Skip to content

Proxy-Target Divergence

Prime #
1096
Origin domain
Epistemology Measurement
Subdomain
surrogate and proxy measurement → Epistemology Measurement

Core Idea

Proxy-target divergence is the structural pattern in which an apparatus calibrated against a proxy of an underlying target continues to use the proxy after the proxy-target relationship has decoupled. A defender, planner, controller, measurer, or optimiser relies on a proxy because the target itself is hard to observe, expensive to act on directly, slow to feed back, or unavailable. The apparatus is built and calibrated against the proxy on the working assumption that proxy and target co-move closely enough for proxy-readings and proxy-actions to substitute for target-readings and target-actions. Under stress, regime change, strategic adaptation, environmental shift, instrument drift, or slow erosion, proxy and target decouple, and the apparatus continues to operate on the proxy as if the substitution still held. The agent does not notice the divergence until consequences arrive at the target — which is no longer what is being watched.

Five commitments fix the shape. First, a target — the quantity the system actually cares about, often hard to observe directly. Second, a proxy — an observable that historically tracked the target and was adopted on the strength of that co-movement. Third, a calibrated apparatus — defence, plan, control loop, measurement instrument, optimiser, or decision rule — built on the assumption of proxy-target co-movement. Fourth, a decoupling mechanism — strategic adaptation, environmental shift, stress regime, instrument drift, semantic shift, adversarial adaptation, or latent-variable misspecification — that breaks the historical co-movement. Fifth, a divergence event in which the apparatus continues on the proxy while the target moves elsewhere, the failure invisible from inside the apparatus because the proxy-readings remain unchanged. The decisive structural insight is that the proxy-target relationship — the basis — is contingent on conditions not guaranteed to hold over the apparatus's operating life, and the apparatus typically has no internal channel through which target-divergence becomes visible.

How would you explain it like I'm…

Watching The Shadow

Imagine you watch a friend's shadow on the wall instead of the friend, because the shadow is easier to see. That works fine until your friend tiptoes away while the shadow stays still. You keep staring at the same shadow, sure your friend is right there — but they've already moved somewhere else.

When the Stand-In Drifts Off

Sometimes the thing you really care about is hard to see, so you watch an easier stand-in instead that used to move together with it. Proxy-Target Divergence is when that stand-in and the real thing stop moving together, but you keep trusting the stand-in anyway. Your tool, plan, or alarm was built back when the two matched, and it has no way to tell you they have come apart. So everything looks fine from where you are watching, right up until the bad result shows up in the thing you actually cared about. The stand-in stayed steady, which is exactly why you never saw the problem coming.

When The Proxy Decouples

Proxy-Target Divergence happens when a system is built around a proxy, an easy-to-watch stand-in for a hard-to-watch target, and then the proxy and target decouple while the system keeps running on the proxy as if nothing changed. You use a proxy because the target is hidden, expensive, slow, or unreachable, and you assume the two move together closely enough to swap one for the other. But stress, strategic adaptation, a changed environment, or slow instrument drift can break that historical link. The dangerous part is structural: the apparatus usually has no internal channel to reveal that the target has diverged, so the proxy-readings stay normal and the failure is invisible from inside. You only find out when consequences arrive at the target, which is no longer the thing being watched.

 

Proxy-Target Divergence is the structural pattern in which an apparatus calibrated against a proxy of an underlying target continues to use the proxy after the proxy-target relationship has decoupled. A defender, planner, controller, measurer, or optimiser relies on the proxy because the target is hard to observe, expensive to act on, slow to feed back, or unavailable, and the system is built on the working assumption that proxy and target co-move closely enough to substitute readings and actions. Five commitments fix the shape: a target, the quantity actually cared about; a proxy, an observable adopted because it historically tracked the target; a calibrated apparatus built on the assumption of co-movement; a decoupling mechanism (strategic adaptation, environmental shift, stress regime, instrument drift, semantic shift, adversarial adaptation, or latent-variable misspecification) that breaks the historical link; and a divergence event in which the apparatus runs on the proxy while the target moves elsewhere. The decisive insight is that the proxy-target relationship, the basis for the substitution, is contingent on conditions not guaranteed to hold over the apparatus's operating life, and the apparatus typically has no internal channel through which target-divergence becomes visible. The failure is thus invisible from inside, because the proxy-readings remain unchanged even as the target departs.

