The Evolutionary Functions
Create, Explore, Attack, Defend: the 65 recurring jobs that evolving systems do, how they are tested one system at a time, and what 12,520 rows of testing show about when each job was first done.
Written 26 October 2026, before we began walking each way a system takes in its world (8 November 2026). The figures below will be re-measured when the re-walks are in. Rows this paper once called "shape" are the done level: the job done by plain physics, with nothing holding it.
Abstract. Any system that varies and is selected on will, if it lasts, come to do a small set of recurring jobs. We call these the Functions, and we group them into four families that together make a survival toolkit: Create (make new things), Explore (find out), Attack (take or outcompete), Defend (hold what you have). The 2025 paper proposed about sixty of them by comparing three systems that share no substrate: the global financial market, the human species, and information security. This version reports what happened when the list was made exact (65 Functions, each with a definition, a selection pressure and a place in a skill tree) and then tested against every system the engine has walked, from the Big Bang to TCP/IP: 202 systems, 12,520 rows, each row sourced. Four findings. Below life the Functions are at work at the done level: for six rungs plain physics does the jobs, with nothing holding them; at the fold into life the held level appears at once and climbs. The four families do not arrive together: Create's done-level rows are present at the universal floor, Defend's and Attack's within three rungs, and Explore's not at all until life, so nothing senses before something lives. Credit flows down to first-doers: six early systems own most of what later systems inherit. And the list itself is now a tested object: 36 candidate Functions were reviewed this month, 15 were not Functions, 17 fitted existing ones, and 3 were added. We give the rule that decides what counts, the reading rules below the fold, and the Functions the record has not yet found anywhere. Everything here is read from the live record and labelled as ours.
Contents: About this version · The claim · The four families · The 65 Functions · The skill tree · How a system is tested · Below the fold · What the record shows · What counts as a Function · What changed · Open questions · How to read the data
About this version
The 2025 paper argued for the four families by analogy. It laid humans, markets and cybersecurity side by side in four tables and showed that each had evolved, by different means, a mechanism for the same job: a market's circuit breaker and an animal's startle reflex both do Quick Harm Avoidance; a honeypot and a cuttlefish both do Deception. The analogies were drawn from other people's results, and the list of jobs came from a spreadsheet of traits Jerry had assembled over years. This version replaces it. The list has been made exact and loaded into an engine that tests every system against all 65 Functions by the same method, and the evidence is that record, read live on the day of writing. No outside paper is cited.
Figures as of 26 October 2026: 65 Functions in four families; 166 prerequisite edges; 202 systems walked in full on 13 rungs; 12,520 Function rows, of which 1,042 expressed, 177 done-level and 934 inherited; 36 candidate Functions reviewed; 1,332 sources.
The claim
Evolution is not only biology's. Anything made of parts that can be arranged many ways, that keeps producing new arrangements, and that keeps some arrangements and drops others by how well they work, will evolve. Markets do. Technologies do. Cultures, immune systems, and the malware they fight do. And whatever is evolving, the problems it must solve to keep going fall into the same four kinds. It must make new things, because the arrangements it has will eventually stop fitting. It must find out about its surroundings and itself, or it will spend its novelty blind. It must take or outcompete, because resources are finite and others want them. And it must hold what it has against damage, error and attack. Within each kind there are specific jobs that recur wherever the kind is needed, and the same job, done by a gene, a firm, a protocol or a court, is the same Function. The pattern is fractal: it repeats at every scale, in different material.
That was the 2025 claim. The engine turned it into something that can fail. Each Function was given a definition tight enough to say no with, and a selection pressure saying what favours it. Every system walked is tested against every Function, and the row says whether the system does the job (or does it by plain physics alone), inherits it, could do it, lacks it by a source's say-so, or is unknown. If the four families were an artefact of looking at three human-scale systems, the record would show it as rows that will not fit. So far it shows the opposite, and it shows something the 2025 paper could not have seen: when each job was first done.
The four families
| Family | The job | Functions | Root of its tree | First done-level row | First expressed |
|---|---|---|---|---|---|
| Create | generate novelty: new units, new arrangements, new environments to work in | 15 | Generation / Synthesis | rung 0, the universal floor | rung 6, life |
| Explore | find out: sense, search, map, learn, remember | 18 | Sensing / Scanning | rung 6, life | rung 6, life |
| Attack | gain at another's expense: take, breach, evade, deny, co-opt | 14 | (none: every Attack Function needs an Explore or Defend parent) | rung 2, fields and particles | rung 6, life |
| Defend | hold a persisting state: block, detect, repair, contain, deter | 18 | Physical / Virtual Barrier | rung 0, the universal floor | rung 6, life |
The four families as the record now holds them. "First done-level row" is the lowest rung at which any Function of that family is booked at the done level; "first expressed" the lowest at which any is booked as expressed.
