Reference-Point Dependence

Version

v4 · 2026-08-30 · History

Prime #

1429

Origin domain

Decision Science

Subdomain

behavioral economics → Decision Science

Related primes

Comparison

Core Idea

Reference-point dependence is the evaluation regime in which an outcome takes its sign and weight from the gap between it and a reference state rather than from its position on an absolute scale. The identifying claim concerns a second variable that ordinary talk about outcomes does not carry: leave the outcome exactly where it is, move the reference alone, and the same state is reclassified — gain to loss, improvement to deterioration, adequate to failing — with nothing about the world having changed. ^[1]

Five elements must be present together before the label earns its keep. An outcome variable whose absolute level can be written down. A reference state. A difference operation applied to the two. A value or response function defined over that difference rather than over the level. And a genuine possibility of the reference moving while the outcome stays put. Strike the last and what remains is a choice of origin, a units convention that changes no prediction and no behaviour. Unequal treatment of the two sides — steeper below the reference than above it — is a frequent and consequential addition, but it is not constitutive: a system responding symmetrically to signed departures is reference-dependent in full. ^[2]

What follows is a measurement problem rather than a curiosity. Because the response is jointly determined by two quantities and only one of them is normally recorded, absolute outcome data underdetermine both prediction and diagnosis. A workforce whose pay rose and whose satisfaction fell, a market selling off on record earnings, a patient reporting worse pain on an unchanged dose: the level is known, the response is known, and the variable reconciling them is the one nobody wrote down. This prime asserts that the missing variable is real, movable, and often the only thing that moved.

Structural Signature

Absolute outcome held fixed + reference state relocated → the signed departure changes, and the evaluation or response tracks that departure rather than the level. ^[1]

Two counterfactuals sit inside the formula and both have to be run. The relocation test must bite: shifting the reference alone changes the response. The translation test must not: moving outcome and reference together by the same amount should leave the response untouched. Systems passing the first and failing the second are mixed, part difference-reader and part level-reader, and mixture is the ordinary case.

Recurring features:

• An outcome variable with a statable absolute level. Money, temperature, throughput, dose, vote share: something reportable without mentioning anything else, so that there is a fact for the reference to be set against.
• A reference state with a nameable source. Prior period, current endowment, an announced target, a published forecast, an aspiration, a claimed entitlement, a peer benchmark, or a running average of recent inputs. Naming the source is what makes the reference auditable rather than a fitted residual.
• A signed difference rather than a bare magnitude. The sign is categorical and load-bearing: it selects which branch of the response applies, and the branches usually differ in slope, in salience, or in what they license next.
• A response function defined over the difference. A value curve, a firing rate, a controller output, a satisfaction rating, an error term handed to an optimizer. The difference is what enters; the level enters only through it.
• An update rule with a time constant. References move, and how fast is a parameter rather than a detail. One that tracks instantly makes the system a pure differentiator registering only change; one that never moves makes it an absolute meter with an offset. Everything of interest lies between. ^[3]
• Independent settability. The reference must be movable without altering what the outcome delivers. If raising the target also raises the budget, two things moved and no attribution survives.
• Exclusion: curvature in the level alone. A response that saturates as an absolute quantity grows is level-dependent without being reference-dependent, and imitates reference effects over narrow ranges unless the relocation test is run.

What It Is Not

The prime does not claim that absolute levels stop mattering. Nearly every real system reads both, and the mixture is the interesting quantity. A person moved from starvation to sufficiency is not merely registering a positive departure, and a reactor whose coolant climbs toward a physical limit is not made safe by recalibrating a setpoint. The assertion is that a differencing channel exists, carries weight, and can dominate over ordinary ranges — not that the level has been abolished. Where a system reads differences only, that is a strong and separately checkable property, and it usually holds locally at best. ^[4]

It does not claim that the two sides are treated unequally. That asymmetry is a further commitment about the shape of the response on either side of zero, and it can be present or absent without changing whether the evaluation is referenced at all. Merging the two ideas hides the many systems — controllers, sensory channels, variance reports — that difference against a reference and respond evenhandedly.

