Skip to the page
Rendered by flife · not drawn
Tick
2,400
Sim time
+0.0 s
Alive
262,144
View
whole

Artificial life · Rust · bit-exact

Evolution, bit for bit.

A real-time artificial-life laboratory. Every organism in this film carries its own heritable neural network; they feed, signal, move, split and die through one shared field of energy and chemistry, and nothing is scripted or trained. Pause it, inspect it, branch it, replay it: the same world, to the last bit, every time.

Open source on GitHub. Every figure on this page names where it comes from.

parameters per brain
9,988The default network, 104 → 64 → 32 → 36. No two organisms share one.
senses to decisions
104 → 36Field energy, 32 chemical channels and their gradients in; reproduction, emission and spending out.
CPU = Metal
Bit for bitControllers, field, sensing and uptake on the Mac's GPU reach the CPU reference's world exactly.
ticks, 100k organisms
≈10/sThe Desktop ecology world on 16 workers of an M5 Ultra. Measured, not a passed gate.
Chapter 01 · The world

Every speck is someone.

The film shows 262,144 organisms at once, and each is an individual: a body, an energy budget, offspring and its own neural network. Zoom in until you reach one, watch what it does in a tick, then run a small world of your own.

01.1Powers of ten

From the world to one brain.

Five steps from the film's whole world to the 9,988 numbers inside one of its organisms. Pick a step, or focus the picture and use the arrow keys.

The film's first frame: the whole torus at tick 2,400, 262,144 organisms coloured by founder lineage.
Rendered by flife · not drawn
×180 × 80 units
20 units
262,144 alivetick 2,400 · frame 0

01 / 05 · The whole torus

262,144 organisms, one world.

The film's first frame: the whole 80 × 80-unit torus at tick 2,400, after 75 simulated seconds of evolution. Every speck is one organism, with a body, a metabolism and its own neural network. flife render drew it from the committed world.

Alive
262,144
the render's count for frame 0
View
80 × 80 units
the whole torus; every edge wraps
Founders
10,954
each lineage wears one of 12 colours
Field
320 × 320
cells of energy and 32 chemical channels
  1. Step 1, The whole torus: 262,144 organisms, one world. Alive 262,144, the render's count for frame 0. View 80 × 80 units, the whole torus; every edge wraps. Founders 10,954, each lineage wears one of 12 colours. Field 320 × 320, cells of energy and 32 chemical channels.
  2. Step 2, A neighbourhood: Lineages become tissue. In view ≈ 8,000, derived: 262,144 alive over 6,400 square units, times 196. View 14 × 14 units, the middle of the social card's crop. Tick 3,872, frame 184 of 0–400. Camera zoom 3, flife render's record for frame 184.
  3. Step 3, One organism: One body, one field cell wide. Body 0.25 units, the preset's [contact] diameter. Field cell 0.25 units, 80 units ÷ 320 cells. This one (29.3, 70.4), picked on the card: a bright body with space around it. In view ≈ 370, derived the same way, for 3 × 3 units.
  4. Step 4, Its network: Its own network, every tick. In 104, energy, 32 chemistry channels, their x and y gradients, itself. Out 36, split rate, child share, emission gate, direction ×32, spending. Evaluated every tick, [controller] evaluation_interval_ticks = 1. Stored as f16, [controller] canonical_weight_dtype.
  5. Step 5, Its numbers: 9,988 numbers of its own. Parameters 9,988, 6,720 + 2,080 + 1,188, layer by layer, in the founders' shape. Shared with no one, each organism carries its own (INV-13). A child differs in ≈ 20, derived: 0.002 × 9,988 ([mutation]). Training none, [controller] autograd = false.
Steps 1–3 are flife's pixels, from lib/media.json: the poster (frame 0, tick 2,400, camera zoom 1, the whole torus) and the social card (frame 184, tick 3,872, camera zoom 3 centred at (28.125, 67.813): the 1080² frame scaled to 1,200 wide and cropped to 1,200 × 630, whose middle 14 × 14 units are shown). Between steps 1 and 2 the picture changes frame as well as zoom. The organism at (29.30, 70.39) was picked on the card as a bright body with space around it. Counts in view are the frame's live count spread evenly over the torus, an average, not a count. Steps 4–5 draw presets/confluent-world.toml's [controller], [mutation] and [architecture_mutation] and the sensor and action ABIs; their weights are illustrative, from a fixed seed, and are no organism's genome.
01.2One organism, one tick

A life in nine steps.

