The Measurement Anchor

I. The credit that was not a tonne

A carbon credit is treated as a quantified claim about the territory: one credit equals one tonne of CO₂e reduced or avoided relative to a counterfactual baseline world. The unit is tradable. Once tradable, the proxy is under direct consequence — every actor in the system has reason to optimize the proxy.

Probst and colleagues, in a 2024 Nature Communications synthesis covering roughly a fifth of credit volume issued to date and almost a billion tons CO₂e, estimated that less than 16% of credits issued to the investigated projects constituted real emission reductions. The variation by project type was large: 11% for cookstoves, 16% for SF₆ destruction, 25% for avoided deforestation, 68% for HFC-23 abatement, no statistically significant reductions for Chinese wind power and US improved forest management in the investigated set. The mechanisms behind the gap include additionality failures, inflated baselines, leakage, weak usage measurement, methodology problems, and verifier-incentive misalignment.

This is not reducible to fraud. Some failures are deliberate; most are predictable structural responses to attaching tradable consequence to a counterfactual quantified claim without anchoring the proxy-territory relation institutionally. The case asks the four questions this essay names: Causal audit path — how exactly does the project cause reductions relative to the counterfactual? Validity envelope — where do baseline and leakage assumptions hold? Bounded proxy status — is the credit treated as a conditional quantified claim, or as atmospheric subtraction made solid? Independent verification — do auditors test gaming and favorable-assumption strategies, and what is their incentive structure?

A credit is not a measured physical tonne the way a scale measures mass. It is a counterfactual proposition about a baseline world. When the institutional discipline that maintains the link from the proposition to the territory fails, the credit drifts away from the tonne. The drift is not exotic. It is the structural pathology of consequential proxies.

II. What a measurement anchor is

A metric is a pointer from a measurable surface (the proxy) back to a causal surface (the territory) the institution wants to act on. The pointer is useful only as long as the proxy-territory relation is stable. Once the metric becomes consequential — once an actor’s reward, funding, status, legitimacy, or operational closure depends on the score — actors apply selection pressure to the proxy. The territory may also change under reactivity: rankings reshape institutions, targets reshape services. The structural point is that the proxy’s information content changes under pressure — the same number no longer means what it meant before consequence attached.

The anchor is not a demand to weaken measurement. It is a demand to preserve the measurement relation when measurement starts to govern.

A measurement anchor is the set of institutional conditions that keeps a consequential proxy connected to the territory it stands in for. It has four:

• Causal audit path — a specified causal chain from the proxy surface back to the underlying territory. The audit path is not “we measured X carefully”; it is “we can show the chain through which X is causally connected to Y, the thing we want.”
• Validity envelope and recalibration — the domain in which the proxy-territory relation has been validated, with periodic re-anchoring when the relation drifts under pressure or context change.
• Bounded proxy status — the institutional record names the metric as a proxy, states the causal surface it tracks, and limits the consequences that can be attached outside its validated scope. A school score cannot be used outside the population it was validated on. A hospital target should not become the sole basis for funding unless the institution has validated the proxy for that use and installed stronger verification. A carbon credit cannot be treated as atmospheric subtraction unless the claim class allows.
• Independent or adversarial verification — a verifier outside the measured actor’s reporting chain, with an explicit gaming hypothesis: the check assumes the actor is optimizing the proxy, not the territory, and designs the audit accordingly.

The conditions are necessary, not sufficient. Even well-anchored measurement can be defeated by determined gaming. They specify the minimum institutional spec any consequential proxy should pass before deployment.

Alfred Korzybski’s 1933 formulation — “the map is not the territory” — is honored prior art at the level of representation. The Measurement Anchor is one layer downstream: not whether the map is the territory (it isn’t), but what institutional discipline holds the map’s relation to the territory once the map starts governing. Korzybski warned against confusing maps with territory; this essay identifies the structural mechanism by which institutions are forced to do exactly that whenever a proxy becomes consequential.

III. Consequence as selection pressure

Selection pressure on a proxy is not a moral failing of the actors. It is the predictable structural response of any system where actors are evaluated against a measurable surface.

Charles Goodhart’s 1975 monetary-policy paper named the dynamic: any observed statistical regularity tends to collapse once pressure is placed on it for control purposes. Donald Campbell’s 1979 social-indicators paper made the parallel claim: quantitative indicators used for social decision-making become subject to corruption pressures and distort the processes they monitor. Marilyn Strathern’s 1997 paper “Improving Ratings: Audit in the British University System” supplied the popular paraphrase — when a measure becomes a target, it ceases to be a good measure — and a stronger claim alongside it: audit is not merely neutral checking; it reorganizes institutions around auditable surfaces while jeopardizing what it claims to monitor.

Theodore Porter’s Trust in Numbers (1995) supplies the steelman against treating quantification as a pure mistake: institutions turn to numbers because personal trust, professional discretion, and institutional legitimacy are politically unavailable. Numbers are a governance technology adopted under conditions of distrust. The Measurement Anchor does not argue against measurement; it argues that consequence-attached measurement requires institutional discipline to remain a working pointer rather than degenerating into a target.

