Gradient Walker: an organization eating its way downhill

Organizations are built out of people and expected to outlive every one of them. That is the strange thing about them. This is a working model of what happens as one grows: a system that eats resources while moving through time, made of parts that can be swapped, keeping books on itself that drift steadily away from the truth.

Not a metaphor. A simulation you can run.

The shape of it

Resources are scattered unevenly across a landscape. Time runs one way, so ground you have already fed on is gone. The organization consumes as it moves and leaves depleted ground behind.

Three roles. Workers do the work. Supervisors hold about eight workers each. Managers hold about seven and a half supervisors. Those ratios are lifted from observed span-of-control research, not invented to make the model behave.

Every run opens with three founders and nobody else. Founders sense poorly — about a quarter of what a worker senses. So the organization can only hire as many people as the founders can personally hold: twenty-four. Growing past that takes structural change, and structural change is where the trouble starts.

There are two operations, and both are one-way.

• Diffuse breaks a supervisory role into interchangeable junior units. Any single loss gets cheaper.
• Consolidate builds a supervisory role out of several units. It is the only thing that turns resources into development.

Each round trip through the two burns mass permanently. You never get back what you spent.

Distribute your mass into interchangeable units and no single loss can hurt you. What you give up is the ability to last, because the concentrated structures that made losses expensive were the same ones that turned food into growth.

That is the claim under test. Most of it held. One part did not, and it is stated below rather than quietly tuned away.

Three phases, and a measurement problem

A run moves through three phases. Foraging: small, sensing the ground finely. Reading: big enough to matter, still able to resolve detail. Committing: mature past the resolution of the ground it stands on.

Span of control

Many small sensors keep fine resolution no matter how big you get. That is not what saturates. What saturates is how much of the edge the center can actually hear.

Units past every supervisor's reach keep working. They keep eating. Almost nothing they find makes it back. The interface shows this plainly as coordinated 64 / 141 — the organization is 141 people and hears 64 of them.

The accounts

The center never sees the organization. It sees reports.

Two errors run in opposite directions. Unheard units make the center undercount itself. Units that are good at describing themselves make it overcount. And new hires are recruited and socialized to resemble whoever looks strongest on the reports.

Nobody is cheating. Average reported performance rises anyway.

The ladder of sacrifice

When runway gets short, the organization makes the smallest cut that will do, then repeats. It spends its final, largest cut only when nothing smaller buys enough time.

Layoffs damage the people who remain, for a long time. The organization's own projections do not price that in.

What the accounts say versus what is true

Ten runs, sampled every second. Average self-accounting climbs steadily with size, and average skill falls over exactly the same span:

At 280 units the organization believes it is taking in about 25 percent more than it actually is. The overstatement grows with size. It does not wobble and it does not correct.

Nothing dishonest happens. The measurement system keeps what it can see, and what it can see has been selecting for visibility the whole way up.

Two numbers that pull apart

Every person carries two values that have nothing to do with each other: how good they are at the work, and how good they are at accounting for themselves. The organization never measures the first. It responds entirely to the second.

The two get handled in opposite ways, and neither is a decision anyone made.

• Visible reporting success gets copied upward. New people are recruited and socialized to match whoever looks best on paper. So average reporting ability rises.
• Skill gets hired to the mean. Founders were selected for high ability. Later hires are recruited against a standard — and a standard is an average. So average skill falls.

No individual hire is a mistake. Nobody games anything. Growth just replaces selective hiring with a hiring bar, and the bar sits lower than the people who set it.

What layoffs actually do

Here is the part that surprised me. A deliberate cut removes people of roughly 0.90 the average skill. Slightly weaker than the organization as a whole — not stronger, not sharper. Output is partly visible, so weaker workers report a little less and are a little likelier to go.

Which locates the damage precisely: dilution happens in hiring, not in firing.

A shrinking organization is roughly doing what it thinks it is doing. A growing one is quietly lowering its own standard with every successful hire, and nothing internal records it.

The two columns move apart for the entire life of the organization. Skill at the work goes down. Skill at describing the work goes up. Neither is anyone's policy. Both are what growth does when the only thing you can observe is a report.