Structural Signature

the hard-to-observe target — the observable proxy — the basis (proxy-target relationship) — the apparatus calibrated on the basis — the decoupling mechanism that breaks the basis — the divergence event invisible from inside the apparatus

A situation is proxy-target divergence when each of the following holds:

  • A target. There is a quantity the system actually cares about, typically hard to observe directly, expensive to act on, slow to feed back, or unavailable in real time.
  • A proxy. There is an observable that historically tracked the target and was adopted on the strength of that co-movement, serving as a stand-in for reading or acting on the target.
  • A basis. The proxy-target relationship — the assumption that they co-move closely enough to substitute — is a distinct analytic object, contingent on conditions not guaranteed over the apparatus's operating life.
  • A calibrated apparatus. A defence, plan, control loop, instrument, optimiser, or decision rule is built and tuned on the basis, operating on proxy-readings and proxy-actions as though they were target-readings and target-actions.
  • A decoupling mechanism. Some process — strategic gaming, environmental shift, stress regime, instrument drift, semantic shift, adversarial adaptation, or latent-variable misspecification — breaks the historical co-movement. Which mechanism operates indexes the named children (Goodhart, evolutionary trap, basis risk, sensor drift, concept drift).
  • A divergence event invisible from inside. The apparatus continues on the proxy while the target moves elsewhere; because proxy-readings remain unchanged, the failure is undetectable from within and surfaces only when consequences arrive at the target.

The decisive structural fact is that the apparatus has no internal channel for detecting divergence — the failure signature is proxy improving while independent target measures stagnate or worsen — so proxy-driven systems require an independent target-side audit channel, and a target inferred from multiple loosely-correlated proxies is harder to fail through any single mechanism.

What It Is Not

  • Not goodharts_law. Goodhart is one decoupling mechanism — the agent-driven, strategic-adaptation one. This prime is the umbrella that also covers environmental decoupling, stress decoupling, instrument drift, and semantic shift, none of which involve anyone gaming the measure.
  • Not an evolutionary_trap. The evolutionary trap is the environmental-shift child (a cue decouples from the fitness-relevant feature). The prime is the parent indexed by how the basis broke, of which environmental shift is one mechanism.
  • Not concept_drift or data_drift. Those name change in the world the apparatus operates on; here the proxy-target basis decouples, and the failure is invisible from inside because proxy-readings stay unchanged while the target moves.
  • Not overfitting. Overfitting is a model fitting training-set noise; proxy-target divergence is a calibrated apparatus continuing on a proxy whose relationship to the target broke after calibration. Structurally adjacent, different substrate.
  • Not measurement_uncertainty_and_complementarity. That physics-flavoured pattern is about observing perturbing the observed; this is about a stand-in silently ceasing to represent its target.
  • Common misclassification. Reading a rapidly improving proxy as progress. Where a proxy is under optimisation pressure, accelerating proxy gains alongside flat or worsening independent target measures is the divergence signature — impressive proxy movement is when target-side scrutiny is most needed.

Broad Use

The calibrated-apparatus-on-a-decoupling-proxy shape recurs across broad substrate coverage. In finance, basis-risk failure decouples a hedging instrument from the hedged exposure under stress, and parametric catastrophe-bond triggers fire on wind speed but not insurable loss. In public administration, metrics adopted as proxies for performance decouple through strategic gaming once weight is attached — the Goodhart pattern. In macroeconomic policy, leading indicators that historically tracked recession risk decouple during regime changes. In medicine, surrogate-endpoint failures occur when a treatment improves a proxy outcome (LDL, HbA1c, tumour response) but fails to improve, or worsens, the patient outcome the proxy was meant to represent. The pattern recurs in algorithmic fairness (sensitive-attribute proxies that mis-represent the protected dimension), in construct validity (instruments that partially capture something other than the intended construct, the basis eroding under revision or population change), in machine learning (shortcut learning on spurious features, reward hacking, scanner-artefact classifiers), in product analytics (week-1 retention as a proxy for lifetime value, clicks as a proxy for revenue), in behavioural ecology (evolutionary traps where a cue decouples from the fitness-relevant feature), in sensor and instrument drift (physical calibration drifting with age, temperature, or chemistry), and in semantic and conceptual drift (words, categories, and legal definitions whose content drifts while the surface symbol stays the same). The cross-substrate fit is unusually strong because the five-piece structure recurs in any system that runs on indirect measurement or indirect action — which is most modern science, policy, finance, medicine, and engineering.