Wendy Espeland and Michael Sauder’s “Rankings and Reactivity” (2007), studying law-school rankings, showed how public consequential measures recreate the social worlds they purport to describe: institutions reorganize around the metric, expectations shift, and the metric’s relation to the underlying object changes through the act of attaching consequence to it. The reactivity is the structural mechanism the anchor preserves the relation against.

The literature already names the phenomenon. What it does not always supply, in one place and in deployable form, is the institutional discipline that keeps the pointer working.

IV. The Goodhart family

The canonical taxonomy is Manheim and Garrabrant’s “Categorizing Variants of Goodhart’s Law” (2018), which distinguishes four mechanisms:

• Regressional Goodhart — proxy-target correlation breaks at extreme values. The correlation held at moderate intensity; extreme optimization selects metric-favorable noise rather than the underlying territory.
• Extremal Goodhart — the proxy is used in a regime where the proxy-target relation was never validated. The metric was validated in one regime; consequence pushes actors into a regime where its information content differs.
• Causal Goodhart — proxy and target share a common cause. Intervention on the proxy doesn’t propagate to the territory because the proxy is downstream of, not upstream of, the underlying causal surface.
• Adversarial Goodhart — actors actively game the proxy.

The variant taxonomy classifies failure mechanisms. The Measurement Anchor classifies maintenance failures. They overlap, but they do not map one-to-one.

Goodhart variants name how the proxy fails. Anchors name what the institution failed to maintain. The diagnostic move is to ask which anchor failed, not which variant the failure was — because the variant is the diagnosis, and the anchor is the repair obligation.

A short walk through specimens:

• Financial VaR models — a value-at-risk model validated for frequent, ordinary market-risk estimation operates accurately in its calibration regime. When pushed into systemic-tail regimes, the proxy-target relation does not hold. Jón Daníelsson’s “Blame the Models” criticized this exact failure: statistical risk models are useful for frequent small events but unreliable when applied to systemically important events. Extremal Goodhart; validity-envelope failure.
• School quality ratings — ratings often correlate with student outcomes through common causes (student demographics, family income, selection effects) rather than through school value-added. Angrist, Hull, Pathak, and Walters (2024) found that school performance ratings’ correlation with white enrollment shares in Denver and New York reflects selection bias rather than causal value-add. Intervention on the rating doesn’t propagate to school quality because the rating is downstream of the common cause. Causal Goodhart; audit-path failure.
• NHS target gaming — Bevan and Hood (2006) documented the canonical adversarial case under the English NHS “targets and terror” regime: ambulance response-time correction and reclassification below the eight-minute threshold, A&E gaming through extra staffing during measured periods and cancellation of operations to keep within waiting-time limits, trolley-wait gaming by treating trolleys as beds, and inappropriate waiting-list adjustments affecting nearly 6,000 records. The targets were validated metrics at the population level; under consequence, the institutional reporting chain optimized the score while the underlying service worsened or stayed unchanged. Adversarial Goodhart; independent-verification failure plus bounded-proxy-status failure.
• Academic citation indices — Fire and Guestrin (2019) document publication inflation, longer titles and reference lists, self-citation, h-index gaming, and reactive academic publishing. This is a mixed case: regressional component at extreme citation levels (where the score selects for field size and citation-network position rather than research quality), but also strong reactivity in Strathern’s sense and gaming in Campbell’s. It does not fit a single variant cleanly.

GDP is a different kind of case worth naming. Stiglitz, Sen, and Fitoussi’s 2009 Commission report argued that GDP, useful as a market-production indicator, is treated as a welfare, sustainability, and distributional indicator it was never validated to be. Bounded-proxy-status failure at scale: the metric is used outside its validated domain because no better-grounded number is institutionally available.

V. The four anchors in detail

Causal audit path. A deployed metric should specify the causal chain from the proxy surface to the underlying causal surface it stands in for. Without this, the institution cannot say what a change in the score proves about the thing it governs — what is the territory the academic citation count tracks? Donabedian’s 1966 structure-process-outcome framework for healthcare quality is the canonical example of a discipline that requires this explicitly: structural and process measures are valid only when their causal chain to outcomes is specified. The audit path is not statistical fit; it is the named causal mechanism.

Validity envelope and recalibration. The institution declares the domain in which the proxy-territory relation has been validated, and re-anchors periodically when the relation drifts. Drift is not always temporal. A metric can fail because consequence pushed actors into a region where the proxy was never validated. Without an envelope, extremal and regressional drift accumulate silently. Generalizability theory in psychometrics is the technical ancestor; the envelope concept translates it into institutional scope-of-use: the institution declares where the metric may authorize action, and what happens when use exceeds that scope.

Bounded proxy status. The institutional record names the metric as a proxy, states the causal surface it is meant to track, and limits the consequences that can be attached to it outside its validated scope. This is stronger than language hygiene. It carries institutional force: do not reify the score, do not use the proxy outside its validation envelope, do not let the proxy become the sole basis for funding, ranking, closure, or legitimacy. For a carbon credit, bounded proxy status means the buyer may record the credit as a conditional offset claim, not as a physical tonne removed, unless the project class has passed the stated baseline, leakage, and additionality tests. Messick’s 1989 unified validity theory treats score interpretation and use as part of validity, including consequential validity; the governance rule here is stricter: no consequence attaches outside the validated use.