It does not require the reference to be chosen, deliberate, or conscious. Usually nobody selects it: it accumulates from exposure, is inherited from last year's number, or is a physical property of an adaptive channel. Treating it as something an evaluator picks smuggles in an agent the structure does not need.

It does not deliver a verdict. Calling a reference-driven evaluation an error presupposes that the correct object of concern is the absolute level, which is a normative position rather than a finding. People plausibly do care about change as such — a cut in pay is a different event from never having had the higher pay — and an organism tuned to change is not malfunctioning. The structure stays silent on whether a case is a mistake, a defensible priority, or an exploitation. ^[5]

Nor is it satisfied by the mere presence of a baseline in a report. Budgets and control limits are everywhere, and most are arithmetic conveniences leaving decisions untouched. The pattern is present only when relocating that baseline, with performance unchanged, would change what someone concludes or does.

Broad Use

The pattern surfaces wherever a response is generated from a difference — which is to say wherever a channel of limited range must represent a quantity of wide range, and wherever an evaluator has a history.

Perception and sensory coding: receptors and neural channels encode departures from a running adaptation level rather than absolute intensity, which is why a room feels cold on entering and unremarkable ten minutes later at the same temperature. That adaptation level moves on a measurable time course. ^[3]

Regulation, control, and physiology: a thermostat, a PID loop, and an autonomic reflex all act on error — measured value minus setpoint — and are indifferent to the absolute reading except through it. Set points themselves drift with acclimatization, training, or chronic load, so a value once defended as normal becomes the deviation that triggers a response.

Finance and markets: prices are read against purchase price, prior close, consensus, or a high-water mark. A firm setting an absolute record while missing consensus is punished, and holding losers while selling winners sorts positions identical in every forward-looking respect.

Performance measurement and compensation: attainment against plan, variance against budget, quotas, and stretch goals are deliberately installed references, and the installed one decides who counts as a success at fixed output.

Negotiation: reservation points and competing accounts of entitlement are references, and concessions are scored as losses from them — which is why an opening number keeps working long after everyone knows it was theatre.

Policy and measurement regimes: emissions targets against a chosen base year, poverty lines relative to median income, school ratings against prior cohorts. The base year is a policy instrument; moving it redistributes credit and blame without moving a single tonne.

Machine learning and optimization: advantage estimation subtracts a state-value baseline from a return before the update is applied, and reward shaping redefines progress relative to a potential function. The subtraction can leave the optimum untouched while transforming the learning dynamics — reference dependence in a system with no experience of gain or loss whatsoever. ^[6]

Clarity

The confusion this prime dissolves is the belief that an unexplained response to a known outcome must be explained by something inside the responder. When output rises and morale falls, or a market drops on good news, the reflex is irrationality, ingratitude, or hidden preferences — all of which locate the discrepancy in the evaluator's disposition. The referenced reading is cheaper: the response is a well-behaved function of a difference, and the difference moved because its second term moved. ^[7]

It also splits two habitually merged questions. Did the situation change, or did the standard we read it against change? A fall in reported satisfaction after a strong year and one after a weak year look identical in the data and call for opposite responses.

It explains why improvements so often fail to register. Where the reference tracks the outcome, a permanent gain produces a transient signal, leaving whoever delivered it holding a higher level and an unimpressed audience. Managers read this as ingratitude and analysts as noise; it is the predicted behaviour of a differencing channel with a moving origin.

And it disarms the pretence of a neutral yardstick. Choosing a base year, a comparison period, a peer group, or a plan number is choosing a reference, and every report of a change has already made that choice. The claim to be presenting facts rather than a framing does not survive the observation that the same facts against a different origin support the opposite headline. ^[8]

Manages Complexity

What the abstraction lets a system stop tracking is the absolute level, and the saving is not rhetorical. A channel with a few dozen distinguishable output states cannot represent a quantity varying over many orders of magnitude, but it can represent local departure from a slowly moving origin at high resolution and re-point that origin as conditions change. Differencing also cancels whatever is common to signal and reference — ambient illumination, general inflation, a shock shared across every business unit — so much nuisance variation never has to be modelled because it never enters the difference.