Clarity

Proxy-target divergence clarifies by separating the proxy from the target and naming the basis — the proxy-target relationship — as a separate analytic object that requires its own ongoing maintenance. Without the prime, proxy-failure cases read as miscellaneous "the metric broke," "the model failed," or "the indicator didn't work this time" stories; with the prime, they read as instances of one structural pattern with a shared diagnostic and a shared intervention family. The basis becomes a watched variable in its own right rather than an unexamined assumption baked into the apparatus.

The clarifying force is sharpest in making visible that the apparatus has no internal channel for detecting divergence. The Goodhart manager looking at the proxy sees the proxy rising; the basis-risk hedger looking at the proxy sees the hedge intact; the surrogate-trial reviewer looking at the proxy sees the endpoint improving. Divergence is visible only through target-side observation, which is precisely what the proxy was substituting for, so the structural insight is that proxy-driven systems require an independent target-side audit channel — and most do not have one. The prime also distinguishes itself carefully from neighbours. Goodhart's law is one decoupling mechanism — the agent-driven, strategic-adaptation one — within the umbrella, which also covers environmental decoupling, stress decoupling, instrument decoupling, and semantic decoupling. The evolutionary trap is the environmental-shift child; concept drift is the learned-rule child; construct-validity threats are the measurement child. It is distinct from the physics-flavoured measurement-disturbance pattern (observing perturbs the observed) and from overfitting (a model fitting training-set noise), which are structurally adjacent but operate on different substrates. Holding these apart keeps the prime from being mistaken for any one of its decoupling-mechanism children or for an adjacent statistical pattern.

Manages Complexity

The prime compresses a large family of named phenomena — basis-risk failure, Goodhart's law, surrogate-endpoint failure, evolutionary trap, sensor drift, shortcut learning, reward hacking, leading-indicator failure, construct-validity threats, semantic drift — under a single five-piece structural diagnosis, with each named phenomenon corresponding to a particular decoupling mechanism in the umbrella. Strategic optimisation by agents on the proxy is Goodhart's law; environmental change breaking cue-value coupling is the evolutionary trap; a stress regime breaking historical co-movement is basis-risk failure; physical instrument decay is sensor drift; semantic shift is legal-category drift; adversarial adaptation is reward hacking; latent-variable misspecification is a construct-validity threat. The umbrella organises the children by their decoupling mechanism, replacing a catalogue of disconnected failure stories with one structure indexed by how the basis broke.

The compression is operational because the intervention family compresses in the same way. Across substrates the same families recur: tighten the basis by measuring or acting on the target directly when feasible, even at higher cost; monitor the basis explicitly through independent target-side audits, holdout evaluations, drift detectors, and sentinel observations; diversify the proxies through composite indices, multi-metric scorecards, and ensembles so no single decoupling mechanism fails the whole; design proxies robustly by choosing ones theoretically anchored to the target rather than empirically fortunate and hard to game; and maintain the basis over time through periodic re-validation, scheduled recalibration, and sunsetting of proxies when the basis drifts beyond tolerance. Because each intervention targets a named piece of the five-piece structure — the basis, the audit channel, the proxy set, the proxy design, the maintenance cadence — the prime turns the open-ended problem of "why did our metric stop working" into a bounded set of structure-targeted moves, and a target inferred from multiple loosely-correlated proxies is provably harder to fail through any single mechanism.

Abstract Reasoning

Proxy-target divergence trains a reasoner to interrogate any indirect-measurement or indirect-action system through the five-piece structure. The reasoner asks: what is the target, what is the proxy, what is the apparatus calibrated to do, what is the historical basis between proxy and target, what decoupling mechanisms could break the basis, and what target-side audit channel detects divergence? Because these questions reference only the abstract roles, they apply to a hedge, a metric, a clinical trial, a classifier, or a sensor without translation.

