Steering by Value: The Signal That Isn’t There
A single feature is easy to govern. One person can hold it in their head, knows roughly why it matters and whether it is worth doing, and needs no apparatus to decide. The trouble is that work does not stay that size. A large initiative, a portfolio, a year of competing demands eventually crosses the point where no one can hold the whole of it at once, and at that point something has to stand in for the judgement a single mind used to make.
The candidate the field reaches for, almost by reflex, is value. Measure it, steer by it, and let the numbers carry the decisions that no longer fit inside one head. This is not a foolish move. It is the reasonable one, and that is exactly why it is worth taking apart. The thing reached for at the moment it is most needed turns out to be the thing that cannot do the job. And the reason is not that value is hard to measure. It is that the word, at the moment of reaching, is quietly doing two different jobs, and only one of them is being asked to do work it cannot do.
Two jobs the word does
Picture a portfolio review. Someone looks at an initiative and asks, “is this valuable?” The sentence is ordinary, and the answer can mean two entirely different things.
In the first sense, value is a control signal: a reading you take every cycle to close a feedback loop, the thing a dial would show if you could mount one on the work. You watch the number, and when it moves the wrong way you correct. This is value as the input to a steering mechanism, consulted continuously, the quantity the system is driven by.
In the second sense, value is an owned judgement: an accountable person looks at something they cannot measure, makes a call, names who owns the resulting bet, and proceeds. This is a one-shot act, not a continuous reading. It does not pretend to be a measurement. It is a person putting their name against a decision and carrying it.
Only the first sense is at stake in what follows. The second survives everything in this essay; it is necessary, legitimate, and the thing accountable people are there for. It simply belongs outside the control loop rather than inside it, which is a matter taken up in a later piece, The Only Decision You Control.¹ For now the point is narrower and sharper: the two senses are different in kind, and almost every confusion in this territory comes from sliding between them, borrowing the respectability of the owned judgement to license the continuous steering signal, or borrowing the apparent rigour of the signal to dress up what was only ever a judgement.
So a rule, binding on the reader and the writer alike for the rest of the series: from here, every appearance of the word names one sense or the other, and the slide between them is the move to watch for.
The field already equivocates, and it does so in a place with a standardised name. Earned Value Management is a decades-old technique that carries “value” in its title while measuring something else entirely. Its central quantity, earned value, is defined as “the measure of work performed expressed in terms of the budget authorized” for that work.² That is conformance to a plan. Its two headline indices read the same way: the cost performance index is earned value over actual cost, and the schedule performance index is earned value over planned value, so both report cost-efficiency and schedule-conformance against a committed baseline. A project can hold both at exactly 1.0 while building something nobody wants. The vocabulary has equivocated before any argument has begun. That is the whole use of Earned Value Management here; the technique itself is not under investigation and is not the target. It is a standing specimen of the word dressed as value while reading as cost.
What a control signal must be
If value is going to serve as a control signal, the thing that drives a feedback loop, then it has to meet the requirements any such signal meets. There are three, and value fails each one, by a different argument.
A control signal must be observable on the loop’s timescale. You cannot correct every cycle on a reading that only resolves quarters later. Value fails this immediately. It is lagging and confounded: a feature ships this month and whatever revenue or retention it produces arrives much later, tangled inseparably with a price change, a seasonal swing, a competitor’s move, and a dozen other things in motion. By the time the signal resolves, it is no longer attached to the action that supposedly caused it. It is off-loop.
A control signal must be attributable to the action. Steering means knowing that this move produced that result. And here value fails not for want of effort but by construction, because what you would need to observe is a counterfactual: what would have happened had you not made the move. That comparison is unobservable in principle. The formal version of this is old and exact. The fundamental problem of causal inference is that you can never observe both potential outcomes on the same unit, the outcome with the action and the outcome without it, and so you can never directly observe the effect of the cause on that unit.³ You see the world in which you shipped the feature; the world in which you did not is gone. The worth of the specific bet is precisely the difference between them, and that difference is not available to measurement.