For an analyst the corresponding saving is that comparison stops requiring the evaluators' baselines to be known or matched. Two evaluators with different histories can be set beside each other on their signed departures without anyone establishing where each one's origin sits, which is much of what makes cross-person, cross-site, and cross-period data usable — though differencing cancels a baseline offset, not a difference in how the two use the scale itself, so what it buys is comparability of changes rather than of levels. ^[9]

What cannot be dropped is the reference and its update rule. The compression is bought on the level side and carries a bookkeeping obligation: a departure recorded without its origin cannot be reinterpreted later, set against a departure from a different origin, or audited. A series of variances whose base kept moving is not a shorter description of the underlying series but a lossy one. Nor can a level guard be dropped, since a system reading only differences cannot notice that its origin has walked somewhere dangerous.

Abstract Reasoning

The prime licenses a diagnostic that runs on observable responses and needs no access to anyone's reasons.

Step one: separate the level from the response. Write the outcome's absolute value and the evaluation beside it, at two or more times or for two or more evaluators. If one level maps to different responses, or different levels to one response, a second variable is in play.

Step two: propose a reference and name its source. A prior period, a target, an endowment, a forecast, a peer set. The proposal must be independently checkable — a published number, a state the system demonstrably occupied — because an origin inferred only from the response it explains is unfalsifiable.

Step three: run the relocation test. Change the reference while holding the outcome fixed and state the predicted direction in advance. Republished targets, base-year revisions, a retuned setpoint, and an evaluator moved into a different benchmark group are naturally occurring versions.

Step four: run the translation test. Shift outcome and reference together by the same amount. A pure difference-reader is unmoved; any residual response is the level-sensitive component, and quantifying it is how the mixture gets estimated instead of assumed. ^[4]

Step five: apply a step change and watch the decay. Move the outcome to a new value and hold it there. A response that peaks and then relaxes toward neutral while the outcome stays put is the fingerprint of an adapting origin, and the relaxation time reads the update rule's time constant straight off the curve. A response that persists at its new level indicates a fixed origin — a different system, with different remedies.

Step six: eliminate the rivals. Curvature in the level alone, which imitates reference sensitivity over short ranges; belief updating, where what the outcome means changed rather than what it is measured from; comparison-set effects, where an ordering among present options moved rather than the valuation of one; instrument saturation; and fatigue.

Knowledge Transfer

What crosses substrates is the role skeleton and the tests built on it: an outcome, a reference, a signed difference, a response function over that difference, and an update rule with a time constant. Instantiate those in a retina, a thermostat, a bond trader, a quota-bearing salesperson, an advantage estimator, or a national emissions inventory, and the relocation, translation, and step-and-decay tests all run unmodified and mean the same thing in each. The countermeasures travel too, since disclosing an origin, freezing it, reporting several, or adding an absolute guard are operations on a measurement procedure and require no theory of what is being measured. ^[10]

What does not cross is the phenomenology the behavioural literature attaches to the pattern on its home ground. Felt loss, regret, entitlement, ownership, the sense of having had something taken: these belong to one class of bearer, and a differencing amplifier that satisfies every structural role exhibits none of them. Carrying them across turns a structural finding into a metaphor and licenses interventions aimed at feelings in a system that has none.

The normative reading travels least of all. In a control loop, responding to error rather than level is the design intent and its correctness is not in question. Only against a standard of absolute-level evaluation — native to welfare economics, and not obviously exportable — does the same structure become a bias.

Examples

Formal/abstract

Take a system with outcome x at time t, reference r, and a response defined only on the difference, v = x − r. Let the reference update toward the realized outcome at rate α = 0.5, so the next reference is r + α(x − r). Start at rest with x = 100 and r = 100, so v = 0.

