Losing Your Shape on Purpose

Risk as adaptation, and why organisms manage it differently to organisations

Human risk management holds shape. It defines a target, gathers information until the model looks stable, and then commits — a blueprint, a budget, a Gantt chart, a loss function. Biological risk management does something else: it loses shape on purpose, and gets paid for it. Adam Timlett's argument is that the difference is not temperamental but formal. The preference for shape conservation is built into the mathematics that economists, engineers and machine learning researchers reach for, and it makes an entire class of strategy invisible.

Mass, shape, and compounding error

Fire an arrow at a target. A fixed angular error at the moment of release produces a larger absolute miss the further the target is. The error rate is constant; the displacement grows with everything invested in the original decision.

The same structure appears wherever information is frozen at a point in time and then scaled: a tower that leans further the taller it gets, a JPEG that degrades as it's enlarged, a royalty line whose data errors matter more as the sums attached to it grow. Timlett's decomposition names three ingredients — the shape of an organisation, the mass added to it (any investment of money, energy or action), and the information added as it grows. Error compounds because mass is added to a single model without new information entering alongside it.

Two consequences follow directly. First, unless the eventual mass is known in advance, the required precision cannot be specified in advance — which rules out the up-front-accuracy strategy for anything whose scale is open-ended. Second, an exponential growth curve is better read as the loss of shape than as a shape: it is the signature of a system being scaled faster than corrective information is entering it. In a population, exponential growth adds genetic variety as it adds mass. In a mono-model, it only amplifies whatever error was there at the start.

The information dividend does not come from uncertainty

An agent occupies a world with two locations, Home and Park, and two states, sun and rain. Rain favours Home, sun favours Park, and transitions cost something. Given a forecast, the agent either moves or stays.

The result that matters: the payoff from moving around survives a perfectly accurate long-range forecast. A ten-day prediction of 80% sun can be exactly right and still fail to specify which hours are sunny. The dividend from responding to daily information is therefore not a function of ignorance about the weather. It requires two conditions instead — that the distribution is non-fractal, meaning the one-day forecast doesn't resemble the ten-day one, and that transition costs are low enough for the difference to be worth acting on. If the short-range forecast is simply a shrunk copy of the long-range one, there is nothing to collect, however uncertain either may be.

This inverts the standard justification for agility. Agility is usually defended as a hedge against unreliable planning: forecasts fail, so stay light. On this account it is the mechanism by which accurate planning is realised. High-level estimates should be as precise as possible, precisely so that everything below them can be left free to move — resource allocation predicted long, task selection decided short. Locking scope to protect an estimate confuses the two: as long as items entering and leaving scope balance, the resource forecast survives intact while the dividend is collected.

It also relocates the standing critique of decision theory. The radical-uncertainty tradition — Ellsberg onward — argues that decision theory assumes more knowledge of distributions than anyone has, and is therefore insufficiently cautious. The Weather Agent argues the opposite failure: decision theory under-exploits, leaving value uncollected even when the world behaves exactly as forecast. Noisy information argues for adding inertia and smoothing, which is the hedging response. Reliable information argues for reducing inertia. Radical uncertainty is close to insoluble; this is not.

Note that feedback does not capture the distinction. A person self-righting on a rocking deck and an agent relocating to exploit sunshine both look like closed loops. One is using information to preserve shape; the other is sacrificing shape to exploit information. The vocabulary of feedback, linear or non-linear, cannot tell them apart, and neither can a dynamical description that classifies systems by the equations that fit their trajectories.

Meta-modularity, and the insufficiency of redundancy

Resilience is usually discussed as redundancy: spares, backups, slack. That collapses two different things.

Multi-modularity is interchangeable standardised components — swap a failed bulb for an identical one. It solves damage, repair and cheap manufacture. It cannot solve novelty, because every replacement is identical to what it replaces, so nothing about how the function is performed can change.

Mono-modularity is the single general model applied across all domains, with each situation treated as data: Bayesian decision theory, an optimisation function, a large language model. It scales well and, by Wolpert's No Free Lunch results, cannot dominate specialist alternatives across all problem classes.

Meta-modularity is a population of overlapping, non-identical models of the same or adjacent domains: genetic diversity across a population, a plant acclimating by switching gene expression to local soil and moisture, E. coli metabolising eighty-odd carbon sources, two colleagues each holding partial and inconsistent knowledge of the same records. Redundancy explains survival. Only overlapping-but-different models explain adaptation, because only they contain information the primary model doesn't.

The exclusion is sharp enough to be checkable: on this definition a neural network is not a meta-model, whatever the surrounding literature calls it, because inspecting it does not reveal the domain it models. Nor is a hierarchy of data types. Going up a level of abstraction produces something that contains no information about any domain at all.

Friction as stored optionality

Apoptosis — programmed cell death — should be a one-bit channel: self-destruct, or don't. Measured against its extrinsic signalling pathway, the channel capacity comes in at roughly 0.92 bits. It does not reliably transmit the instruction it exists to transmit.