A control signal must be robust to being optimised against. The instant a value measure becomes the thing the organisation is steered by, it stops measuring value and starts measuring what people do to the measure. “When a measure becomes a target, it ceases to be a good measure.”⁴ This is not a defect you can engineer out with a better-designed metric. It is what steering-by-a-metric is. The scores fill with whatever clears the gate, and the gap between the score and the thing it stood for widens in proportion to how seriously the score is taken.
The structure here is what makes the demolition hard to beat. The three failures are independent. They are not one objection stated three times; they are three different arguments, and an opponent has to defeat all three. One of them, attribution, fails in principle rather than in practice, and that is the leg that will carry the heaviest weight in a moment.
Why more rigour cannot help
The natural reply to all this is to reach for precision. Measure it better. Add criteria. Adopt a more sophisticated framework. Surely the problem is that the measurement is crude, and a finer instrument would close the gap.
It would not, and the reason is structural. The thing being measured is a counterfactual, and a counterfactual is unobservable however fine the instrument. You cannot raise the precision of a measurement of something that is not there to be measured. Adding rigour to an unobservable does not converge on the truth; it manufactures false precision, estimates dressed as measurements, which is worse than honest ignorance, because honest ignorance does not get defended. A number does. The single line worth holding onto is this: more detail sharpens the proxy and does nothing to the gap between the proxy and the counterfactual it stands in for, and the error lives in that gap, where no amount of resolution reaches it.
It is worth turning this same scrutiny on the claim this essay is making, since the section has just leaned hard on the word type. The claim is that value’s failure as a control signal is a type error and not merely a difficulty, and honesty requires admitting that the three legs do not earn that word equally. The attribution leg earns it: Holland’s unobservability is in principle, not a matter of instrument quality, and that is a difference of kind. The off-loop leg and the gaming leg are severe and structural, but they sit closer to “practical and very hard” than to “impossible in principle”. So the strong framing, type error rather than difficulty, rests hardest on attribution, while all three legs remain load-bearing for the weaker and still-sufficient claim that value cannot be a control signal. Saying this plainly does not weaken the argument. It shows the type claim surviving its own scrutiny on the leg that matters, rather than being asserted evenly across three legs that cannot each bear it.
There is a sting in the tail. The more quantitatively ambitious the apparatus, the more dangerous it becomes, because it lends the gaming and the false precision more authority, not less. Rigour applied to an unobservable raises confidence without raising correctness, and that divergence, confidence climbing while correctness stays put, is the indictment.
One root, three surfaces
It is tempting to treat the three failures as separate obstacles to be reduced one at a time: a measurement problem here, a gaming problem there, an attribution problem to be modelled around. They are not separable, because they are one phenomenon seen from three angles.
Get Bruno Baketarić’s stories in your inbox
Join Medium for free to get updates from this writer.
Follow the chain. Attribution fails, because value is a counterfactual, which is why you fall back on estimates. Estimates are judgements dressed as numbers, which is what invites gaming. Gaming is people acting on the signal rather than on the world, which is why the political dimension takes over the whole exercise. Underneath every link sits a single fact: value is an unobservable counterfactual, and everything downstream is people negotiating over an unobservable. Measurement was never going to settle that, because there is nothing there for the measurement to land on.
This is why you cannot fix one surface without the others reappearing. Tighten the estimates and the politics simply move to where the estimates are set. Police the gaming and it migrates to the definition of the metric. The surfaces are not independent leaks to be patched; they are the same root showing through in three places. A prioritisation meeting where the value scores are visibly the residue of who argued hardest is not exhibiting two problems, a politics problem and an estimation problem. It is exhibiting one, twice.
The measurability inverse law
From that single root falls a regularity worth stating on its own. Measurability falls as strategic weight rises.