Each organism is a small neural network with a body, a metabolism and offspring, living in one shared field of energy and chemistry. Nothing is scripted and nothing is trained: whatever lineages survive, they survive by these rules alone.
FieldEcologyV1 · one organism, one tick at a time

Sense, decide, feed, move, emit, divide, die.

Diagram · not simulation output
phase 2SourcesFieldSenseDecideFeedMoveEmitDivideDie

01 · the shared field

Sources

Seeded Gaussian patches drift across the torus on smooth paths. Each injects a fixed power per simulated second, with a fixed chemical direction. Apart from a researcher's brush, it is the only food that enters the world.

The ring is the field. In the baseline model an organism never touches another directly: no collisions, no attraction tables, no predation. Everything one does to another, it does by eating from the field or writing into it.

README “How the world works” and “One tick”; plan §4–§8. Steps in the order the ordered tick runs them.
01.3A toy world, live

Run a small world yourself.

Before the real rules, a toy written for this page: about a thousand organisms, each with its own tiny network, under four of FrankenLife's rules in miniature. Select one to watch its brain, brush energy into the field, then replay from the seed and land on the same digest.
A toy world · 96 × 96 cells · 100 weights an organism

Run a world in your browser.

A toy in your browser · not FrankenLife

Every dot is an organism: a body with an energy store and its own network of 100 weights, which no other organism shares. Each tick it senses the field around it, decides where to move, how much to eat and whether to divide, and pays for all of it. Food enters only at four drifting sources, and wherever you brush. A child inherits a mutated copy of its parent's weights; nothing is trained.

tick 00 alive
Tap an organism

Lineages over time. Each column, the living by founder lineage; 12 of 12 lineages left.

last 1,920 ticks

births
—
deaths
—
divisions blocked
—

World digest

—

Rebuilds this world from seed 3 and runs it to the tick you are on, then compares digests.

Rules you can flip

Brains

Children

One shared runs the first founder's network in every body, the shortcut FrankenLife forbids. Fresh weights is one of its controls: a child draws new weights instead of inheriting. A switch restarts the world from tick 0 with the same seed, so you compare one changed rule; what the toy then does is not evidence about FrankenLife.

Tap an organism to read its own network: what it senses, its eight hidden units, and what it decides. Its kin stay lit; the rest of the world dims.

The ledger, this tick

exact, every tick
injected, ever
—
in the field
—
in bodies
—
dissipated, ever
—

The energy in the world now

last tick: +0 in from the sources, −0 out as heat

Upkeep, motion, births and the field's slow loss are dissipated; a death returns the body's energy to its cell. Nothing tops anything up.

A toy written in TypeScript for this page: not FrankenLife's code, and nothing it does is a FrankenLife result. It borrows four rules in miniature: uptake in two passes, in whole quanta by largest remainder, ties to the lower identity (plan §5.2, fixture C.1); a ledger that closes exactly on every tick (INV-03); birth as a transaction, cost paid and the rest split (C.5); and each organism its own weights, changed only by inherited mutation (INV-13). It leaves out chemistry and its 32 channels, emission, the sensing pyramid, the 9,988-parameter networks, the GPU kernels and the ordered ten-phase tick with its candidate world. Founders start from one forager's genome with every weight moved at random. Its step uses only +, −, ×, ÷, rounding down and 32-bit integer arithmetic, which every browser computes to the same bit (never exp, tanh or sin, whose last bit can differ), so a seed reaches the same digest anywhere. The real rules: check them by hand, the real networks, one real tick.
Chapter 02 · Why it is different

Beautiful is easy. Earned is hard.