At t = 0 the outcome jumps to 130 and stays there permanently. The response is +30. But the reference is now 115, so at t = 1 the identical 130 yields +15; then the reference is 122.5 and the response is +7.5, then +3.75, then +1.875. The outcome is permanently thirty units higher and the response has returned to zero. Summed over all time the entire permanent improvement is worth 60 units of signal — a finite total, extracted once, from a change that persists forever.

Now let the outcome fall at t = 3 to 115, still fifteen units above the level the system occupied for its whole prior history. The reference by then sits at 126.25, so the departure is −11.25: a loss, and one larger in magnitude than the +7.5 recorded a step earlier for a state that was strictly better. The same absolute value would have been a substantial gain at t = 0 and is a substantial loss at t = 3 — the relocation counterfactual arriving through the passage of time rather than an experimenter's hand. ^[11]

The treadmill result falls out of the update rule. Holding the response at a constant positive c requires the difference to stay at c, so the reference climbs by αc each period and the outcome must climb with it. A constant signal requires unbounded growth in the level.

Mapped back: x is the outcome variable with a statable absolute level; r is the reference; v is the response function over the signed difference; α is the update rule's time constant, and it does the work that makes the case non-trivial. The relocation test is satisfied by construction, since one outcome value maps to opposite signs at t = 0 and t = 3. The translation test is satisfied exactly — a pure difference-reader, the limiting case real systems only approximate.

Applied/industry

A regional sales organization pays bonus on attainment, the percentage of quota achieved, and sets each quota in January by taking the rep's prior-year actual and adding a growth factor.

A rep closes 4.0 million against a 4.0 million quota, then next year closes 4.4 against the same quota: 110 percent, full bonus, named at the sales conference. The committee sets the following quota from 4.4 plus ten percent, or 4.84 million. The rep closes 4.6. That is the best year of their career in absolute terms — 200 thousand more than the year they were celebrated for, 600 thousand more than the year they were merely on plan — and it records as 95 percent attainment, below threshold, and opens a performance-management conversation.

Reps who grasp the update rule stop maximizing the outcome and start managing the origin: holding late-December deals into January, declining to overachieve past the threshold, resisting territory expansion. Today's outcome is tomorrow's reference, and the coherent response to that costs the firm booked revenue.

A second limb concerns which origin is used at all. The same 4.6 million is a 4.5 percent improvement on prior year, a 5 percent shortfall against plan, and — if regional peers averaged 4.2 — a nine percent outperformance of the benchmark. Three references, three signs, one unchanged fact. Which appears on the dashboard is settled by whoever owns the dashboard, and that is not a decision about the rep.

Mapped back: booked revenue is the outcome variable, quota is the reference, attainment percentage is the signed difference expressed as a ratio, and the bonus schedule is the response function — one with a hard threshold, so the sign matters more than the magnitude. Prior-year-plus-growth is the update rule, with a time constant of one year and a coefficient above one, which is what lets the reference overtake the outcome. Independent settability holds, and is exactly the problem: the committee moves the origin without moving anything the rep delivers. ^[12]

Structural Tensions

T1 — The reference is latent and inferred from what it explains. It is rarely published and almost never measured directly, so in practice it gets reconstructed from the very response it is invoked to account for. That is a free parameter wearing the costume of a finding: any anomalous reaction can be accommodated by positing whatever origin makes the sign come out right. What keeps the construct empirical is independent evidence — a stated target, a documented prior state, an adaptation curve measured on its own terms — before the reference enters the explanation. Most applied uses skip that step.

T2 — Whoever sets the reference owns the verdict. Because evaluation is a function of the difference, the party controlling the origin controls the sign of the result without touching performance. Quota committees, index constructors, base-year negotiators, and expectation-managing executives all hold this lever, and it is more attractive than improving the outcome because it is cheaper and looks technical rather than political. It invites gaming from both ends: the setter shades the origin toward the verdict it wants, and the evaluated party withholds effort to keep the next one low.