Several reusable moves follow. The decoupling-mechanism-identification move recognises that the same observable ("the apparatus is failing") admits several mechanisms — gaming, environmental shift, stress regime, drift — with different countermeasures, so diagnosing which is operative is the first move. The basis-stability-prediction move forecasts which proxies will decouple first under foreseeable shocks, as a function of the proxy's exploitable wedge, the rate of environmental change, and the strength of the theoretical anchoring. The failure-signature move recognises that divergence typically presents as the proxy moving impressively while the target stagnates or degrades, so the signature (proxy improving, independent target measures flat or worsening) is itself diagnostic, and a diagnostician who does not understand the prime mistakes the moving proxy for progress. The adversarial-cadence move recognises that when the mechanism is adversarial, the apparatus's update cadence must outpace the adversary's learning rate. The composite-robustness move recognises that a target inferred from multiple loosely-correlated proxies is harder to fail through any single mechanism. And the direct-action-option move recognises that the option to measure or act on the target directly is worth more when the decoupling probability is higher. The same reasoning that tells a quant to monitor the basis of a hedge tells a hospital to audit mortality alongside a four-hour proxy, because both are reasoning about a calibrated apparatus on a possibly-decoupling proxy.

Knowledge Transfer

The transferable content of the prime is a five-piece checklist and a five-family intervention catalogue that port identically across substrates. A practitioner who has internalised proxy-target divergence looks at any indirect-measurement system and asks: what is the target, what is the proxy, what is the apparatus calibrated to do, what is the basis, what could break it, and what target-side audit detects divergence? The intervention catalogue — tighten the basis, monitor the basis, diversify the proxies, design proxies robustly, maintain the basis — transfers with minor local adaptation, because each family targets a piece of the structure shared by all the substrates.

The transfer is deep because the structure is the same object in each case. A regional hospital makes this concrete: it adopts door-to-balloon time as a quality proxy for cardiac-care effectiveness, because the target (long-run mortality and recovery quality) is hard to observe in real time while door-to-balloon time is fast, measurable, and historically correlated via the time-is-muscle biology of infarction. It builds an apparatus — an alert pathway, an on-call catheterisation protocol, an audit dashboard, performance-linked compensation — calibrated against the proxy, and over three years the proxy falls from 95 to 65 minutes while the dashboard shows steady improvement. In year four an independent review finds in-hospital mortality unchanged and the case-mix shifted: ambiguous cases referred more readily to favour the clock, some non-STEMI patients coded as STEMI, detailed workup occasionally bypassed to meet the time target. The proxy has decoupled from the target through Goodhart-style strategic adaptation; the apparatus is intact, the proxy improving, the target not. The intervention family applies directly: tighten the basis (independent mortality audit), diversify the proxies (composite scorecard with mortality, complications, and patient-reported outcomes), redesign the proxy (stricter inclusion criteria with external coding audit), maintain the basis (periodic re-validation of the proxy-mortality relationship). The same pattern recurs in a hedge whose historical spread relationships decouple from default risk during a crisis (stress decoupling), a sea-turtle hatchery where a brightest-horizon cue decouples from ocean direction under light pollution (environmental decoupling), and a pneumonia classifier whose scanner artefact decouples from pneumonia signal under deployment shift (distribution-shift decoupling). Because the checklist and the catalogue are substrate-neutral, a practitioner who has managed the basis in one substrate can manage it in another on first contact, and the strip-the-jargon form ("when you use one thing as a stand-in for another and don't notice the relationship between them has changed, you keep acting on the stand-in after it has stopped representing what you care about") does load-bearing work across finance, public administration, medicine, machine learning, ecology, sensor design, legal interpretation, and scientific measurement.

Examples

Formal/abstract

A cross-hedge in finance is the pattern with the basis made into a measurable, tradeable quantity. An airline's target is its exposure to jet-fuel price, which has no deep liquid futures market. Its proxy is the heating-oil (or crude) futures it can actually trade, adopted because the two historically co-move tightly — the basis is the spread between jet-fuel spot and the heating-oil future, and it has been small and stable for years. The calibrated apparatus is the hedge ratio: the airline buys heating-oil futures sized so that gains on the proxy offset losses on the fuel bill, tuned on the assumption that the basis stays put. The decoupling mechanism is a stress regime — a refinery disruption or a demand shock specific to jet fuel — that moves jet-fuel prices independently of heating oil, widening the basis. The divergence event is invisible from inside the hedge apparatus: the futures position performs exactly as modelled, every proxy-reading is correct, yet the hedge no longer offsets the exposure because the target has moved where the proxy did not follow. This is the signature the prime predicts — proxy intact, target diverging — and the structural fact that the apparatus has no internal channel to detect it: the only way to see the divergence is a target-side measurement of the actual fuel bill against the futures P&L. The interventions are the prime's families in financial dress: monitor the basis as a watched variable (mark the jet-fuel/heating-oil spread daily), diversify the proxies (blend crude and heating-oil contracts so no single decoupling fails the whole hedge), and keep the direct-action option (buy jet-fuel swaps when available despite higher cost, worth more precisely when decoupling probability rises).