Artificial-life imagery is easy to fake: one network copied into every body, energy from nowhere, a quiet cull, a renderer doing the work. FrankenLife forbids each shortcut, and its controls show the structure going when a cause is taken away.

02.1Four shortcuts

Each one, forbidden.

Each of these shortcuts can produce a gorgeous video of nothing evolving. FrankenLife's rules forbid all four, and its tests hold it to them.
  1. 01 · the shortcut

    One network, copied

    Hundreds of thousands of organisms that are really one shared policy: cheap, fast and smooth.

    The rule

    Every organism carries its own weights. Organisms with the same topology are batched to share a kernel and a memory layout, never a parameter, and a rare topology gets a correct fallback, never a different controller.

    INV-13plan §6.5

  2. 02 · the shortcut

    Energy from nowhere

    A world that thrives on food nobody paid for, or an overspent cell quietly zeroed.

    The rule

    An energy ledger closes on every committed tick: total energy changes only by injection, removal and dissipation, within a declared residual. No correction may add energy to make it balance.

    INV-03plan §4.4

  3. 03 · the shortcut

    A quiet cull

    Ugly regions removed, an extinct world reseeded, an aesthetic score steering selection.

    The rule

    No hidden culling, no reseeding after extinction and no aesthetic fitness inside the world. A capacity limit is explicit and recorded, and failed and extinct runs are kept.

    INV-12plan §7.5, §15.5

  4. 04 · the shortcut

    A renderer with no brains

    A smooth frame rate from a world that is not thinking, or one that skips ticks under load.

    The rule

    The Studio reports committed ticks per second and simulated time apart from frames per second. The engine never skips ticks, enlarges Δt or lowers neural cadence under load while keeping the model's name.

    plan §12.2§22.5

What already exists
  • Neural cellular automata learn one shared local rule.
  • Lenia, Flow-Lenia, Particle Lenia show what continuous local rules, mass conservation and interactive exploration can do.
  • ScriptBots-style simulators give agents direct interactions and pluggable brains.
  • Neural-Kimchi-4B shows the target look at the target scale.
What FrankenLife composes
  1. 1An independent heritable brain for every organism. Weights start random and change only by inherited mutation. Nothing is trained by gradient.
  2. 2Field-only interaction under an energy ledger. Organisms touch each other only by eating and writing the shared field, and the books balance every tick.
  3. 3A deterministic tick that survives cancellation and crashes. Checkpoint, kill, resume and branch: the same commands reach the same world, bit for bit.
  4. 4Live, inspectable, recordable animation of that same state. FrankenManim draws committed epochs. Watching, zooming or recording never changes the world.

Target, not a result: all four at the scale of the reference screenshots, about 400,000 organisms and four billion independent parameters, where memory bandwidth rather than arithmetic is the limit. Why that is hard.

What it is not
  • Not a port of Neural-Kimchi-4B, and no claim of numerical parity with it.
  • Not a decorative particle animation: no shapes are painted in.
  • Not a shared-policy, distilled or low-rank shortcut presented as independent networks.
  • Not trained by gradient, and not scored by any aesthetic fitness function inside the world.
  • Not a claim about life: “species”, “colony” and “cooperation” are analysis labels with operational criteria.
02.2The model and its controls

Remove a cause, lose the structure.

The film's world, run again twice with one rule changed. Without emission nothing ever leaves an organism, and the tissue becomes a haze; with fresh offspring instead of inherited ones, the world dwindles to scatter. Same preset, seed, founders and food.
Plan §5.4 · the model and its two controls · seed 42

Three worlds, one rule apart.

Rendered by flife · not drawn

One preset, one seed, the same 10,954 founders and the same food, run 2,400 ticks three times. Each control changes one rule and nothing else. flife drew all three the same way: the whole 80-unit torus, each organism a lit sphere in its founder's colour. The tissue is not painted on: take away the networks' decisions, or inheritance, and none forms.