Where the assumptions come from

Almost nothing here is original. It is a synthesis, and naming the sources is both credit and a statement about how far each one can be trusted.

The two poles

The organization sits between slime mould and water. Physarum polycephalum follows food trails and can rebuild transport networks about as efficiently as engineered ones. Water follows the landscape and remembers nothing. A real organization is neither, and the tension between memory and pure gradient-following is the whole middle of the thing.

Next to that, stigmergy — Pierre-Paul Grassé. Termites coordinate with no controller by acting on traces other termites left. Headless coordination is treated here as the substrate, not a figure of speech.

Interchangeable units

The core claim sits close to Harry Braverman in Labor and Monopoly Capital: management splits conception from execution and turns skilled work into interchangeable labor. That is why every person here carries a skill value the accounts cannot read, and why it falls as the organization grows.

One honest wrinkle. Diffuse — breaking a supervisor into junior units — is de-layering, and it weakens the managerial layer. Braverman's deskilling strengthens it by moving conception up. The two run opposite on the axis he cared about. Diffuse captures the consequence he described, interchangeability and cheap losses, not the mechanism. The mechanism lives in hiring.

Span of control

V. A. Graicunas and Lyndall Urwick argued in the 1930s that a supervisor's load grows combinatorially with direct reports. The model uses modern empirical numbers instead: engagement peaks around eight or nine direct reports, Amazon typically runs six to eight, Google seven to ten. Here it is eight per supervisor and roughly seven and a half supervisors per manager.

Who actually decides

Joseph Bower and Robert Burgelman showed that real strategy comes out of the pattern of resource commitments made throughout a firm — not out of strategy statements. Senior management's actual job is setting the structural context that selects among proposals.

The clearest case is Intel leaving memory. A rule maximizing margin per wafer start, plus monthly ranking by gross margin, moved capacity out of DRAM and into microprocessors before senior management decided to exit the memory business. The decision was already made by the time anyone made it.

Clayton Christensen supplies the corollary: allocation rationally follows current margins, so strategically important work loses the competition for resources. Good management practice is what kills it.

Politics inside the ledger

David Scharfstein and Jeremy Stein showed division managers rent-seek, headquarters ends up paying them off with preferential budgets, and you get a kind of socialism in internal capital markets where weaker divisions are subsidized by stronger ones. A unit's share of resources reflects its standing, not just its output. That is where the self-accounting mechanic comes from.

Experience, and what follows a cut

T. P. Wright's learning curve and BCG's experience curve: unit costs fall by a fairly consistent fraction with every doubling of cumulative output. A mature organization gets more out of the same ground.

Pulling the other way, research on layoff survivors: roughly three-quarters report their own productivity dropping afterward, with recovery taking four months to a year, while leadership typically expects three. That gap is modeled directly. The organization's projected cost of a cut is deliberately more optimistic than the measured outcome.

How the numbers get made

Every claim was tested headlessly, before and after implementation. The simulation runs with no rendering, sixteen to twenty-four seeds at a time. Analysis scripts report peak size by policy, where the economy turns over, whether a deliberate cut reaches a supervisor, and how far the accounts have drifted from the organization.

Being plausible is not enough to survive. Several claims did not:

• Growth blinds the organization. Implemented as bulk-driven coarse-graining. Measured: growth stalled at thirty units and the middle phase vanished entirely. The degradation belonged in the channel between edge and center, not in the sensing organ.
• A smaller cut before the terminal cut buys time. Measured against no ladder at all: 30.6m versus 30.7m. On average it buys nothing. Cutting eagerly costs four minutes and a third of the runs.
• Cuts can restore solvency. The ladder never fired across eleven runs. An organization starving on depleted ground has almost no intake, and no cut makes income exceed cost. Cuts buy runway. They cannot make revenue.

Current results. Sixteen seeds per policy, forty-minute cap. eff is the share of what the edge sweeps up that the center can actually use.

The pattern is an arch, not a slope.