The default reading is noise: a defect, something better engineering would remove. The alternative reading is that the channel is not performing one computation. It is simultaneously damping over-response to an unreliable upstream signal and smoothing the response across a local population of cells, since synchronised death across a dense region would leave a void more damaging than the infection. Two units of selection — the cell and the tissue — are being optimised at once, and the compromise appears as unreliability.

Generalised, this becomes a claim about what channels are for. Shannon's framework treats a message's entropy against a channel's capacity, with shortfall attributed to noise. A computational channel rejects inputs meaningfully rather than corrupting them, and its entropy rate — bits submitted per bit accepted — measures work done rather than information lost. A rate of 1.5 means a third of submitted bits are rejected; a password validator with strong requirements has a high rate by design, and that is the point of it.

The consequence is counterintuitive and does real work. What matters is not the rejection rate but the cost of rejecting. Two processes with identical entropy rates can differ enormously in the flow ratio between accepted and rejected inputs: writing 226 words before binning a page versus 680 before binning a page, at the same 66% acceptance rate. The cheaper the rejection, the more options can be considered at no additional cost — which is a measure of how efficiently a system entertains alternatives.

Read this way, the friction between a generator language and an acceptor language is not waste but latent optionality: the rejected inputs are proximate options, already produced by a generator that carries information about them, and reachable by small adjustments to the acceptor. Efficiency drives that reduce this friction — in bioengineering, in organisational process design — are removing the substrate's capacity to adapt, which is one explanation for why organisms resist being optimised in the directions engineers push them.

This sits at an angle to the free energy principle rather than against it. Both describe homeostasis as the central object. But where surprise minimisation treats deviation as the quantity to be suppressed, this framework treats the capacity to deviate as the thing being purchased — robustness bought in order to afford surprise, because that is where the dividend lives. The disagreement is substantive and is not settled by either formalism.

Meta-accuracy, meta-precision, and crossing valleys

Statistical precision is the relation of attempts to each other; accuracy is their relation to a reference. Lifted a level: meta-precision is the apparatus that makes attempts commensurable — the standard distance, the shared metric, the fair archery contest where every competitor shoots the same distance at the same target. Meta-accuracy is hitting any of several available targets, each with its own encoding.

For artificial problems these coincide, because the metric that ranks attempts is also the metric that defines success. For natural problems they trade off. A natural problem can therefore be defined as one where raising meta-accuracy necessarily costs meta-precision — and the diagnostic signature is loss of commensurability. Alternative splicing, where introns are resolved differently by context and a gene contributes to multiple proteins; disordered proteins with no fixed conformation participating in several pathways; overlapping coding sequences. In each case there is no single distance-to-fitness that can be computed for a gene, which is exactly what makes them hard to model and exactly what raising meta-accuracy predicts.

Which sets up the harder problem. Complex functions cannot be reached one improvement at a time: append characters to a valid word and you mostly produce invalid strings, with the next valid word some distance away. This is the fitness valley in evolution and the local minimum in optimisation — no path on which every step improves.

Wagner's answer is neutral drift. Phenotypic robustness tolerates mutations that do nothing, letting a population wander a high-dimensional neutral space until it lands near an innovation. The objection is that it costs reproductive ground in a Red Queen race, and that a backup pathway targeting the same niche doesn't explain how new environments get explored.

The alternative runs on fragility. Suppose the broken intermediate is not neutral but positively beneficial — in a simpler niche. There is evidence: yeast in a nutrient-complete medium accumulates mutations disabling amino-acid synthesis pathways it doesn't currently need, and shedding the metabolic load raises growth rate. Deleterious in the wild, beneficial in the lab. Simple environments are not laboratory artefacts either; parasitic species outnumber free-living ones, and environmental simplification is among Nature's most common strategies.

The mechanism, in schematic form: a sequence valid in a complex niche is extended and becomes invalid there, but each intermediate is valid — advantageous, as a disabled pathway — in a simpler niche. A further extension restores validity in the complex niche, now with a longer, more complex gene. Every step paid, in whichever niche the organism was occupying at the time. The two niche languages are related as a hedge: what breaks one benefits the other, so the fitness landscapes are inversely correlated rather than independent. Only minimal agency is required — a population distributing across niches each generation is sufficient; more agency, in the form of sampling and choosing environments, raises efficiency.

The test

The definition of a natural problem is sharp enough to be uncomfortable: it is a problem where making attempts comparable to each other is what prevents solving them. The standardised metric that permits ranking, benchmarking and gradient-following is the same apparatus that forbids hitting any of the other targets available.

If that is right, the indictment is broad. It covers optimisation, where single-objective methods stall in local minima not by accident but because complex functions have gaps that a single metric cannot cross. It covers project management, where scope discipline protects a forecast at the cost of the dividend the forecast was supposed to enable. And it covers scientific practice, where standardising terminology before a domain is understood is precisely the trade being argued against — meta-precision purchased at the expense of meta-accuracy, in a field whose problems are natural ones.

The framework summarised here is developed in Adam Timlett's work at Turing Meta, including a technical appendix formalising computational entropy, stressor-driven design, and a deliberately negative theory of consciousness.