Proxy-Target Divergence¶
Core Idea¶
An apparatus calibrated against a proxy of a hard-to-observe target keeps operating on the proxy after the proxy-target relationship has silently decoupled — a failure invisible from inside, because proxy-readings stay unchanged while the target moves elsewhere.
How would you explain it like I'm…
Watching The Shadow
When the Stand-In Drifts Off
When The Proxy Decouples
Broad Use¶
- Finance: Basis-risk failure decouples a hedge from the hedged exposure under stress.
- Public administration: Metrics adopted as performance proxies decouple through strategic gaming — the Goodhart pattern.
- Macroeconomic policy: Leading indicators that tracked recession risk decouple during regime changes.
- Medicine: A treatment improves a surrogate endpoint (LDL, tumour response) while failing to improve, or worsening, the patient outcome.
- Machine learning: Shortcut learning and reward hacking — a policy scores high reward while violating the intended behaviour.
- Behavioural ecology: Evolutionary traps where a cue decouples from the fitness-relevant feature.
- Sensors: Physical calibration drifting with age, temperature, or chemistry.
Clarity¶
Names the basis — the proxy-target relationship — as a watched variable, and makes visible that the apparatus has no internal channel to detect divergence, so an independent target-side audit is required.
Manages Complexity¶
Organizes basis risk, Goodhart, surrogate-endpoint failure, sensor drift, and reward hacking as one structure indexed by how the basis broke, with a shared intervention family.
Abstract Reasoning¶
The failure signature is directional and counterintuitive: the proxy moves impressively while independent target measures stagnate — so an accelerating proxy is exactly when target-side scrutiny is most needed.
Knowledge Transfer¶
- Finance ↔ medicine: "Monitor the basis" tells a quant to mark the hedge spread daily and a hospital to audit mortality alongside a four-hour proxy.
- Across substrates: "Diversify the proxies" — a composite of loosely-correlated proxies is provably harder to fail through any single mechanism.
- Across substrates: "Price the direct-action option" — measuring the target directly is worth more precisely as decoupling probability rises.
Example¶
A hospital drives door-to-balloon time from 95 to 65 minutes while the dashboard shows improvement, but a year-four review finds mortality unchanged and the case-mix gamed — the proxy decoupled from the target through Goodhart-style adaptation, invisible until a target-side audit.
Relationships to Other Primes¶
Parents (1) — more general patterns this builds on
- Proxy-Target Divergence is a kind of Proxy–Target Fidelity — child of emergent proxy_target_fidelity
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
Not to Be Confused With¶
- Proxy-Target Divergence is not Goodhart's Law because it is the umbrella over many decoupling mechanisms, whereas Goodhart is the one agent-driven, strategic-gaming mechanism.
- Proxy-Target Divergence is not Concept Drift because the proxy-target basis decouples (invisible from inside), whereas concept drift names change in the world the apparatus operates on.
- Proxy-Target Divergence is not Overfitting because it is a calibrated apparatus continuing on a broken basis, whereas overfitting is a model fitting training-set noise.