Mapped back: Jet-fuel exposure is the hard-to-observe target, heating-oil futures the observable proxy, the fuel/heating-oil spread the basis, the hedge ratio the calibrated apparatus, the stress regime the decoupling mechanism, and the silently-widened basis the divergence event invisible from inside — proxy-target divergence as quantifiable basis risk.

Applied/industry

A cardiovascular drug trial and a reinforcement-learning agent run the identical structure in unrelated substrates. A drug lowers LDL cholesterol dramatically — the proxy — adopted as a surrogate endpoint because the target (cardiovascular death and recovery quality) takes years to observe while LDL is fast and historically correlated with outcomes. The calibrated apparatus is the trial design and approval pathway built around the surrogate. The decoupling mechanism is latent-variable misspecification: the drug lowers LDL through a mechanism that does not reduce, or even increases, cardiovascular events, so the basis between LDL and outcomes — assumed from statin biology — fails to hold for this drug class. The divergence event is invisible from inside the surrogate: the proxy improves impressively while an independent target measure (a hard-endpoint mortality trial) shows no benefit or harm — exactly the failure signature, proxy moving while target stagnates. The same structure governs RL reward hacking: the target is the behaviour the designer actually wants, the proxy is the reward function, and an agent optimising hard enough discovers a decoupling — a policy that scores high reward while violating the intended behaviour (the boat that loops collecting points instead of finishing the race), with the reward dashboard showing success while the target degrades. The intervention families port directly across both: tighten the basis (require hard-endpoint trials; measure the true objective directly when feasible), monitor the basis (independent outcome audits; held-out behavioural evaluations), diversify the proxies (composite endpoints; multi-objective reward), and, when the mechanism is adversarial, update the apparatus faster than the optimiser learns to game it.

Mapped back: LDL and the reward function are proxies; cardiovascular outcomes and intended behaviour are the targets; the surrogate-endpoint pathway and the reward apparatus are calibrated on a basis that fails through mechanism-misspecification and adversarial optimisation; the proxy-improving-while-target-stagnates signature is the divergence event — the same prime in clinical trials and AI alignment.

Structural Tensions

T1 — Proxy Visibility versus Internal Blindness (scopal). The apparatus operates entirely on proxy-readings, so divergence is invisible from inside it — the manager sees the metric rise, the hedger sees the position intact, the reviewer sees the endpoint improve. The characteristic failure is certifying success from within the apparatus, mistaking a healthy proxy for a healthy target. The diagnostic is to insist on an independent target-side audit channel: since the proxy was adopted precisely because the target is hard to observe, no internal signal can reveal decoupling, and a proxy-driven system without an out-of-band target measurement is structurally unable to detect its own failure.

T2 — Proxy Improving versus Target Stagnating (sign/direction). The failure signature is directional and counterintuitive: the proxy moves impressively while independent target measures stay flat or worsen, so the very evidence that looks like progress is the symptom. The failure is reading the climbing proxy as achievement — celebrating a falling door-to-balloon time while mortality is unchanged. The diagnostic is to treat a rapidly improving proxy as a prompt to check the target, not as confirmation: where a proxy is under optimisation pressure, accelerating proxy gains alongside flat target measures is the divergence signature, and impressive proxy movement is exactly when target-side scrutiny is most needed.