Independent or adversarial verification. The metric is checked by a body outside the measured actor’s reporting chain, with an explicit gaming hypothesis. Self-reporting is the canonical failure mode. The adversarial frame matters: the verifier must assume the measured actor is optimizing the proxy, not the territory, and design checks accordingly. Reliability and replication, in the psychometric and audit traditions, ask whether scores are stable across raters and occasions. Adversarial verification asks something different — whether the actor under consequence can manufacture the score while severing it from the territory. Daniel Koretz’s Measuring Up (2008) is the canonical analysis of how high-stakes accountability under self-reporting produces score inflation and the illusion of progress.

The four conditions are validity-under-use written as an institutional checklist for carbon credits, ESG scores, AI evaluations, NHS targets, university rankings, KPIs, fiscal indicators, and other consequential proxies that do not always arrive packaged as psychometric tests. The psychometric tradition since Cronbach and Meehl’s 1955 “Construct Validity in Psychological Tests” already contains the technical vocabulary of validity, reliability, generalizability, calibration, and (after Messick) consequential validity. The work here is translation and packaging — putting that tradition into deployable form for institutions that have to attach consequence to proxies and want to know what holds the proxy-territory link.

VI. Wrong repairs

When boards, agencies, funders, or managers see a metric failing, they usually add metrics, granularity, transparency, or automation. The instinct is wrong when the missing condition is not metric volume:

• More metrics without causal audit path → more degrees of freedom for gaming, not better measurement. A new ESG sub-score inherits whatever audit-path failure broke the headline ESG score.
• More granular metrics without validity envelope → finer-grained drift outside the validated range. Breaking the school rating into ten subscales does not extend the population it was validated on; it just splits the same selection bias into ten pieces.
• More transparent metrics without bounded proxy status → audience and actor both treat the metric as the thing itself. Publishing the carbon-credit methodology does not stop buyers from booking the credit as a physical tonne when the institution never bounds the claim.
• More automated metrics without independent verification → the actor authoring the data is still the actor being measured. NHS trust dashboards updated nightly do not solve self-reporting; they speed it up.

Donella Meadows’s “Leverage Points” (1999) ranks numerical parameters as the lowest-leverage intervention in a system. Anchored measurement does not replace rule, goal, or feedback redesign. It only prevents the metric layer from severing the institution’s contact with the governed thing. Higher-leverage moves sit at other layers of the corpus and are not the subject here. The anchor is the floor under any institution that wants to use a consequential metric at all.

The diagnostic before any reform: ask which of the four anchors failed, not which metric to add or improve.

VII. Relation to the corpus

The Measurement Anchor sits at Stack Layer 4 (Measurement). Adjacent corpus essays at neighbouring layers:

• Mechanism Space — the philosophical parent: the metric is the semantic-space pointer; the territory is mechanism-space. The Anchor is the institutional discipline that preserves the relation under consequence.
• Response Vector — formal gaming is one of four response channels actors take when a metric becomes target. The Anchor is upstream: the measurement-side failure that activates the gaming channel.
• Constructive Diagnosis — the §VII wrong-repair grid contains the closest existing compressed version of this thesis. The Anchor is the dedicated expansion for the measurement layer.
• Bad Equilibria Are Not One Thing — “Payoff failure” is the diagnostic gate. The Anchor supplies the measurement-layer mechanism beneath it.
• The Severed Map — the discipline-level pattern: when the map stops updating from the territory. The Anchor is the metric-object specialization of the same failure.
• Cargo Cult Epistemology — ritualized verification decoupled from causal contact. Anchor failure on the verification side.
• Implementation Ledger — the verification field requires an anchored metric. The Anchor is upstream of any conditional-closure design.
• The Stack — Layer 4 anchor.

VIII. Close

A consequential proxy remains usable only while it is institutionally bounded by a declared validity envelope, a causal audit path, and independent adversarial verification. Outside those conditions, the institution must either recalibrate, reduce the consequence attached to the metric, or explicitly treat the metric as advisory.

A metric is a maintained surface of contact. The anchor keeps that surface from becoming the thing it was meant to observe.

Related:

• Mechanism Space — the metric is the semantic-space pointer; territory is mechanism-space.
• Response Vector — what actors do when a metric becomes target.
• Constructive Diagnosis — the wrong-repair discipline this essay specializes for the measurement layer.
• Bad Equilibria Are Not One Thing — “Payoff failure” gate; the Anchor supplies its measurement-layer mechanism.
• Implementation Ledger — conditional closure requires anchored verification.
• The Refusal to Compute — the Layer 7 sibling: how institutions avoid producing or admitting the constraining relation in the first place, before a proxy could even drift.
• The Severed Map — discipline-level map-detachment; the Anchor is the metric-object specialization.
• Cargo Cult Epistemology — ritualized verification decoupled from causal contact.
• The Stack — Layer 4 anchor.