The Gauge

Abstract

Why do evaluation reforms all fail the same way? Goodhart explains why a measure dies; he does not explain why the next one dies the same. Capture comes first: a measure from above makes the value below undepositable. Then the pressure law, by which inherited bite sets the quantity of damage, which a gauge reform only displaces or converts. Then the remedy: loosen, by the threshold, by abundance, or by time; or damp when you cannot loosen. And the instrument, finally: new failures among safe actors, the loosening signal hardest to imitate. Two empirical cases, one prediction signed in advance, a formal appendix with its breaking conditions.

I. The tourniquet

A tourniquet squeezes an arm. Downstream, the hand turns blue: what lives under the constraint dies of what it no longer lets through. People take alarm. They reform. They move the tourniquet: two fingers higher, a different angle. The hand dies differently; it dies as much. That is migration.

So the band is widened, to “hold better,” to “read wider”: the strap covers more of the arm, and it is the whole arm that goes numb. Necrosis has been traded for ankylosis. That is conversion.

Only one move gives the blood back: loosening. And that move cannot be counterfeited, because it has a sign, the pins and needles in the limb, which nothing else produces. The return of life declares itself in disorder, tingling, clumsy moves, before declaring itself in strength. No one mistakes pins and needles for a worsening. Who thinks he is treating a blue hand by moving the tourniquet toward the elbow?

The gauge is the position of the tourniquet, which chooses the place of the wound. The bite, that survival pressure the tier above makes weigh on the gauged tier, is its tightness. It sets the size of the wound. The tier below is the limb. And new failures are the pins and needles (the hardest to imitate, so the most significant).

II. Capture comes from above

A big maille*, a university, an administration, a company, wants to see itself, stay accountable, measurable, alive. It gives itself a gauge embodied in indicators, evaluation cycles, dashboards, transparency requirements. At its own scale, the move looks healthy; it is even what gets prescribed against closure. But this gauge reads only part of what happens below (the publishable, the countable, the visible, the fast). And everything it does not read becomes, on the tier below, undepositable and illegible: intuition without a scenario, care, and so on. What does not count for the gauge can no longer make a trace for those who live below. When it durably makes the value it does not read undepositable, capture above produces dissolution below: the audit dissolves the craft, the indicator the team, transparency the trust.

* Maille (French): a stitch, a single loop of knitting or of chain mail, the smallest place where the thread passes and holds. The word names the local place where a move can enter, be received, and change what follows. See The Maille.

Capture is not a third death of the maille. The Maille knows only two, and the count holds. Take up its three grounds again: the rock, where nothing enters; the water, where nothing stays; the living ground, which retains just enough for the trace to take. One would be tempted to add a fourth surface, the mold, which keeps only what fits its template. But the mold is not a surface you put your foot on: it is the form of the tier above, the one that decides which prints the surface below will be allowed to keep. If the rock and the water are states of a maille, the mold is a relation between two mailles.

And everything is decided below. From the small maille’s point of view, capture is indistinguishable from water. The workshop, the care team, the classroom feel dissolution plain and bare: we do work that has meaning and it evaporates, every year starts from zero. What they live is the second death, without remainder. What they do not see is its cause: a gauge installed one notch above. So capture is not a state of the small maille, but a relation. The capturing maille can stay alive on its own tier while dissolving the mailles placed beneath it, often without knowing it, often believing it is improving them.

Hence a diagnostic rule, which reverses the usual order of questions. Before a wasting maille, the question is “can something still enter here?” If nothing enters, it is closure, and you reopen the crossing. Can what enters stay and become operative? If nothing holds, the real question comes: who sorts what remains, and on which tier? Because if the trace is impossible here, very often a gauge is deciding elsewhere what counts. The dissolution you suffer is someone else’s captured selection.

This reversal has a guardrail, because “it’s a threshold for the larger form” would otherwise become the alibi of every constraint. The criterion is double. A/ a legitimate threshold opens, on the tier it serves, a passage that can actually be crossed; B/ it leaves, on the tier it constrains, enough to deposit a trace. A closure that opens nothing higher up, or makes everything undepositable lower down, is not a threshold. It is a gauge that takes itself for a form.