Tap a world to place a loupe; swipe to the next, and it stays on the same place.

  1. The film's world, drawn by flife at tick 2,400: lineage domains of labyrinth, foam and packed spheres.1/3tick 2,400262,144 bodies

    Nothing removed

    The film's world

    Every rule on: each organism's own network chooses its actions from what it senses, and a child inherits its parent's genome, mutated.

    lineage domains of labyrinth, foam and packed spheres · the record, by eye

    Alive
    262,144
    the ceiling
    Distance
    0.74
    to the references; lower is nearer
  2. Decisions removed, drawn by flife at tick 2,400: a uniform mottled haze of small clumps: no domains, labyrinths or foam.2/3tick 2,400262,144 bodies

    Removed: the networks' decisions

    Decisions removed

    Every network still runs and is still charged for, but every organism's actions decode from an all-zero output, so no network decides anything.

    a uniform mottled haze of small clumps: no domains, labyrinths or foam · the record, by eye

    Alive
    262,144
    the ceiling
    Distance
    1.44
    to the references; lower is nearer
  3. Inheritance removed, drawn by flife at tick 2,400: sparse scatter.3/3tick 2,4001,222 bodies

    Removed: inheritance

    Inheritance removed

    Each newborn gets the genome a founder would have at its identity, unmutated, instead of its parent's mutated genome.

    sparse scatter · the record, by eye

    Alive
    1,222
    fewer than the 10,954 founders
    Distance
    5.00
    to the references; lower is nearer

Colour: founder lineage lit by energy: the lineage's colour (lineage-energy), lighter as the organism holds more of the energy it needs to divide, black with none, white with 1.8 times it. So the decision-free haze is pale because its bodies hold much of the energy they would need to divide, and the scatter is dark because there is almost nothing on the torus to draw.

The loupe spans 10.7 of the torus's 80 world units. A body, 0.25 units across, is 3.4 of flife's pixels.

What the record concludes

From the record

“The domains and their textures come from emission, chosen by the networks, and inheritance is necessary.”

And the decision-free haze fills the torus yet scores farther from the reference screenshots than the film's world, “so the preset's distance is not only a filled frame.” The showcase record.

What it is not. One seed and one tick per world: the controls drawn, not a statistical test. The distance is the showcase search's image statistic against the reference screenshots, lower being nearer; it chose the preset, so it is no evidence of evolution by itself. From 10,954 founders the film's world reaches its ceiling of 262,144; without inheritance, 1,222 are alive at the same tick. The seeded experiments with declared rules are FL-048.

Each still is flife render CONFIG --ticks 2400 --every 2400 --width 1080 --height 1080 --no-field --look sphere --color lineage-energy --out FRAMES on the film's preset (presets/confluent-world.toml) with only its model changed, coloured as the film is (lineage-energy, sphere look). At tick 2,400 each world's SHA-256 digest equals the one the record names for its trial (w80-o25, cw-chem, cw-fresh in experiments/showcase/confluent.json), and the first is the film's own first frame. Distances and the words in quotes: experiments/showcase/README.md, "### Controls". Rendered 2026-11-21 on Apple M5 Ultra, macOS 27.0.1, flife SHA-256 a159e6bfbf1e…; lib/controls.json records the rest.
02.3Provenance

What you are watching.