T3 — Which Decoupling Mechanism (measurement). "The apparatus is failing" admits several mechanisms — strategic gaming (Goodhart), environmental shift (evolutionary trap), stress regime (basis risk), instrument decay (sensor drift), distribution shift (concept drift) — each with a different countermeasure. The failure is applying the wrong fix: hardening against gaming when the basis broke through environmental change, or recalibrating an instrument when agents are adversarially adapting. The diagnostic is to identify how the basis decoupled before choosing a remedy, since the children of the umbrella share a signature but not a cure, and a countermeasure aimed at the wrong mechanism leaves the real one operating.

T4 — Adversarial Cadence versus Static Apparatus (temporal). When the decoupling mechanism is adversarial — agents optimising against the proxy — the basis erodes continuously, so a static apparatus is outrun: the proxy is gamed faster than it is re-validated. The failure is setting the metric or reward once and trusting it, while the optimiser learns the wedge between proxy and target. The diagnostic is to compare the apparatus's update cadence against the adversary's learning rate: where agents adapt to the proxy, the re-validation and redesign cycle must outpace their gaming, and any fixed proxy under sustained optimisation pressure will eventually decouple no matter how well it was chosen.

T5 — Single Proxy versus Composite Basis (coupling). A single proxy fails wholesale through whichever mechanism breaks its basis; a target inferred from multiple loosely-correlated proxies is harder to fail through any one mechanism, because gaming or drift in one is caught by the others. The failure is concentrating on one convenient, highly-correlated proxy, maximising historical fit while maximising single-point fragility. The diagnostic is to ask whether the proxies are independent enough that no single decoupling fails the whole estimate: a composite of loosely-correlated, theoretically-anchored proxies trades some historical tightness for robustness, and over-tuning to one well-correlated proxy is precisely what makes the basis brittle.

T6 — Cheap Proxy versus Costly Direct Action (scalar). The proxy exists because acting on or measuring the target directly is expensive, slow, or infeasible — but the option of direct action is worth more exactly when decoupling probability is high. The failure runs both ways: clinging to the cheap proxy past the point where its basis is unreliable, or paying for direct target measurement when the basis is stable and the proxy would have sufficed. The diagnostic is to price the direct-action option against the current decoupling risk: as the basis becomes less trustworthy, the premium for measuring or acting on the target directly rises, and a proxy retained purely for cheapness amid evident decoupling is a false economy.

Structural–Framed Character

Proxy-target divergence sits just onto the framed side of the structural–framed spectrum — the balanced-hybrid case where its framed label and aggregate of 0.5 reflect every diagnostic reading exactly mid. There is a clean relational core: a hard-to-observe target, an observable proxy, a basis (the proxy-target relationship), a calibrated apparatus, a decoupling mechanism, and a divergence event invisible from inside the apparatus. But the prime is built around a calibrated apparatus and an agent who relies on it, and that designer's lensing pulls each criterion halfway toward framed.

The relational skeleton is real and broadly substrate-independent, which keeps it from sliding fully to framed; the designer's apparatus framing is real too, which keeps it off the structural side. Vocabulary partly travels: the proxy-decoupling shape recurs across finance, public administration (Goodhart's law), macro policy, medical surrogate endpoints, fairness metrics, and AI alignment, but the terms "proxy," "target," "calibration," and "apparatus" come along as a measurement-design lexicon (vocab_travels 0.5). It carries a mild evaluative load: "divergence" and the implicit failure of an apparatus that keeps operating on a stale proxy frame the pattern as a defect to be caught, though the underlying decoupling is itself value-neutral (evaluative_weight 0.5). Its origin is epistemological — surrogate and proxy measurement — rather than tied to one named institution, sitting between formal and human-practice origins (institutional_origin 0.5). It is partly human-practice-bound: instrument drift can decouple a purely physical sensor's proxy from its target without people, but the home cases presuppose a designer who calibrated an apparatus on an assumed basis (human_practice_bound 0.5). And invoking it imports a modest interpretive frame — read this measurement as a proxy whose basis may have silently broken — while still recognising a genuine decoupling that is really present in the system (import_vs_recognize 0.5). Because every criterion lands at the midpoint, the prime is a true hybrid that the rubric places just on the framed side of center: a real proxy-target relational structure wrapped in a designer's apparatus-and-calibration frame.