III. The setup

Louise is finishing her night in geriatrics. Since midnight she has turned bodies heavy with sleep, redone two dressings, recognized by smell what no number was saying yet, held an 87-year-old hand, soft as silk, that would not let go. At six ten she sits down at the screen and makes her night legible: every act has its code, every code its drop-down menu. The held hand has none. Forty minutes later, her night exists for the tier above, minus what did the caring.

This scene contains the whole setup. What remains is to name the parts, each part with its counter.

The tiers. Numbered k, zero at the very bottom: the laboratory, the classroom, the care team. Tier k+1 gauges tier k.

Value. Everything that really settles on a tier, dimension by dimension: care, slow craft, the fast measurable, the publishable. Counter: the independent, low-stakes audit (the NAEP for schools, the share of work that reorients a field).

The gauge. A legible set of dimensions, and a scalar score computed on them, a number, not a vector. To command access to a resource, any gauge must finally produce an ordered decision (a score, a ranking, a verdict comparable to a threshold), and the model keeps its scalar form: the state score, the watch-time, the impact factor. Being read costs, per dimension read (documenting, formatting, complying of course), paid in capacity by the gauged actor. Counter: the share of care time spent coding the care.

The threshold. To receive its resource (a position, a grant), tier k must cross a threshold, set by tier k+1.

The bite. The quantity that commands the whole setup. The share of an actor’s capacity that the necessity of crossing his survival threshold commands, in other words how much the threshold weighs in every choice. It rises with the height of the threshold, with the scarcity of the resource, with the cadence of evaluation. At saturation, survival commands everything. Counter: the grant rejection rate; the frequency of cycles; the share of time under the gauge.

Damage, in two forms, which are not summed, for want of an exchange rate between them. Exported dissolution: the value the gauge does not read and which, not being paid by survival, declines until it becomes undepositable below. Counter: the gap between the high-stakes test and the low-stakes audit, up to four to one. Imported closure: the capacity that legibility itself consumes. Counter: defensive documentation, reporting hours.

As the bite rises, the actor on tier k works less directly at value and more at the score his survival depends on. What the score does not pay then becomes what the pressure risks dropping.

IV. The pressure law

Goodhart explains why a targeted measure gets corrupted. Once it becomes a target, it is gamed. Almost everywhere the same lesson was drawn: we need better measures, and twenty years of reforms have reformed the direction of the gauge, producing a profusion of charters and other so-called responsible metrics. The law here stands one notch behind, and explains why the replacement corrupts in its turn if the pressure remains.

The gauge chooses the place of the wound; the pressure sets its size.

Pass-through proposition. In a chain of tiers subject to survival thresholds, inherited pressure determines the quantity of damage exported downward, in both of its two forms. Under the stated conditions and in steady state, at constant pressure, changing the gauge displaces or converts the damage without durably reducing it. At saturating bite, the wound is total outside the legible, wherever the legible is. Only a drop in pressure reduces the damage. And pressure has a time. A lengthening horizon, a tolerance for early failure are drops in bite per unit of time.

This proposition says neither that all gauges are equal, nor that better aim never reduces damage. It says that with pressure maintained, under the stated conditions and once the regime is installed, improving the gauge displaces, converts, or delays the damage more surely than it abolishes it. A better measure can buy time; only a drop in bite durably changes the regime.

The law derives in three steps (the tier, the direction, the chain), detailed in the appendix. The essential holds here, in two lines. Under pressure, the actor crosses his threshold at least effort, and at saturation nothing remains outside the legible. The wound is total outside the gauge, wherever the gauge is. Better aim first makes the defect fall; then the read dimension bends toward the score-without-value (Koretz’s inflation: the score climbs while performance does not move) and the defect climbs back to the level the pressure pins. Goodhart is not an anomaly of the law; he is the mechanism of its regime. And pressure is inherited: each tier’s threshold is set by the bite of the tier above, cascading down from the scarcity at the head. The chosen directions change nothing there, and the damage at the very bottom durably declines only if the head pressure declines, or if a tier passes on less than it receives.

The conditions carrying this are named: survival crushes vocation, legible effort does not spill over onto the illegible, costs are comparable, the regime is stationary. The appendix treats each lifting. A law that does not say where it breaks is a dogma; this one breaks at named places.