Not an illustration: flife render drew each frame from the committed world at its tick. The world is deterministic: the same commands reach the same digests, below, on any machine.
The world
Preset
presets/confluent-world.toml
Model
field-ecology-v1-contact
Seed
42
Founders
10,954 on a 320 × 320 field
Tick length
1/32 s of simulated time
Evolved first
2,400 ticks (75 simulated seconds)
Alive in the film
262,144
The film
Ticks
2,400 → 5,600, a frame every 8
Frames
401 at 30 fps, 1080 × 1080
Colour
founder lineage lit by energy: the lineage's colour (lineage-energy), lighter as the organism holds more of the energy it needs to divide, black with none, white with 1.8 times it
Field
not drawn, sphere look
Camera
10 keys, chosen from the run
First digest
3f279f51662ba456db526d2828b3faf19f93a7bd160deeb6f8e5f7f86f5134ed
Last digest
7fa0cc68e168a7277016aabd5a68b3322d4b074bdd2d99ede61228e48d3c9fc7
How this world was chosen

The film's world is not the baseline model. It runs field-ecology-v1-contact, a registered variant with one added rule: every organism is a body, here 0.25 world units across, and after each tick overlapping bodies push apart while bodies within 1.5 diameters pull together by the square of how alike their inherited chemistry is. The rule moves organisms, never energy. It was added because the reference screenshots' organisms look like packed spheres, which organisms that are points cannot be. Measured, the bodies stack two or three deep rather than pack: the pull, summed over every neighbour, outweighs the push.

Its numbers were chosen by searches outside the simulation, in several stages, against 14 statistics of the reference screenshots that inspired the project. The last stage kept the rules and set the scale. The field is four times finer than in the stage before, so a field cell holds about 2.6 organisms rather than about ten, nearer the references' one. Food is plentiful, so the population sits at its ceiling of 262,144, as the references' counters show theirs do at a million. Children land up to 2.5 world units from their parent, so colonies fill the 80-unit torus instead of leaving most of it black. It scores 0.74 on the statistics' scale, against 1.93 for the same rules at the coarser scale and 1.03 for the same ecology without bodies, scored the same way; the nearest any world of the last stage came was 0.73. Every trial was scored coloured by chemical affinity; this world was chosen by eye, coloured by lineage. The film colours each organism by its lineage and lights it by its own energy: black with none, white with nearly twice what it needs to divide, so pale patches are well fed and dark ones starving. Coloured that way, the scored frame comes to 0.72, against 0.90 coloured by lineage alone. The film shows the chosen world itself, from the tick where it was scored.

Nothing the search measured feeds back into a run. The organisms have no aesthetic fitness: they feed, emit, split and die by energy alone. The seed is the Desktop preset's 42, not a pick. On three seeds the search never saw, the world scores 0.81 to 1.57, against 1.92 to 2.68 for the coarser world on the same seeds.

Emission, chosen by the neural networks, makes the structure, and contact shapes it at the scale of bodies. Organisms return about half of what they eat to the field, in a chemical composition their networks choose and kin choose alike, so each lineage paints its surroundings with its own chemistry. Adhesion then sorts lineages by that chemistry, as real tissues sort by adhesion, into domains of labyrinths, foam and packed spheres. With every controller's decisions zeroed, nothing is emitted and the bodies settle into a uniform mottled haze, with no domains, labyrinths or foam. With every newborn's genome drawn fresh instead of inherited, the population falls from its ceiling to about 1,200.

The camera is chosen from the run, not by hand. The same ticks are first rendered small; each shot then eases into the region where the organisms changed most during it, and the film opens and closes on the whole torus. The camera changes no result: the preview and the film end on the same digest.

The search, every sample's score and the held-out seeds.

Make it yourself
# flife is `cargo run --release -p flife-cli --` in the repository
flife run presets/confluent-world.toml --ticks 2400 --run-dir RUN
flife render --run-dir RUN --ticks 3200 --every 8 --width 1080 --height 1080 --no-field --look sphere --color lineage-energy --camera CAMERA --out frames
# CAMERA: lib/media.json's render.camera keys, as {"keys": [...]}

Rendered at revision b4a6d62d77fa with controllers on cpu (frankentorch population_affine_f32_into, 24 lanes); no result depends on the route. The first frame shows 262,144 organisms and the last 262,144; the render's end record names the last digest above. The PNG frames are flife's; this page's video encodes them with flife's own BT.709 colour chain at CRF 24 for the web, and its SHA-256 is 65b03ad6b91921f30f5fabc8d9f02e929bb6b0ac36c3c095afba87ddff57e800.