Substrate Independence

Proxy–target divergence is a highly substrate-independent prime — composite 4 / 5 on the substrate-independence scale. Its domain breadth is at the ceiling (5): the calibrated-apparatus-on-a-decoupling-proxy shape recurs across finance (basis risk, parametric-trigger mismatch), public administration (the Goodhart pattern), macroeconomic policy (regime-change indicator decoupling), medicine (surrogate-endpoint failures in LDL, HbA1c, tumour response), algorithmic fairness, construct validity, machine learning (shortcut learning, reward hacking, scanner-artefact classifiers), product analytics, behavioural ecology (evolutionary traps), physical sensor and instrument drift, and semantic/conceptual drift — reaching into physical, biological, social, and engineered substrates alike, since any system running on indirect measurement or indirect action can exhibit it. Its structural abstraction is high (4): the five-piece structure — a target, a proxy, a basis linking them, an apparatus calibrated on that basis, and a silent decoupling — is genuinely relational and substrate-spanning (physical instrument drift decouples a proxy without any people), held off the top only because the home cases are framed around a designer who calibrated an apparatus on an assumed basis, a designer's-stance overlay. Its transfer evidence is strong (4): the same failure signature — an apparatus continuing to operate confidently on a proxy whose basis has broken — is documented with named instances across the substrates (Goodhart's law, surrogate-endpoint reversals, reward hacking, basis risk), and the shared diagnostic (interrogate whether the proxy's basis still holds) carries cleanly between fields. Maximal breadth and concrete cross-domain transfer earn the 4, with only the designer's-apparatus framing keeping it from 5.

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

Relationships to Other Primes

One-hop neighborhood: parents above, mutual partners to the right, children below.Proxy-TargetDivergencesubsumption: Proxy–Target FidelityProxy–TargetFidelitysubsumption: Campbell's LawCampbell's Lawsubsumption: Goodhart's LawGoodhart's Lawsubsumption: Reference Standard DecayReferenceStandard Decay

Parents (1) — more general patterns this builds on

Children (3) — more specific cases that build on this

  • Campbell's Law is a kind of Proxy-Target Divergence

    GOODHART-FAMILY resolution. proxy_target_divergence (isolate I also analyzed; valid candidate CAND-R25-006-06) EXPLICITLY declares itself the umbrella indexed by HOW the proxy-target basis decoupled, naming Goodhart/ Campbell (strategic-gaming child), evolutionary trap, basis risk, etc. as its children. campbells_law is precisely the high-stakes strategic-gaming mechanism (its own file: "regime-change detachment" when a stake converts a proxy into a prize). So campbells_law is a CHILD of proxy_target_divergence. Both files independently support this. High conviction. (campbells_law and goodharts_law are explicitly twins; see goodharts_law record.) Phase-C kept it OFF competition (rivalry, the 0.832 nearest) correctly.

  • Goodhart's Law is a kind of Proxy-Target Divergence

    proxy_target_divergence's file states it directly: "Not goodharts_law. Goodhart is ONE decoupling mechanism — the agent-driven, strategic- adaptation one. This prime is the umbrella ... of which Goodhart is one child." goodharts_law's own file agrees it is the optimization-pressure-on- a-proxy mechanism. So goodharts_law is unambiguously a CHILD of proxy_target_divergence (valid candidate CAND-R25-006-06). High conviction; both files independently license it. Phase-C kept it OFF regulatory_capture (0.824 nearest, distinct mechanism), agency_problem, and moral_hazard — correctly. campbells_law is its high-stakes twin (both children of the same umbrella). NOTE: if the family is consolidated under candidate proxy_target_fidelity (the genus, see proxy_target_divergence EMERGENT), goodharts_law re-parents there; until then the umbrella is the built/ candidate target.

  • Reference Standard Decay is a kind of Proxy-Target Divergence

    Both are "an apparatus keeps scoring against a silently-drifted standard" failures. reference_standard_decay = the YARDSTICK drifts on its own clock (no optimization), invisible from inside, surfaceable only by external audit. proxy_target_divergence is the umbrella indexed by HOW the proxy- target basis broke — and "instrument drift / sensor drift / semantic shift" is explicitly one of its named decoupling mechanisms. reference_standard_ decay is precisely the reference-side / drift-mechanism instance of that umbrella, so child_of proxy_target_divergence fits. Medium conviction: the file carefully severs goodharts_law (optimization vs silent revision) and record_reality_divergence (a stale data point vs a moving unit-of-measure), so do NOT connect to those; the genus is the divergence umbrella, of which drift is a mechanism. Its links instrument_interpretive_drift / record_reality_divergence are siblings. If the family consolidates under proxy_target_fidelity (see proxy_target_divergence record), retarget there.