V. The two displacements of damage

At constant pressure, damage moves in two ways. Migration: change the direction of the gauge at equal rank; the damage relodges from one dimension to another. Conversion: read more dimensions; exported dissolution is exchanged for imported closure, along one same pressure. One death for the other. The three gauge remedies every reform proposes line up there.

Measure what was being missed. Finally aim at the high-value dimensions. The fall is real, and it has impact; then the read dimension becomes the target and empties as it is gauged. Koretz measures it: real performance fallen back half a grade level while the score climbed. Better aim, at best, delays the migration.

Measure more. Each dimension read takes one out of the shadow, and pays for it in closure. Compliance thickens, surveillance grows heavier to the point of suffocation. That is conversion; the open-plan office is its pure case. Everything was to be read, and most people sealed themselves.

Measure at random. The unpredictable audit. Impossible to game the gauge; you would think you had the parry. The law predicts something else. Randomizing spreads the pressure over all dimensions without reducing its sum. A conversion in expectation, where every move acts as if it were about to be read, hence defensive documentation and compliance anxiety.

Three failures of different shapes. None touches the pressure, none durably reduces the damage. They only redistribute. The whack-a-mole game at the fair, where what you hit with the mallet is a head and a spring.

VI. Loosen, or damp

Within the law’s domain, the only way to durably reduce total damage is to reduce the bite. Changing the gauge can temporarily modify its distribution, or limit some of its forms, but not remove the pressure that reproduces it. Three main levers reduce it directly. They are the threshold, scarcity, and time.

Time is the main door to real loosening. Lengthen the evaluation horizon, tolerate early failure. The long window strictly contains the short one, so it weighs less per period. It is a drop in bite per unit of time, and it is what houses the HHMI properly: no threshold abolished, no abundance (evaluation remains, entry stays ferocious), but a temporal slackening, hence partial, hence a partial fall in damage.

The threshold: the tier stops being gauged, de-gauging. Rare, costly, and the only move critique ordinarily demands, for want of seeing the other two. Scarcity: the resource stops being rationed, and the bite goes out. Abundance is the end of pressure, outside the law’s domain.

There remains the reader in the middle of the chain, who can loosen neither the threshold he suffers nor the scarcity at the head. Going up one notch serves to localize the bite; and when you cannot loosen it (you almost never can), you can decline to pass it on whole. Loosening reduces the pressure at its source; damping reduces the pressure passed to the next tier. Every tier is a reducer or an amplifier of pressure: what it demands of the tier below is not what it suffers from the tier above; it is what it chooses to pass along. Absorb variance for the tier below; lengthen the local horizon; fund the illegible out of your own margin; take some of the threshold onto yourself. That is what a good boss does, a good dissertation advisor, a good parent, and the formalism gives it a name: the damper, a wedge between the pressure received and the pressure passed on, paid out of the tier’s own slack. It does not destroy the damage but takes it on, and converts it into a cost borne by whoever chose to bear it. Its guardrail is The Hold’s: the tier that damps beyond its slack no longer protects anything; it collapses. Appendix C gives it its formal place: it falls out of the cascade instead of being added to it.

There remains the common-sense objection: stop measuring, and no one does a damn thing anymore. It confuses two moves. Loosening is not de-gauging. At the HHMI, evaluation remains; only the horizon lengthens, and the funded fail more while succeeding bigger. The lazy man, for his part, produces no new failure. A failure costs, can be verified, steps outside the repertoire, and idleness pays none of those prices. The true part of the objection has its place in the law. Slack has its bell curve, too little kills exploration, too much feeds complacency, and the question is never measure-or-no-measure, but which side of the bell you stand on. Many contemporary institutions (high rejection rates, shortened cycles, reporting that eats the time of care) appear to stand on the side where loosening pays.

VII. The failure test

How do you know whether a loosening has taken place? Everything in a reform can be imitated. The charters, the committees, the barometers: form. The new indicators: another gauge direction. And even the successes, which selection supplies: picking stars yields successes, never failures. Whoever judges a reform by its successes judges a signal that several indistinguishable causes produce.

Show your new failures, and show who makes them.