The mechanism is not mysterious. The work whose value you can plainly read is systematically the work that least needs steering: low-level, technical, locally coupled, with a short and legible path from the change to its effect. The work that carries real strategic weight, a market entry, a platform bet, a change of direction, is exactly the work whose value is least readable, because the path from the decision to any outcome runs through years and through everything else the world is doing. The two properties move in opposite directions as a matter of structure, not coincidence.
The consequence is awkward for the whole programme. Value-based steering is most rigorous precisely where it is least necessary, and least available precisely where it is most wanted. That makes it a method aimed at the wrong target, rather than a sound method with rough edges aimed at the right one. And the law disarms the obvious counterexample before it is raised. The cases where value-steering appears to work are the cases where the stakes stayed low enough that the coupling stayed short. That is not a concession to be explained away; it is the boundary this essay closes on, and it is worth being precise about it shortly.
The pragmatist’s objection
The strongest opponent here is not the one who insists the measurement can be made rigorous enough. That position has already lost. The strongest opponent is the pragmatist, and the pragmatist says something harder to answer.
Everything is a proxy, the pragmatist points out. Flow metrics are proxies. Value metrics are proxies. Cycle time is not the thing you care about either; it is a stand-in, just like a value score. So this whole argument is not epistemology at all. It is a preference for one set of proxies over another, dressed up as a difference of principle. Why is steering by a flow-proxy any more defensible than steering by a value-proxy? Pick your imperfect number and stop pretending one is purer than the other.
The first move is to concede what must be conceded, because the concession costs nothing. Flow is not the thing ultimately cared about, and there is no intention of defending it as a better proxy for value. That defence is the one that loses, and it should be refused outright.
The actual answer refuses the proxy frame rather than trying to win inside it. Flow metrics are direct readings of a real object: the system’s own behaviour. Work in progress, cycle time, throughput, the age of an item, all are observable properties of the machine, and they are wanted for their own sake. You want the machine to be predictable and unclogged whether or not any particular item passing through it turns out to be worth anything. A value metric is a different kind of thing entirely: a proxy reading of something that is not directly observable, and it fails at the proxy step, because the thing it proxies is the counterfactual that is not there to be read. So the distinction is not that flow is a cleaner stand-in for the same underlying quantity. It is that flow reads a different object, an observable one, and value reads a signal that does not exist.
This matters, because saying “flow is the more principled value-proxy” invites exactly the rejoinder the pragmatist is fishing for: that you merely stopped the proxy chain one link short and called your stopping-point direct. The defensible claim is the stronger and stranger one. Flow is not a better proxy for value. It is a reading of a different thing, wanted in its own right. The image to hold is two dials. One reads the machine, faithfully, and tells you something true every time you look. The other is wired to nothing, and its needle moves because people move it. The difference between them is one of type, not of quality.
And there is a price to pay for this answer, paid openly: it means you are not steering by value at all. Which is precisely the thesis, so the price is no price. It does leave a question standing, though. If flow is not proxying value, what is it for? It has a job, and the job is a real one, but it is built elsewhere, in You Can Kill a Bet, Not Confirm One.¹ For now it is enough that flow’s job is not the one the pragmatist assumed it was failing at.
The boundary: coupling distance
A flat, totalising claim is easier to attack than one that draws its own edge. So it is worth naming the variable that decides when value-steering can work at all, and drawing the boundary from the inside.
Consider an internal IT organisation. It is structurally better placed than almost anyone to steer by something value-like, and the reason is specific. It can measure the very thing it exists to change: the business process. Process cycle-time, error rates, manual touches, rework, all are directly measurable. Process cost follows from them by arithmetic. Money follows from that. The chain from the thing changed to the money is short, and every link in it is mechanical rather than behavioural.
