent and recovery duties are recorded. Bounded deliberation means both sides are material; set a cutoff, assign the missing evidence and decide who will own the commitment. Stop and escalate means an irreversible action lacks clear authority or creates an accountability burden the current process cannot carry.

A low score is not reassurance. A decision may be outside the instrument because the affected people, rights or consequences are not represented. The journal is most useful before meetings, model runs and urgency have already narrowed the vocabulary of the choice.

function assessFinitude(input) {
  const n = value => Math.max(0, Math.min(4, value)) / 4;
  const R = n(input.irreversibility);
  const U = n(input.uncertainty);
  const L = n(input.learning);
  const D = n(input.delay);
  const S = n(input.scarcity);
  const A = n(input.affected);
  const C = n(input.continuity);
  const K = n(input.recovery);
  const H = n(input.authority);

  const optionValue = U * L * (0.55 * R + 0.45 * (1 - K));
  const delayHazard = D * (0.75 + 0.25 * S);
  const accountability = R * (A + C + (1 - H)) / 3;
  const tunnelRisk = S * (U + A + (1 - K)) / 3;

  let move = 'bounded-deliberation';
  if (R >= 0.5 && H < 0.5) move = 'stop-and-escalate';
  else if (optionValue > delayHazard + 0.12) move = 'stage-and-learn';
  else if (delayHazard > optionValue + 0.12) move = 'commit-with-receipt';

  return { move, optionValue, delayHazard, accountability, tunnelRisk };
}

Use computation to widen futures, not own them

The next time a consequential model returns a ranked answer, do not begin by asking whether its recommendation is intelligent. Ask what kind of world the recommendation assumes. Can errors be averaged across repetitions? Are losses fungible? Can the affected people be restored? Will waiting reveal information that changes the action, or will it merely consume the remaining option? Who retains authority, and who will still be present to answer after the model version has changed?

Use machines to expand the visible future, not to erase the fact that a particular future must be owned. Let them enumerate alternatives, stress assumptions, estimate sensitivity, watch deadlines and maintain receipts. Keep irreversible authority with a person or institution whose mandate, reasons and duties survive the action. This is not a claim that humans are inherently wiser. It is a design response to the current asymmetry between computation that can be rerun and lives that cannot.

For reversible, low-stake actions, speed and automated execution are often rational. For irreversible actions under material uncertainty, seek the smallest probe that can change the decision. When delay itself becomes irreversible, commit at a declared cutoff with dissent and recovery obligations intact. In every case, preserve the difference between a model proposal, a permitted action and a verified outcome.

Finite life contributes neither certainty nor moral purity. It contributes a structure: time runs out, attention narrows, some relationships and identities cannot be replaced, and consequences adhere to bearers rather than averages. Wisdom is the discipline of acting inside that structure without pretending it is a simulation.