A new failure has three properties, all verifiable: it costs (real time, budget, reputation); it can be verified (an auditable administrative trace, not an anecdote); it is new (outside the already profitable repertoire of the one who tried it). Under bite, it is strictly dominated: no gauge optimizer manufactures one. Taken in isolation, though, a failure proves nothing; what discriminates is a joint modification of the distributions, relative to a counterfactual. Because the nearly dead fail too. Whoever can no longer cross his threshold by the safe routes plays the desperate bet, even at negative expectation. The two regimes separate at two marks. The position: slack failures come from safe actors, far from the threshold; distress failures, from the nearly dead. The other tail: exploration raises the big successes along with the failures; distress raises the failures alone. Pins and needles are thus the sign of loosening that nothing else produces. They are made of failures the way a tingle is made of pinpricks. And their false twin is the convulsion, which comes from the limb that has no blood left and that no strength accompanies.

Hence the complete signature, on two coordinates (who fails, and what the other tail does):

The last line cuts the contemporary cult of failure. A failure that costs nothing, cannot be verified, or stays inside the already profitable repertoire has become display. Goodhart applied to failure itself. So the test cuts both ways. Your successes prove nothing; neither do your failures, if they are papier-mâché. A failure is not subsidized; it is authorized. And the complete empirical protocol (floors, positions, control groups) is developed in the operating note The Failure Test.

VIII. In practice, does it really work like that?

Migration in its pure state: the operating room. New York State’s cardiac surgery report cards: once operative mortality was published, surgeons selected the less severe patients, the score improved, and the care of heavy cases degraded; Dranove and his coauthors established it. The value the gauge does not read, operating on the hard cases, became undepositable. Migration in its chemically pure state, with lives at the end of it.

Temporal loosening: the HHMI. The program funds people instead of projects (long horizon and no deliverables fixed in advance). The temporal face of the bite, lowered. The NIH, short cycles and merciless renewals, keeps its guard high. Among matched researchers, HHMI grantees produce more high-impact work, explore newer directions, and fail more (quantity of work below their own prior floor). The two tails, together, among actors far from the threshold: the signature of slack. Selection? The controls are matched on prior accomplishment, and above all selection does not explain the failures: anointing a star does not make it fail more; only risk does. The HHMI is not zero constraint; it is precisely a partial slackening, and a partial fall, which the law, graduated, predicts.

The open prediction: the Netherlands. The Recognition & Rewards reform (2019) changed the direction of the gauge, from the impact factor toward societal impact, open science, teamwork, without touching the head pressure: international competition, rankings, the scarcity of positions. The law predicts migration: capture relodges under the new dimensions. What is observed to date is compatible and settles nothing; the window is too short to separate a migration from a slow healing. What will settle it is the observable the test designates, and it must be preregistered: the failure register. Count, per researcher, before and after the reform, the risky directions opened then abandoned, the publications below their own prior floor. The prediction is sharp: no new failures among the safe. Let them appear, with successes rising, and the law is wrong somewhere.

IX. Limits and refutation

The law says where it breaks, in its very statement. Where vocation holds a floor against survival (the nurse, the craftsman: betray rather than die is not their order), pass-through is partial; the law bounds, it no longer measures. Where legible effort spills over onto the illegible (teaching to the test teaches a little), pass-through is bounded, and choosing gauges with strong spillover is a real lever, though a bounded one: it does not touch the pressure. The time of impact, before the regime: better aim buys time. And abundance, which is not its favorable case but its exit.

The test has its own. It does not survive its own bureaucratization: if “counting failures” becomes a target, cheap failures get manufactured by relabeling. The parry is in the definition, real cost and verification, and in the authorization rule. It requires a counterfactual: an external shock makes everyone fail, and the before-after comparison must hold a control group, or it holds nothing. And it detects loosening, not wisdom: a slackened pressure can be badly used; the test says whether the blood comes back, it does not say what the hand will do with it.

And there remains one quantity I do not know how to measure directly: the bite itself. The rejection rate, the cadence, the share of time under the gauge are only proxies for it, and each can be gamed, like everything that becomes a target. A theory of pressure whose central variable is observable only by its shadows stays exposed there; I do not know how to close that loop.