T3 — Adaptation erases the gains it records. Where the reference tracks realized outcomes, permanent improvements yield only transient signal, and whoever delivered them is left with a higher level and an audience registering nothing. Welfare accounting becomes genuinely ambiguous, because an integral over the transient response and a comparison of absolute states can rank the same policy in opposite orders, and neither is obviously the right object to score. Escalation turns structural rather than pathological: a constant positive signal demands an outcome that keeps rising, which nothing does indefinitely.

T4 — Bias or legitimate preference. Calling reference dependence an error requires the premise that only absolute states should count, which is a normative commitment rather than a result. Against it stands the observation that a reduction is a different event from never having had the thing, that prior possession and entitlement are real features of a situation, and that change is what a bounded system can afford to encode. The same evidence supports a diagnosis of irrationality and a defence of a considered priority, and no structural test adjudicates.

T5 — Difference-reading is blind to where the level has drifted. A system responding only to departures cannot see slow common-mode drift, because the origin moves with it and the error stays near zero throughout. Creeping tolerance for noise, ratcheting exposure limits, degraded service nobody complains about, and set points following chronic load all have this shape, invisible from inside the loop. The remedy — an absolute alarm outside the differencing path — costs precisely the dynamic range that made differencing attractive, so the trade has no dominant setting.

T6 — Several references hold at once and nothing ranks them. One outcome normally stands in relation to prior period, plan, forecast, peer benchmark, entitlement, and aspiration at the same time. These routinely disagree in sign, and no principle within the structure says which governs. Arguments about whether performance was good are therefore arguments about origin selection conducted in the vocabulary of fact, and the selection is usually made after the outcome is known, exactly when motivated choice is easiest. Fixing the origin before the period opens is the only cheap defence.

Structural–Framed Character

Reference Point Dependence sits on the framed side of the structural–framed spectrum, labeled mixed-framed at an aggregate of 0.5, with every criterion reading half. The skeleton is small and exact: an outcome variable with a representable absolute state, a reference state, a signed comparison of the two, a value or response function over that difference, and an update rule governing when the reference moves and when it stays sticky. The identifying prediction: relocating the reference while holding the outcome fixed can change whether it counts as gain or loss.

Human-practice-bound does the most work at 0.5. The signature names an evaluator, and its reference states — expectation, current possession, prior state, aspiration, entitlement, framing — are largely the furniture of an agent's situation. What holds it to half is that the response need not be conscious: perception, control systems and adaptation-level models preserve the outcome–reference–signed-change roles with nothing deliberating.

Domain vocabulary travels partially at 0.5: gain, loss and value function carry a behavioral tint, though the signed-difference machinery is statable without them. Evaluative weight reads 0.5 because the difference is fed to a value function, making evaluation part of the identity, not a downstream use. Institutional origin sits at half: human choice and finance supply the paradigm cases, though no institution moves a reference. Import-vs-recognize is 0.5 — imported in finance, recognized in perception and control.

The grade means the prime exports, but not for free: name the reference state and its update rule. Without them what remains is baseline deviation, which records departure without requiring valuation.

Substrate Independence

Reference-Point Dependence is about as substrate-independent as a prime can be — composite 5 / 5 on the substrate-independence scale. Its core is a three-term relation with no medium inside it: an outcome whose absolute level can be written down, a reference state that can be relocated independently, and a response defined over the signed gap rather than over the level, so moving the reference alone reclassifies a state that never changed. Choice under risk, sensory adaptation, controllers reading error against a setpoint, target-based performance measurement, and prices judged against prior levels or announced forecasts all pass that relocation test. Nothing requires the comparison to be conscious, the spread runs from receptor to market, and the instances are measured rather than asserted — abstraction, breadth and evidence agree.

• Composite substrate independence — 5 / 5
• Domain breadth — 5 / 5
• Structural abstraction — 5 / 5
• Transfer evidence — 4 / 5

Relationships to Other Abstractions

Current abstraction Reference-Point Dependence Prime

Parents (1) — more general patterns this builds on

• Reference-Point Dependence presupposes Comparison Prime

  Reference-point dependence presupposes comparison because an outcome obtains its sign and value only through contrast with the selected baseline.