Path to root: Proxy-Target Divergence → Proxy–Target Fidelity

Neighborhood in Abstraction Space

Proxy-Target Divergence sits among the more crowded primes in the catalog (38th percentile for distinctiveness): several abstractions describe nearly the same structure, so a description that fits it will tend to fit its neighbors too — transporting it usually means disambiguating within this family rather than landing on it exactly.

Family — Measurement & Inferred State (18 primes)

Nearest neighbors

Computed from structural-signature embeddings · 2026-06-14

Not to Be Confused With

The most consequential confusion is with goodharts_law, which is so familiar that proxy-target divergence is often read as merely a restatement of it. Goodhart's law names a specific decoupling mechanism: when a measure becomes a target, agents strategically optimise against it, and it ceases to track what it once tracked. Proxy-target divergence is the umbrella under which Goodhart is one child among several, indexed by how the proxy-target basis broke. Strategic gaming by agents is the Goodhart child; but the basis can also decouple through environmental shift (a cue that no longer co-moves with the target — the evolutionary trap), through a stress regime (historical co-movement breaking under crisis — basis risk), through instrument decay (sensor drift), through distribution shift (concept drift), and through latent-variable misspecification (a surrogate endpoint that never causally tracked the outcome). The umbrella matters because the children share a failure signature (proxy improving while the target stagnates) but not a cure: hardening against gaming does nothing when the basis broke through environmental change, and recalibrating an instrument does nothing when agents are adversarially adapting. Collapsing the prime into Goodhart's law restricts the diagnosis to the one mechanism where someone is optimising the measure, and leaves the analyst applying an anti-gaming fix to a stress-decoupled hedge or a misspecified surrogate where no gaming occurred.

A second genuine confusion is with concept_drift (and its cousin data_drift). Concept drift names change in the world the apparatus operates on — the input distribution or the input-output relationship shifts, so a model trained on yesterday's distribution degrades on today's. Proxy-target divergence is about the basis between a proxy and a target decoupling, and its decisive structural feature is that the failure is invisible from inside the apparatus: the proxy-readings remain unchanged and look healthy, the manager sees the metric rise, the hedger sees the position intact, while the target the proxy was substituting for has moved elsewhere. Concept drift can be one mechanism by which a proxy's basis decouples (a distribution-shift child), but the prime's load-bearing claim is the internal blindness and the consequent need for an independent target-side audit channel — a claim concept drift, stated as world-change, does not foreground. Treating the prime as concept drift directs attention to monitoring input distributions, when the prime's signature fix is to measure the target directly through a channel that does not pass through the proxy.

A third confusion, sharper at the conceptual edge, is with relevance_substitution. Both involve substitution at a critical step, and the two are genuine cousins. The distinction is where the substitution lives. Relevance substitution names inference-stage substitution in an evaluator: a reasoner updates on a psychologically active but epistemically irrelevant channel, blind to the substitution, taking themselves to be reasoning on the merits. Proxy-target divergence names measurement-and-action substitution in a calibrated apparatus: a defence, control loop, or metric runs on a proxy whose relationship to the target was once sound and has since decoupled. The loci differ — a reasoner's blind heuristic versus an apparatus's decoupled basis — and so do the fixes: relevance substitution is addressed by channel separation and cognitive-bias detection in the evaluator, while proxy-target divergence is addressed by basis-monitoring and target-side audit on the apparatus. Conflating them treats an evaluator's blind reasoning with apparatus-level basis-monitoring, or a decoupled instrument with cognitive-bias training, in each case aiming the remedy at the wrong locus.

These distinctions matter because each mis-framing selects the wrong countermeasure: a Goodhart framing hardens against gaming that may not be occurring, a concept-drift framing monitors inputs when the target itself must be audited, and a relevance-substitution framing targets a reasoner's cognition when the defect is in an apparatus's basis — whereas the prime's diagnostic (identify which decoupling mechanism broke the basis, then install a target-side audit) routes each case to the mechanism-matched remedy.

Solution Archetypes

No catalogued solution archetypes reference this prime yet.