Chapter 03 · How it works

The machinery.

Independent brains, an explicit model with one energy ledger, rules pinned by answers you can compute by hand, and a tick that is published whole or not at all.

03.1Independent controllers

Every organism, its own brain.

Each organism runs its own dense network on what it senses. No two share a parameter, nothing is trained by gradient, and the weights change only when a child inherits a mutated copy.
The controller · plan §6.1–§6.4

104 senses in, 36 decisions out.

Diagram · not simulation output
Hidden layers
qZ ×32∂x ×33∂y ×33self ×5λ splitβ sharegatedir ×32f spend104 in64 tanh32 tanh36 outlinear · σ once, in decoding

Senses · 104

Decisions · 36

Every input reaches every neuron of the first hidden layer, and every layer the next: a dense network with tanh hidden layers and a linear output. Point at a group, or tab to one, to see what it carries. The colours pair what it reads with what it writes: chemistry in, chemistry out; energy in, energy spent.

Its own parameters

104×64 + 64 + 64×32 + 32 + 32×36 + 36 = 9,988

104 → 64 → 32 → 36: 39,952 bytes of F32 weights and biases that belong to this organism alone. Offspring move between the five networks by declared maps; a widened copy computes exactly what its parent did (before the child's own mutations), while a change of depth claims no such thing (plan §6.4, fixture C.16).

No motor output

Nothing here steers. An organism moves by inherited responses to the same energy and chemical gradients, scaled by how much it was just granted to eat (plan §5.4). The network decides when to split, what to give the child, and what to write into the field.

Inputs and outputs in the code's order: crates/flife-neural/src/sensors.rs (energy-moment-gradient-104-v1) and actions.rs. The five networks: crates/flife-genome/src/catalog.rs (dense-104-36-catalog-v1, plan §6.4). Parameter counts include biases and are exact arithmetic; the drawn connections are a sample of the dense layers' full set.
Independent controllers · plan §6.5, INV-13

Same shape, never the same weights.

Diagram · not simulation output

Every organism carries its own parameters, inherited from its parent with a few mutations and shared with no one. Kin look alike, never identical; unrelated organisms look unrelated. Selection acts on exactly those differences.

Inheritance

Diagram · not simulation output

Parent

Its child

Unrelated

From the record

In the reference world's 1,024-tick smoke run, live children keep 99.8% of their parent's parameters bit for bit, while unrelated founders share 1.3%. Grandchildren carry 99.7% of their parent's mutations. Weights start random and change only by inherited mutation: nothing is trained by gradient.

The thumbnails are illustrative weights drawn from a fixed seed, not genomes from a run; each lineage keeps one topology here for clarity. The rule: plan §6.5 and INV-13; the batching: crates/flife-backend-metal and crates/flife-backend-cpu (per-topology parameter shards, FL-026). The heredity figures: README.md, the reference preset's 1,024-tick smoke run.
03.2FieldEcologyV1

The model, explicitly.

An original, explicit rule set: a periodic field of energy and chemistry, and organisms that can only touch each other through it.
World
A two-dimensional periodic torus. Organisms have continuous positions; the field is a regular grid.
Field
Each cell holds energy and a 32-component signed chemical moment. One positive five-point stencil diffuses both, conserving energy before declared decay.
Sources
Seeded Gaussian patches drift across the torus on smooth paths, each injecting a fixed power with a fixed chemical direction per simulated second.
Sensing
Energy, 32 composition values and their gradients at a heritable sensing scale, plus the organism's own energy, age, velocity and recent energy change: 104 inputs.
Brain
Dense tanh networks with a linear output layer. Offspring can move between catalog topologies of 1,164 to 29,316 parameters.
Decisions
36 outputs: a reproduction hazard, a child's energy share, an emission gate, a 32-dimensional emission direction and a spending rate. No motor output.
Uptake
A two-pass demand and grant step. A contended cell grants in whole quanta, so no cell is overdrawn and thread order never decides who eats.
Chemistry
A heritable affinity decides whether each granted packet feeds or poisons. Toxicity is clipped to the energy the organism actually holds.
Motion
Heritable responses to gradients give an acceleration; drag is integrated exactly, and movement costs energy in proportion to speed squared.
Emission
An organism can pay energy to deposit chemistry at an efficiency of at most one. Deposits are first sensed on the next tick, so no loop can amplify energy.
Reproduction
A split makes a mutated child. Identity, genome, slot and energy transfer commit together, or nothing happens.
Death
Zero energy after any stage is death, checked after every stage. Part of the residue returns to the field.
The ledger