Hands-off stays small, stays underdeveloped, and dies anyway. Tightly managed grows large, develops furthest, and dies soonest. The longest-lived is neither: lean, deliberately under-managed, carrying more edge than it can fully hear. It gets there while growing to nearly twice the size and two and a half times the maturity of the hands-off case.

The margin is about six minutes out of twenty-six. But note what it costs to reach: every tight setting does no better than doing nothing, and the tightest does worse. Coordination is not what helps. Some coordination is.

Tighter policy cuts more supervisors as overhead. A supervisor's report says nothing about the eight people who go invisible when that supervisor is removed.

Why large firms don't die on schedule

Early versions of every run ended in starvation. Which implied a lot of real organizations should already be dead. They are not — they move faster than smaller competitors — and a single-gradient model had no way to explain it.

A large firm is not one organization on one gradient. It is several, each working its own ground, with a center allocating between them. Alfred Chandler's multidivisional form, which is the structure all the resource-allocation literature above was written about. One organization on one gradient must eventually exhaust it. That is arithmetic, not strategy.

What matters past size is resolution. A division attends to a narrower domain than the whole, so it can resolve detail the center has already coarse-grained away. A mature organization is blind at its own scale and still contains parts that can see.

Splitting buys that focus and costs a supervisory role. So opening a second gradient competes directly with coordinating the ground you already hold.

The bill for it

The center cannot see what any division actually earned. It sees reports, and it funds reports. Over a run the divisions' average self-accounting roughly doubles — about 1.0 to about 2.0 — with nobody gaming anything. New divisions get founded in the image of whichever reports looked strongest, and strong reports are what got funded.

So the organization stops starving for lack of resources. It starves by funding the division that describes itself best.

That is a worse failure, because every single step looks like ordinary management. Christensen's dilemma shows up as a consequence of the machinery rather than as something asserted from outside it.

The illegibility is a property of the work, not of bad reporting. A division on established ground reports in categories the center already understands. A division attending to detail the center cannot resolve has no shared language for it, and good faith on both sides does not create one.

So the error grows with scale instead of averaging out:

The bigger the organization, the worse it funds its own divisions — and the ones it systematically underfunds are the ones working ground it cannot see.

The only counter is incubation: fund a division too young to have a record according to need rather than demonstrated output, and take the difference from the parts that can already prove themselves. Without it, measurement shows no new gradient ever establishes, and nesting stops paying at all.

Where it breaks

This version of the walker does exactly one thing to its landscape. It eats it.

Real organizations at scale also build landscape — infrastructure, marketplaces, supplier terms, standards — raising their own future yield and lowering everyone else's. Biology has a frame for this: niche construction, from John Odling-Smee, Kevin Laland and Marcus Feldman, where organisms modify their environments and those modifications feed back into selection. A beaver is not a better forager than a mouse. A beaver is an animal that makes ponds.

Nothing in the model can express that yet. It is the obvious next step and a risky one, because landscape capture threatens to make a large organization effectively unkillable. That is why nesting had to come first. Now that an organization can be punished for funding its divisions badly, it can be handed the power to reshape its ground without becoming immortal.

The quieter omission

Each division here works its own ground, so a new division's success never costs an established one. Real portfolios are not like that.

Kodak's problem was never that digital photography was hard to understand. Kodak invented it. The problem was that digital would have destroyed film.

What is modeled is Christensen's neglect of the new: illegible, underfunded, withers. What is not modeled is fear of the new — an incumbent that sees the new thing clearly and suppresses it precisely because its success is the incumbent's loss. Different failures. Different remedies. Only the first one is in here.

Letting divisions overlap would introduce the second: one division's consumption would thin another's ground, giving the allocation rule a real motive to starve it. It would also make the nesting results considerably messier, and every number above would need re-measuring before any claim could be kept.

Colophon

Gradient Walker is built in React Three Fiber over three.js. The landscape is seeded gradient noise with exact CPU and GPU parity, folded per octave so channels appear at every scale, and gated so that detail below the organization's own resolution stops existing — in the geometry and in the simulation at the same instant.

Every figure here is measured output, not illustration.