What would prove me wrong is relatively simple. A real loosening, new failures in evidence, that would not durably raise uptake: the corollary would be false. A gauge substitution at constant pressure that would durably reduce the damage: the law would be false. New failures among safe actors, with successes rising, after the Dutch reform: then the foundation of the test would be wrong somewhere. The Dutch failure register can be computed on open data; the complete protocol (personal floor, novelty condition, control groups) is in the operating note. The three refutations are there for the seeking.

X. Lineages

Three lineages carry the main debts. Kerr, Campbell, Goodhart (in the size Strathern gave him), and Holmström-Milgrom work at one tier: the corruption of measures, multitask distortion, the contract that weakens when what matters is illegible. Cyert and March, Nohria and Gulati, Manso hold the link from slack to exploration, down to the irony that the HHMI tests Manso’s theorem, he being a coauthor of the paper. Spence and Zahavi hold the costly signal; Jensen and Meckling, the resurrection gamble, which Brown, Harlow, and Starks carried to measurement. The hierarchies literature, from Williamson to Mookherjee, computes the dilution of incentives along the tiers; the invariance of downstream damage under rationing, I have not found explicitly assembled in that literature. Scott (legibility that destroys what it simplifies) and March (the coupling of the code and individuals) hold verticality itself.

What the essay adds: the cascade, the conversion between the two damages, identification by the failure tails, and the ontological register: undepositable, not under-incentivized. What the gauge does not read loses a place to make a trace.

Conclusion

Take up the tourniquet again. For a long time people believed the wound came from the position of the band, and so, logically, the band was moved, reform after reform. The wound changed place; it did not change size. Goodhart had said why each measure dies; what remained was to say why the next one would die the same: the gauge chooses the place of the wound, the pressure sets its size. Only one move gives the blood back: loosening, by the threshold, by abundance, by time. And for whoever cannot loosen, damping: passing on less pressure than you receive, out of your own slack. And loosening has a signal particularly hard to imitate: new failures, costly and verifiable, when they appear among safe actors at the same time as the big successes climb.

The blood that comes back tingles.

Formal appendix

The formalism that follows establishes the result in a deliberately conditioned case. The hypotheses are laid down in A, and their breakings treated in E.

A. The model

An actor, on tier k. There are n dimensions of value, i = 1…n. The actor allocates an effort e = (e₁, …,eₙ), eᵢ ≥ 0, under capacity Σᵢ eᵢ ≤ E. Value produced: Vᵢ = fᵢ(eᵢ), increasing concave. The gauge: a legible set L ⊆ {1,…,n}, of rank r = |L|, and a scalar score s(V) = Σ_{i∈L} wᵢVᵢ, wᵢ > 0; a legibility cost κ ≥ 0 per dimension read, paid in capacity: useful capacity is E − κr. Survival: the actor receives his resource only if s(V) ≥ τ, a threshold set by tier k+1. Vocation: a utility u(V), increasing; the vocational allocation e^v maximizes u under capacity alone, and serves as reference. The bite: the shadow price λ, multiplier of the constraint s(V) ≥ τ at the optimum; or the commanded share σ = E_τ/(E − κr) ∈ [0,1]; saturation: σ = 1. The bite rises with τ, with the scarcity of the resource, with the cadence of evaluation.

Four hypotheses, laid down as conditions. A1: survival crushes vocation. A2: no spillover; effort on one dimension produces nothing on another. A3: comparability of the fᵢ (isotropy is only a condition for comparing wound sizes, not a foundation). A4: stationary regime; equilibria are compared. Each lifting is treated in E. Damage, in units of effort: D_diss = Σ_{i∉L} (e^v_i − e*_i), the effort withdrawn from the illegible dimensions relative to vocation; D_clos = κr to first order, the capacity legibility consumes. The two are not summed.