The families are not independent. Create feeds Explore with things to test and Explore feeds Create with places to look; Attack provokes Defend and Defend provokes a cleverer Attack; a secure enclave is where risky exploration can happen. The 2025 paper called this a cycle of adaptation. The skill tree below makes it structural: Attack has no root of its own, because every Attack Function needs something from Explore or Defend first. You cannot evade detection until there is detection, and you cannot breach a barrier until there is one.
The 65 Functions
Each Function has a key, a definition, a selection pressure, a family, and a tier in the skill tree (0 is a root). The three added this month are marked.
Create (15)
Generation / Synthesis · Recombination / Composition · Scaffolding / Enablement · Templating / Standardisation · Niche Construction · Quality Improvement / Refinement · Interoperability / Transduction · Encapsulation & Modularity · Amplification / Scale-Up · Symbiosis / Co-Creation · Evolvability / Plasticity Engineering · Distributed Autonomy · Internal Coordination / Regulation · Reproductive Division of Labour (new) · Scheduled Build-and-Dismantle (new).
Explore (18)
Sensing / Scanning · Hypothesis / Goal Selection · Search / Foraging · Mapping / Modelling · Navigation / Pathfinding · Uncertainty Reduction / Learning · Risk-managed Exploration · Intrinsic Motivation / Curiosity · Active Experimentation / Probing · Multi-Scale Data Fusion · Trail-Marking / Externalised Memory · Surprise / Change-Point Detection · Frontier Outposts / Sensing Extension · Hypothesis Portfolio · Optimal Stopping & Patch-Switching · Offline Simulation & Replay · Social Learning / Public Information Use · Positional Identity (Body Map).
Attack (14)
Evade Detection · Overcome Barrier · Encirclement · Speed · Precision Attack · Ranged Attack · Neutralise Defences · Resource Denial / Attrition · Subversion & Co-optation · Prepositioning & Latency · Battlespace Shaping · Leverage Host / Amplification · Divide & Isolate · Cascade Induction.
Defend (18)
Quick Harm Avoidance · Reactive Response · Proactive Defence · Physical / Virtual Barrier · Technique-Based Identification · Self-Monitoring · Delay / Redirection · Validation of Identity / Intention · Validation of Effectiveness · Enclave / Segregation · Deception · Resilience & Repair · Redundancy & Diversity · Triage & Prioritised Allocation · Altruistic Containment · Deterrence · Quality Control & Error Correction · Bridging Supply Gaps (new).
Two examples of how a definition and its pressure are written. Generation / Synthesis: the production of a new functional unit that did not previously exist in the system's repertoire, assembled from more basic primitives the system already has; favoured where novel threats or opportunities outstrip the existing repertoire faster than rearranging it can cover. Recombination / Composition: new capability by rearranging, fusing, splitting or redeploying existing units; favoured wherever generating new primitives is expensive relative to recombining old ones, so the pressure rises with the size of the primitive set.
The skill tree
The Functions are not a flat list. Some need others first, and the record holds 166 such prerequisite edges, arranged in tiers from 0 to 10. There are three roots: Generation / Synthesis for Create, Sensing / Scanning for Explore, and Physical / Virtual Barrier for Defend. Everything else descends from them. Mapping needs Trail-Marking, which needs Search, which needs a Hypothesis, which needs Sensing: you cannot map what you cannot sense. Evade Detection needs Technique-Based Identification, which needs Learning, which needs Experimentation. Cascade Induction, the deepest Attack Function, sits at tier 10 and needs Multi-Scale Data Fusion, because you cannot trigger a chain reaction you cannot see across scales.
The tree is what makes the Functions a ladder rather than a checklist. A system on a low rung has a short tree available to it, and the record shows exactly that: the deeper the tier, the later the rung at which the Function is first expressed. It is also what a fold acts on. A fold links Functions that were already present, the way protein folding brings distant parts of a chain together, and that linking unlocks the next tiers.
How a system is tested
A walk first records what the sources say the system does, then reads those observations against the Functions, with a reason for each row. Expressed: the system's own machinery does the job (held, selected or valued). Done: plain physics does the job, and nothing holds or regulates it. Inherited: the job is done by something the system sits on, and the row names the owner. Possible and unknown mean what they say; absent is booked only when a source says so, and a job looked for and not found is recorded as not found, never as absent. Each row carries a confidence and the source it rests on.