Children (5) — more specific cases that build on this

• Big-fish–little-pond effect Domain-specific is a kind of Reference-Point Dependence

  The proposed strict upward parent is prime:reference_point_dependence.

• Degree of frost Domain-specific is a kind of Reference-Point Dependence

  The proposed strict upward parent is prime:reference_point_dependence.

• Root (chord) Domain-specific is a kind of Reference-Point Dependence

  The proposed strict upward parent is prime:reference_point_dependence.

▸ Show 2 more

Hierarchy path (1) — routes to 1 parentless root

• Reference-Point Dependence → Comparison → Self Checking

Neighborhood in Abstraction Space

Reference-Point Dependence sits in a sparse region of abstraction space (73^rd percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely rather than landing on a neighbor.

Family — Drift, Decay & Record Fidelity (19 primes)

Nearest neighbors

• Ground Truth — 0.70
• Discrepancy-Driven Correction — 0.70
• Framing — 0.69
• Baseline Deviation — 0.69
• Conservation Event — 0.69

Computed from structural-signature embeddings · 2026-09-10

Not to Be Confused With

Reference-Point Dependence must first be separated from Comparison, its parent. Comparison places items in a shared frame along chosen dimensions and reads off a relation, carrying no commitment about what the relation is for. The child narrows it three ways: one side is an outcome and the other a reference state rather than a co-equal item, the relation is signed, and the result is fed to a response function. Comparing two candidates yields an ordering and no gains or losses.

It is not Baseline Deviation, the closest structural sibling and the one most often mistaken for it. Baseline deviation declares a reference, measures departure from it, and makes that departure a first-class fact — a control chart, a variance line, a residual. Same arithmetic, and it stops there. This prime adds valuation and movability: the departure enters a response function, and relocating the origin changes what the evaluator concludes or does. A process chart whose limits are recentred reports different residuals while nobody's assessment changes; a recentred quota changes who gets paid.

It is not Anchoring. An anchor is a value encountered first that drags a later estimate toward it; its defining features are temporal priority and insufficient adjustment. It works on a single estimate, needs no relevance to what is judged, and produces no gain-or-loss classification. A reference point need not come first, is typically highly relevant, and works by supplying a sign rather than a pull. An arbitrary number can anchor a valuation; it cannot make an outcome a loss.

It is not Loss Aversion, which is a claim about the shape of the response on either side of a reference already in place. Loss aversion says the downside branch is steeper; this prime says there is a branch structure at all and that the dividing point can move. The dependence is more basic and holds with equal slopes, whereas loss aversion cannot be stated without a reference to be asymmetric around. Conflating them hides every symmetric case: controllers, adaptive sensors, variance reporting.

It is not Context-Dependent Preference, the nearest neighbour on the choice-theoretic side, and the boundary is sharp. Context dependence holds two or more focal options fixed and varies something unchosen about the surrounding set or procedure, so that which option wins changes; its relatum is a set. This prime operates on a single outcome evaluated against a state, needs no menu, and moves the valuation of that one thing rather than an ordering. One salary against last year's is reference-point dependence; two offers whose ranking flips when a third appears is context dependence. They compose — a menu can install a reference — without being identical.

It is not Framing. Framing changes the presentation of a situation while holding the situation fixed, and much of its effect runs through salience and interpretation rather than any origin. It is often the instrument by which a reference gets selected — a discount forgone or a surcharge imposed nominates two different origins — but it also does work unrelated to signed departures, and references relocate with wording untouched, through the passage of time or a new published target.

It is not Adaptation or Habituation. Adaptation is the broad adjustment of a system to prevailing conditions; habituation is the decline in response to a repeated inconsequential stimulus, marked by stimulus specificity, recovery, and restoration by novelty. Both describe how a reference comes to move. But they are the update rule, not the evaluation regime: this prime is what makes a moved origin consequential, and it holds equally where the origin is fixed by decree and never adapts.