B. Proposition 1: pass-through

(i) Monotonicity: D_diss is increasing in σ. By A1, the actor solves min Σᵢeᵢ subject to s(f(e)) ≥ τ; the envelope gives dE_τ/dτ = λ > 0. At σ = 1, e*ᵢ = 0 for all i ∉ L: the wound is total outside the legible. (ii) Directional invariance at saturation: at σ = 1, D_diss = Σ_{i∉L} e^v_i; changing L at given rank moves the wound without changing its size as soon as A3 makes the sizes comparable; under symmetric A3, D_diss = (1 − r/n) · E, independent of L. Without A3, the equality of magnitudes falls; the structural point remains: the wound is total outside the legible, wherever the legible is. (iii) The two displacements: at constant σ, changing L at equal rank is migration; increasing r: ∂D_diss/∂r = −E/n < 0 and ∂D_clos/∂r = κ > 0, is conversion, exported dissolution exchanged for imported closure along an iso-bite. (iv) Loosening: in the one-tier model, D_diss strictly decreases if and only if σ decreases, by three routes and three only: the threshold (τ falls), scarcity (λ → 0), and time. The last is an extension of the static model, which by itself establishes only migration and conversion: short window s_t ≥ τ each period versus long window Σ_{t≤T} s_t ≥ Tτ; the feasible set of the second strictly contains that of the first, the per-period multiplier is weaker; strictly, as soon as “fail early, deliver late” paths have value.

C. The chain: Proposition 1′ (cascade)

Tiers k = 0…K, each saturated. Tier k+1 produces its score from the legible outputs of tier k; its constraint sets, through its own multiplier, the effective threshold τ_k and the induced multiplier λ_k = h_k(λ_{k+1}), h_k increasing (recursion by the envelope, top down). By induction, the vector (λ_K,…,λ₁) is determined by the head constraint and the tier technologies, never by the directions L_k. Hence: stationary damage at the bottom depends on the chain only through the inherited multipliers, D₀ = Φ(λ₁), Φ increasing; at saturation and under A3, it is invariant to the gauges chosen at each tier; it decreases if and only if the head pressure decreases, or an intermediate tier passes on less than it receives. This last case is the damper, and it falls out of the formalism: a wedge between λ_{k+1} and λ_k, paid out of tier k’s own slack. Remark (technical lineage): at a single tier, substituting a principal’s objective for the survival constraint, (i)–(iii) recover Holmström-Milgrom’s multitasking; the chain, saturation as a regime, the conversion, and the damper are not, to my knowledge, assembled there.

D. Proposition 2: detection

Addition: a new action (exploration), of cost c > 0 per attempt; failure probability 1 − p; failure yields a null score and is seen; success yields high value, legible with delay. The new failure: a failed attempt, costly, verifiable, outside the actor’s already profitable repertoire; formally, an action outside the support of his previously profitable allocations. Three positions relative to the threshold: the safe (σ < 1), the taut (σ = 1), the doomed (no safe allocation reaches τ). (i) Zero failures under tension: for the taut, any attempt withdraws c from survival effort and yields nothing within the window; exploration is strictly dominated. For the safe, it becomes payable as soon as slack covers c. (ii) The resurrection gamble: for the doomed, failing by a little or by a lot is the same death; any bet whose success would cross τ dominates the safe route, even at negative expectation. This is Jensen-Meckling risk substitution, in tournament form. The failure rate as a function of slack is therefore U-shaped. (iii) Separation of regimes: position (exploration failures come from σ < 1, distress failures from the infeasible zone); the other tail (exploration, chosen under an expectation criterion, jointly thickens the success tail; the distress bet thickens the failure tail alone and lowers the conditional mean). Selection thickens successes and thins failures; gauge-gaming displaces without thickening. Joint thickening among the safe is proper to loosening. (iv) Counterfeiting cost: manufacturing a failure costs c, dominated for the taut, payable only out of slack for the safe; the signal resists Goodhart insofar as a verified failure consumes real capacity that an actor under tension cannot rationally sacrifice.

E. Where it breaks

A1 lifted, vocation floors: if u retains a floor on the illegible, pass-through is partial, and directional invariance falls even at fixed σ; a gauge experienced as control erodes u itself (crowding out of intrinsic motivation: Frey; Deci, Koestner, and Ryan). The law holds in the limit where survival crushes vocation; outside that limit, it bounds from above. A3 lifted, unequal costs: better aim lowers D_diss at first, then endogenous value brings the defect back to the level σ pins; anisotropy does not refute the law, it dates it. A2 lifted, spillovers: if legible effort co-produces illegible value, pass-through is bounded, and direction matters through the spillover matrix; a real lever, which does not touch σ. Saturation lifted, abundance: λ = 0, allocation follows vocation; outside the law’s domain, not its favorable limiting case.