Two rules keep the rows honest across scales. Ownership is first-come. The system that first did a job owns it; later systems inherit. Chemical bonding was first proton and electron, and the electron existed first, so the electron owns bonding and every atom and molecule inherits it. Credit never flows upward: a standing rule turns any row that inherits from a later system into a possibility realised at that later system. The reading changes at the fold. Below life, test a Function's bare mechanism and drop its value-laden clauses, as described next.
Below the fold
Before life, nothing holds a job; the thing a Function names still happens, but nothing else comes along with it. So below rung 6 the walker tests the bare thing. A star's balance of gravity against heat is regulation done by plain physics; a resonance in a carbon nucleus is scaffolding at the done level; the Hoyle gate is a barrier at the done level. None is held. They are booked at the done level, and the record holds 177 of them.
Attack and Defend need a further rule, because their names carry intent. Below the fold an Attack Function is tested as Compete: one system's activity pushing another's valence negative, by taking from a shared pool, stripping, starving, breaching or outpacing, judged from the losing side. A massive star uses up the gas that would-be stars need; nothing aimed, and the small star is starved. Defend is tested as Preserve: holding a persisting state against disturbance. Aim, target, intelligence and "deliberately" are value clauses, and they are dropped exactly as Barrier's "protected asset" became "any persisting state". Compete is usually a triangle, two takers on one pool, and each triangle is walked as a system of its own; there are 223.
What the record shows
1. Below life, every Function at work is at the done level
Of 1,042 held-level rows, none is on rungs 0 to 5. The six rungs of physics and chemistry hold done-level rows (115 of them) and inheritances (334). At rung 6 the held level begins, averaging 6.3 Functions per system, and climbs to the mid-teens from rung 7. This is the fold, visible in the statuses. The 2025 paper assumed the four families were universal requirements of evolving systems; the record says they are universal requirements of living ones, and that before life they exist at the done level, waiting to be held.
2. The families arrive in order
| Family | Done level, rungs 0–5 | Inherited, rungs 0–5 | Earliest done-level row and where |
|---|---|---|---|
| Create | 79 | 255 | rung 0: Generation, Templating and Symbiosis at the Big Bang |
| Defend | 20 | 41 | rung 0: Enclave; rung 2: Barrier |
| Attack | 16 | 34 | rung 2: Neutralise Defences; rung 4: Overcome Barrier, Resource Denial |
| Explore | 0 | 4 | none below life; Sensing's first done-level row is at rung 6 |
Below the fold, the four families are not equally present. Explore has no done-level row anywhere in six rungs and 1,995 rows.
Create is the oldest family. Its roots have done-level rows at the universal floor: the Big Bang generates, templates and, in the record's reading, co-creates. Defend's done level follows at once as enclosure, and as a barrier by rung 2. Compete's first done-level rows are at rung 2 and rung 4. But nothing explores before life. In 1,995 Explore rows below the fold there is not one done-level row. Sensing, the root of the whole Explore tree, first appears at the done level at rung 6 and is expressed at rung 6. Whatever a star or a mineral surface does, it does not find out. Finding out is the first thing life adds, and from the skill tree it is also what unlocks everything above tier 1 of Explore, and through Explore most of Attack and Defend.
3. Expression by family, above the fold
| Band | Create | Explore | Defend | Attack | All expressed |
|---|---|---|---|---|---|
| Rungs 6–10 | 265 | 144 | 220 | 52 | 681 |
| Rungs 11–12 | 165 | 90 | 96 | 10 | 361 |
Expressed rows by family and band. Attack is 6% of all expressed rows; most competition is booked in the 223 competition systems rather than as a system's own Attack Function.
Create leads at every band, Defend second, Explore third, Attack a distant fourth. The 2025 paper gave the families equal weight; the record does not. Part of the gap is the reading rule: competition is booked where it happens, in the triangle, and judged from the losing side, so a predator's walk shows fewer Attack rows than its prey's valence shows losses. But part is the data. The most expressed Functions in the record are Interoperability (49 systems), Recombination (48), Generation (44), Scaffolding (42), Barrier (42), Quality Control (42), Templating (40), Encapsulation (38), Amplification (37) and Validation of Identity (34): seven Create, three Defend, no Explore, no Attack. Systems that last are built, bounded and checked before they are clever or aggressive.
4. Credit concentrates in first-doers
Of 934 inherited rows in walked systems, 389 point to the cell, 107 to the Big Bang, 40 to the star, 37 to colour confinement, 31 to the atomic nucleus and 31 to the central dogma. The most inherited single Function is Templating, with 118 rows, almost all from the Big Bang; then Recombination (82), Interoperability (53), Barrier (52) and Encapsulation (43). A few early owners carry most of what everything later leans on, and the ladder is built from a small number of first-doers.