𝓔ₙ₊₁ − 𝓔ₙ = injection − removal − dissipation + numerical residual

𝓔 is the energy in every organism plus the energy in every cell. Emission, absorption and a child's endowment only move it. No correction may add energy to make the equation balance.

03.3Appendix C, playable

Check it by hand.

Each rule has a plausible-looking wrong version, and the plan pins each one with a known answer small enough to compute by hand. Change the numbers: the page recomputes the rule beside its wrong twin.
Appendix C · 7 of the plan's 17 known-answer fixtures

Every rule has a plausible wrong version. Try both.

The rule, computed here

Each tab is one of the plan's hand-computable fixtures, recomputed here in your browser, in double precision, beside the wrong version it exists to catch. Move the numbers and watch the two part ways; reset to the fixture and the page checks its answer against the one the plan states. This is the arithmetic, not FrankenLife's code: its Rust kernels are tested against the same answers, and each tab names the test.

C.1A crowded cell: two passes, not first come, first served

The ruleSum every request into the cell's demand A, then grant each organism α q with α = min(1, e/A), in whole quanta. No order of threads can change who eats.

Plausible, wrongOne pass: each thread subtracts its request and clamps the cell at zero. Who eats depends on who ran first, and racing threads can spend the same energy twice.

The rule · any thread order

cellA8→4B12→6

Wrong · one pass, A → B

cellA8→8B12→2

Rule: grants

4 · 6

α = 0.5

Wrong: grants

8 · 2

in order A → B

Rule: ledger

−10 = 3.5 + 6.5

field loss = useful gains + dissipation

Wrong: energy

10 of 10

granted from the cell

Quanta here are whole units; in FrankenLife a quantum is the cell's F32 unit in the last place. Leftover quanta go to the largest remainders, ties to the lower identity (try the three-way tie). η for A and B are C.1's; C's and D's are this page's.

Matches fixture C.1: grants 4 and 6, gains 2 and 1.5, dissipation 6.5

Plan §5.2 · fixture C.1 · INV-04 · implemented in crates/flife-field/src/uptake.rs · checked by crates/flife-field/tests/uptake_ledger.rs::c1_contended_uptake_balances_exactly
03.4Ten phases

One tick.

Each tick is prepared as a candidate against the committed world and published whole, or not at all. Cancel it part way and nothing has changed.
Plan §8.1–§8.3 · fixture C.8

Published whole, or not at all.

Diagram · not simulation output

Committed world

tick 1,024

epoch 1,024

  • population
  • field
  • ledger
  • genome quota
  • identities
  • source paths

Authoritative. Observers, the inspector and checkpoints read only this.

Candidate

tick 1,025

none yet

  • spare field buffer
  • reusable scratch
  • its own copies

The only things a candidate writes. None of them is the committed world.

The committed world is tick 1,024. Nothing is being prepared.