Finally, it is not Homeostasis or Rank-Dependent Value. Homeostasis holds a variable near a defended set point; it uses a signed error, but its subject is maintenance of a state rather than valuation of an outcome, and a fixed set point exhibits no reference movement. Rank-dependent value indexes an outcome to its holder's position within a distribution of holders, so a rank-preserving shift of everyone leaves the positional component alone. Here the relatum is a single state, usually the evaluator's own past, and a uniform shift of everyone can still register as a personal loss.

Solution Archetypes

No catalogued solution archetypes reference this prime yet.

References

[1] Kahneman, Daniel, and Amos Tversky. "Prospect Theory: An Analysis of Decision under Risk". Econometrica, 1979. Establishes that outcomes are evaluated as gains and losses against a movable reference point rather than as final states. registry ↩a ↩b

[2] Tversky, Amos, and Daniel Kahneman. "Loss Aversion in Riskless Choice: A Reference-Dependent Model". Quarterly Journal of Economics, 1991. Formalizes reference dependence and separates it from loss aversion and diminishing sensitivity as three distinct components of one model. registry ↩

[3] Helson, Harry. Adaptation-Level Theory: An Experimental and Systematic Approach to Behavior. Harper & Row, 1964. Treats the adaptation level as a pooled, time-varying reference against which stimuli are judged, so the reference itself moves with exposure history. registry ↩a ↩b

[4] Kőszegi, Botond, and Matthew Rabin. "A Model of Reference-Dependent Preferences". Quarterly Journal of Economics, 2006. Derives utility from a consumption-utility component defined on levels plus a gain-loss component defined on departures from a reference, making the level-versus-difference mixture an explicit modelled quantity. registry ↩a ↩b

[5] Kőszegi, Botond, and Matthew Rabin. "Mistakes in Choice-Based Welfare Analysis". American Economic Review, 2007. Argues that classifying a response as a mistake rests on assumptions about well-being that observed choice alone cannot supply. registry ↩

[6] Ng, Andrew Y., Daishi Harada, and Stuart Russell. "Policy Invariance Under Reward Transformations: Theory and Application to Reward Shaping". Proceedings of the Sixteenth International Conference on Machine Learning, 1999. Proves that potential-based shaping leaves the optimal policy unchanged while altering the learning dynamics. registry ↩

[7] Ross, Lee. "The Intuitive Psychologist and His Shortcomings: Distortions in the Attribution Process". Advances in Experimental Social Psychology, 1977. Documents the default of attributing an anomalous response to the responder's disposition rather than to situational structure. registry ↩

[8] Fisher, Irving. The Making of Index Numbers: A Study of Their Varieties, Tests, and Reliability. Houghton Mifflin, 1922. Treats the choice of base period and comparison formula as constitutive of any reported change rather than as neutral bookkeeping. registry ↩

[9] Ferrer-i-Carbonell, Ada, and Paul Frijters. "How Important is Methodology for the Estimates of the Determinants of Happiness?". The Economic Journal, 2004. Finds that differencing out person-specific fixed effects changes estimates substantially while treating the reported scale as cardinal rather than ordinal changes them little. registry ↩

[10] World Resources Institute and World Business Council for Sustainable Development. The Greenhouse Gas Protocol: A Corporate Accounting and Reporting Standard, Revised Edition. WRI/WBCSD, 2004. Requires a reporter to choose, justify and disclose a base year and to publish a base-year recalculation policy, making origin disclosure and origin freezing generic procedural requirements independent of what is being measured. registry ↩

[11] Brickman, Philip, and Donald T. Campbell. "Hedonic Relativism and Planning the Good Society". In Adaptation-Level Theory: A Symposium, ed. M. H. Appley, Academic Press, 1971. Introduces the hedonic treadmill, in which a reference that tracks realized outcomes converts a permanent improvement into a transient signal. registry ↩

[12] Weitzman, Martin L. "The 'Ratchet Principle' and Performance Incentives". Bell Journal of Economics, 1980. Models the ratchet effect in which the principal resets the next period's target upward from realized performance. registry ↩