Proxy–Target Fidelity¶
Core Idea¶
The relation between an observable stand-in and the unobservable thing it stands in for, posing one question: how faithfully does movement in the proxy reflect movement in the target? — a fidelity that is never perfect and always use- and regime-relative.
How would you explain it like I'm…
Does the Stand-In Match?
You can't see how warm someone really feels inside, so you look at a thermometer instead. A thermometer is a stand-in: a thing you CAN see that's supposed to tell you about a thing you CAN'T. Proxy-target fidelity just means asking, how good is that stand-in? When the room number goes up, does the person actually feel hotter, or is the stand-in fibbing?
How Good Is the Stand-In
Most things we truly care about, like how smart someone is or how healthy they are, are hard to see directly. So we pick a stand-in we can see, like a test score or a temperature reading, and use it in place of the real thing. Proxy-Target Fidelity is how faithfully that stand-in follows the real thing: high fidelity means reading the stand-in really is like reading the real thing, low fidelity means it can fool you. A stand-in is never perfect, because if it were a perfect copy it would just be the real thing. And here's the tricky part: a stand-in can be good enough for one job, like describing a whole crowd, yet much too sloppy for another job, like deciding about one single person.
Faithfulness of the Proxy
Proxy-target fidelity is the relationship between an observable stand-in (a proxy) and the unobservable or expensive thing it stands in for (the target), plus the one question that relationship raises: how faithfully does movement in the proxy reflect movement in the target? Almost nothing we actually care about — ability, health, code quality, customer satisfaction — is directly readable, so we substitute something cheap, fast, or available, like a test score, a biomarker, or a click-through rate. Fidelity is how faithful that substitution is, and it is crucial that fidelity is never perfect: the proxy is not the target, or there would be no need for it. It is also use-relative and regime-relative: a proxy faithful enough to describe a whole population may be far too lossy to optimize against one person, and a proxy faithful where it was tested may collapse outside that range. This is more than ordinary measurement error — a thermometer can read temperature perfectly and still be a terrible proxy for comfort — because the question isn't whether the instrument reads its own quantity accurately but whether that quantity, even read perfectly, tracks the different quantity you actually care about.
Proxy-target fidelity is the structural relation between an observable stand-in and the unobservable or expensive-to-observe thing it stands in for, together with the single question that relation poses: how faithfully does movement in the proxy reflect movement in the target? Because almost nothing of genuine interest — ability, health, welfare, code quality, model competence — is directly readable, an actor substitutes a proxy (a test score, a biomarker, GDP, a passing test suite, a click-through rate) chosen for being cheap, fast, or simply available where the target is not. Four commitments define it: a target that is latent, costly, or slow; a proxy put in its place; a standing-in relation, where someone has designated the proxy to represent the target for measurement, optimization, decision, or inference; and a fidelity — the degree of correlation across the regime that matters, on which the whole value of the substitution rides. Crucially the prime names the relation and its gradient: fidelity is never perfect, and it is use-relative and regime-relative, so a proxy faithful enough to describe a population may be far too lossy to optimize against one actor, and one validated in one regime may collapse outside it. Its signature distinguishes it from bare measurement — measurement asks whether an instrument reads its own quantity accurately, while proxy fidelity asks the prior question of whether that quantity, read perfectly, tracks the different quantity actually cared about — and it is the genus over a family of named failures: imperfect fidelity gives measurement error and partial validity; fidelity eroding under optimization pressure gives the Goodhart family; a medical proxy failing to carry a treatment's true effect gives the surrogate-endpoint problem; an organizational metric diverging from real performance gives KPI gaming and Campbell's-law effects.
Broad Use¶
- Psychometrics: A test score proxies for a latent ability; construct validity asks whether it tracks the construct rather than test-taking skill.
- Medicine: A biomarker proxies for a clinical outcome; the surrogate-validity literature reckons with proxies that moved while outcomes did not.
- Machine learning: A loss function, accuracy metric, or reward model proxies for the true objective; misspecification is a fidelity failure.
- Economics: GDP proxies for welfare, CPI for cost of living — and the critiques are fidelity arguments about how far the indicator tracks the good.
- Management: KPIs proxy for performance — test scores for school quality, citations for research impact.
- Finance: A credit rating proxies for default probability; a backtest proxies for live strategy performance.
Clarity¶
Separates is this number accurate? (a measurement question) from does this number track what I care about? (a fidelity question) — a proxy can ace the first while failing the second completely.
Manages Complexity¶
Names both halves of the bargain: a proxy buys tractability and spends fidelity, so the management move is to bound and monitor the fidelity rather than assume it.
Abstract Reasoning¶
Fidelity holds for a target, in a regime, under a use, so fidelity validated for description does not transfer to optimization, and under optimization pressure it tends to erode (the Goodhart move).
Knowledge Transfer¶
- Psychometrics → everywhere: The validation discipline — never trust a stand-in until related to external criteria — ports to biomarkers, benchmarks, and KPIs.
- Medicine → ML/policy: The surrogate-endpoint caution — proxy gains need not be target gains — warns against a model that benchmarks higher but deploys worse.
- ML → management/economics: The guardrail-metric pattern — pair every optimized proxy with an unoptimized target measure — transfers to detect erosion.
Example¶
A team optimizes a recommender against click-through rate as a stand-in for user satisfaction; CTR correlates with satisfaction descriptively but erodes under optimization as the system learns clickbait, so guardrail metrics (retention, complaints) are needed to catch the divergence.
Relationships to Other Primes¶
Foundational — no parent edges in the catalog.
Children (3) — more specific cases that build on this
- Construct Validity is a kind of Proxy–Target Fidelity — The file: construct_validity is the PSYCHOMETRIC specialization (proxy=instrument, target=latent construct) of the cross-domain proxy->target fidelity genus. Clean child.
- Cue Outcome Decoupling is a kind of Proxy–Target Fidelity — child of emergent proxy_target_fidelity
- Proxy-Target Divergence is a kind of Proxy–Target Fidelity — child of emergent proxy_target_fidelity
Not to Be Confused With¶
- Proxy–Target Fidelity is not Goodhart's Law because it is the static genus — the relation and its gradient — whereas Goodhart is the failure mode under optimization that erodes it.
- Proxy–Target Fidelity is not Proxy-Target Divergence because it is the relation whose strength can be high or low, whereas divergence is the realized low-fidelity event.
- Proxy–Target Fidelity is not Construct Validity because it is the cross-domain genus, whereas construct validity is the psychometric instance where the target is a latent construct and the proxy a measurement instrument.