The real reason this works is not “better measurability” loosely stated. It is that the attribution problem, the root of all three failures, is structurally smaller here, because the counterfactual nearly collapses. Re-engineer an invoice process by removing three manual approval steps and the resulting change in throughput is attributable to the change, because no confounding market, no customer psychology, no seasonality stands between the intervention and the reading. The world without the change and the world with it are close enough together that the difference between them is, for practical purposes, observable. This is not a loophole in the fundamental problem; it is the recognised condition under which a causal effect becomes estimable, a controlled change with little standing between cause and measured effect.⁵
So the thesis is better stated as a gradient than as a wall. The viability of value-steering is inversely proportional to the coupling distance between the system you can measure and the value you care about: short for internal IT changing a process it can instrument directly, long for a product thrown against a market it can only guess at. This absorbs the obvious objection, the enterprise software rollout that returned a real, measurable amount, rather than denying it. That case worked because it satisfied the antecedent: short coupling, a near-collapsed counterfactual, a process measured directly. It is the boundary case, not a refutation.
The careful part is that none of this breaches the thesis. Internal IT measures the process, which is a property of the system, not its own value-as-such. It did not solve value-measurement; its domain handed it an unusually faithful proxy for a particular, locally coupled question. This is the steer-the-system posture holding up under the most favourable conditions available anywhere, which is a confirmation of the posture rather than an exception to it. And the same line reappears the moment IT finishes its work. It can measure the capacity it freed; it cannot measure what the organisation does with that freed capacity, which runs back out into the long-coupled world of everything else. Freed capacity the business then declines to redeploy is an organisation problem, not an IT problem, the same way that a customer who buys software, uses it to generate value, and then throws that value out of the window has not refuted the software. That boundary, between the measurable delta and what is done with it, is the same steer-the-system-and-own-the-bets line that runs through everything that follows, and it will turn up again at every scale: as the limit on which premises a delivery loop can falsify, and as the floor under the cost of any probe.
That is the shape of the claim, then, and the boundary is not an escape from it. Value cannot serve as a control signal. This is a type error and not a difficulty rating, and the one regime where the proxy chain is short enough to make the steering look like it works is the regime of short coupling distance. Which leaves the large, awkward remainder: all the work that was never on the value axis to begin with, and has been made to pretend it was.
¹ The Signal That Isn’t There is the first of a four-part series, Steering by Value. References here to later parts (The Work That Has No Number, You Can Kill a Bet, Not Confirm One and The Only Decision You Control) name where each point is developed; links will be added as those parts are published.
² Project Management Institute, A Guide to the Project Management Body of Knowledge (PMBOK Guide), 7th ed. (2021), glossary, s.v. “earned value”; the 7th edition refers the fuller treatment to the dedicated Standard for Earned Value Management. The cost and schedule performance indices (CPI = EV/AC, SPI = EV/PV) read efficiency and conformance against the authorised baseline, never worth.
³ Paul W. Holland, “Statistics and Causal Inference”, Journal of the American Statistical Association 81, no. 396 (1986), p. 947, on the fundamental problem of causal inference, that it is impossible to observe both potential outcomes on the same unit and therefore to observe the effect of a cause on that unit. Holland is explicit that this does not make causal inference impossible, only impossible without untested assumptions; the in-principle unobservability claimed here is for the effect on the single unit, the worth of the specific bet, not for value in general.
⁴ Marilyn Strathern, “‘Improving ratings’: audit in the British University system”, European Review 5, no. 3 (1997), p. 308, giving this formulation of Goodhart’s law (relayed via Hoskin’s gloss on Goodhart’s original, differently phrased, monetary-control observation).
⁵ Holland, “Statistics and Causal Inference”, p. 947 and §4, on the routes past the fundamental problem: the scientific solution, resting on homogeneity or invariance assumptions, and the statistical solution, resting on randomisation across a population. A controlled process re-engineering is the statistical solution in operational dress, which is why the confound nearly vanishes when coupling is short. The coupling-distance framing itself is the author’s own.