5. Late arrivals, and three Functions the record has not found
Most Functions are first expressed at rung 6. A few wait: Mapping, Navigation, Frontier Outposts and Positional Identity until rung 7, the animals; Precision Attack and Hypothesis Portfolio until rung 8; Deterrence, Active Experimentation, Curiosity and Subversion until rung 9, the mammals; Deception, Cascade Induction, Divide & Isolate and Offline Simulation until rung 10, the primates; Battlespace Shaping until rung 11. These are the deep tiers of the tree, and they arrive when the tree below them is in place.
Three Functions have almost no rows at all (no rows means no walk has recorded them yet, not that they're absent). Speed and Reproductive Division of Labour are expressed by no walked system, done by plain physics in none and inherited by none. Bridging Supply Gaps is expressed once, at rung 11. The last two were added this month and the walks that would carry them, at rungs 6 to 9, were walked before they existed; the re-walks now under way will fill them or not. Speed is older and emptier, and is a candidate for the critic: either the walks have under-read it, or it is a performance dimension of other Functions rather than a job of its own, which is how a sibling candidate, "latency compression", was closed this month.
What counts as a Function
The list is a tested object now, and it has a gate. A candidate Function is recorded only where there is a trait that varies and that selection can act on. Physical processes, outcomes, performance dimensions and survivorship do not count, and the default is to fit a candidate into an existing Function. Every walk checks each trait against the list, and a role that will not fit becomes a candidate with its sightings.
In October 2026, 36 candidates were reviewed with an evolutionary biologist on the team. Fifteen were not Functions: breaking apart (a physical process), pipelining (an architecture), the commit threshold (a gate design), blind variation and selective retention (the process that generates Functions, not one of them), the ratchet (an outcome). Seventeen fitted existing Functions, eight after a widened wording: swamping into Redundancy, teaching into Social Learning, costly commitment signalling into Validation of Identity, extended apprenticeship into Risk-managed Exploration. Three became Functions: Reproductive Division of Labour (a part gives up reproducing so a related part does it for both: body and germ cells, worker castes), Bridging Supply Gaps (holding a stock or going dormant through lean times: fat, seed banks, hibernation), and Scheduled Build-and-Dismantle (building a part for a season or stage and then removing it: the placenta, antlers, deciduous leaves, and servers spun up on demand). They were found late because the walks only recently reached the bodies and life stages where they are most visible.
What changed since the 2025 paper
2026-11-07: Functions all the way down. A Function is the job itself, at every rung; each fold adds what comes with it: done (plain physics does the job), held (something keeps it going), selected (sorted by its results) and valued (it means something to its holder). Rows below life, once read as "shapes" or "bare forms", are the done level of the same Functions. The data did not change; the reading did.
| In the 2025 paper | In the record now |
|---|---|
| About 60 Functions, in four tables of analogies across humans, markets and cybersecurity | 65 Functions with definitions, selection pressures and a 166-edge skill tree, tested against 202 systems from the Big Bang up |
| The four families as universal requirements of evolving systems | Universal requirements at every rung; below life, at the done level. The fold is visible in the statuses |
| The families given equal weight | Create 0 → Defend 0 → Attack 2 → Explore 6 in order of first done-level row; Create leads expression at every band; Attack is 6% of expressed rows |
| Attack and Defend applied at every scale by analogy | Below the fold they read as Compete and Preserve, judged from the losing side, with value clauses dropped; competitions are systems |
| A list assembled from a spreadsheet of traits | A list with a gate (the selectable-trait rule), a candidate queue, and a review that added 3, routed 17 and closed 15 |
| Who does a job, by domain | Who did it first: ownership flows down to six early systems, and credit never flows upward |
| A functional-information measure borrowed from others | No borrowed measure. What is counted is rows by status, by family, by rung |
Open questions
We raise a structural question only when the data gives cause. Three things in the record are worth watching. Speed has no rows anywhere; the critic should decide whether it is under-read or is a dimension of other Functions rather than a Function. Attack's small share may be the reading rule or the data; the re-walks will say, as they reach the animals where predators and parasites live. And the tree's deepest tiers predict the rung order of first expression; so far they match, and the match is the kind of thing that could fail. The 2025 paper's prediction that a system missing any one family will be outcompeted stands as a prediction; the record can now test it by asking which walked systems lack a family's expression entirely, and what became of them.
How to read the data
Every Function row carries its status, its reason, a confidence and its source. Rows booked at the done level or inherited name what they rest on. Candidates and their sightings are kept with the walks that raised them, and the review that ruled on them is in the record. The list of Functions, their definitions and the skill tree are the framework's own analysis and are labelled as such; the rows are readings of sources, and the sources are what to check.