  1. 01Admit commands at their recorded tick boundaryseals the commands due at this boundary
  2. 02Transport the field and inject sources into one immutable observation fieldwrites the spare field buffer: the tick's one observation field
  3. 03Build the sensing pyramid and spatial binswrites reusable scratch: the pyramid and spatial bins
  4. 04Sense and evaluate due controllers, all from the same epochwrites scratch and its own copies: observations, held outputs
  5. 05Resolve uptake demand into cell grantswrites reusable scratch: requests and cell grants
  6. 06Settle useful gain, toxicity and maintenancewrites its own copies: each organism's budget
  7. 07Propose affordable motion, then emission, then reproductionwrites its own copies: motion, emission and split proposals
  8. 08Resolve capacity, construct births and collect deaths, in stable identity orderstages births, their genomes held by the candidate; collects deaths
  9. 09Reduce deposits and recycling into the next fieldwrites the spare field buffer: grants debited, deposits and recycling added
  10. 10Validate the invariants and publish state, events and accounting togetherchecks INV-01–03, 06 and 07, then publishes everything at once
crates/flife-engine/src/world.rs: World::advance prepares a Candidate against the committed world; Candidate::publish commits the population as one checked transaction, then the rest, which cannot fail. Dropping or cancelling a candidate changes nothing. The phase where this diagram raises its fault is its own choice; any phase can.
Chapter 04 · What it can show

Claims, kept apart.

Replays that land on the same digest, experiments whose rules were written before they ran, a scale argued in arithmetic, and a status that names what is not done.

04.1Checkpoints · resume · branches

Rewind, branch, compare.

A kept run checkpoints as it goes. Kill it and resume, and it reaches the uninterrupted world. Branch it, and a control reaches the parent's world exactly, while an intervention differs only from its brush onwards. That is what makes an experiment a comparison.
Run directories · FL-037, FL-039

Same commands, same world. Change one, and it differs only after.

Diagram · not simulation output

The run

Killed, resumed

Control branch

Intervention

  • checkpoint
  • killed
  • resumed from the newest checkpoint
  • branched
  • brush applies

The worlds at tick 3,584 (toy digests):

The run1b0e8e3breference
Killed, resumed1b0e8e3b the run's world
Control branch1b0e8e3b the run's world
Intervention527c75c6 differs

Killed part way, a run resumes from its newest checkpoint that restores to exactly the uninterrupted run's world and final checkpoint bytes, across controller phases, cooldowns, newborns and commands still in the future. A control branch reaches the parent's world; its intervention sibling matches it until the brush applies at tick 2,304's boundary, and differs from tick 2,305 on. A branch links the parent's verified chunk files instead of copying them.

flife run CONFIG --ticks 4096 --run-dir run
flife run CONFIG --ticks 4096 --run-dir killed
# SIGKILLed at tick 3200; it resumes from its checkpoint at 3072
flife resume killed --until 4096
flife branch run control --at 1536
flife branch run brushed --at 1536 --inject 2304:64,64:256
flife resume control --until 4096
flife resume brushed --until 4096
README.md (FL-037, FL-039); crates/flife-cli/tests/resume.rs (a_killed_write_ahead_run_resumes_to_the_uninterrupted_state) and tests/branch.rs (a_control_branch_continues_as_its_parent_while_an_intervention_branch_diverges). The digests here are toy 32-bit hashes of each lane's world, not flife's SHA-256 world digests; checkpoints every 512 ticks are this diagram's, a preset sets its own interval.
We ran itObserved · 2026-11-21

The reference preset kept for 512 ticks and stepped to 576, then branched twice from that checkpoint: a control, and a branch with one energy brush (256 energy at cell (64, 64), tick 640). All three resumed to tick 1,024:

  • The parent, resumed2ade32c2f3e74f6f54181019774181e2aa0857bd79bd6c5a6f774f85e46a8183the parent's world
  • Control branch2ade32c2f3e74f6f54181019774181e2aa0857bd79bd6c5a6f774f85e46a8183the parent's world
  • Brushed branch0d910c5d40a76a1511bcef2eafccdf7bb748a60460fb81519aa4d91f785d8cc7a different world

Each line is the run's end record's SHA-256 world digest, on an Apple M5 Ultra, macOS 27.0.1, with a release flife (SHA-256 a159e6bfbf1e8b8e…). The commands are under Run it yourself; the repository's own tests make the same comparison on every change.

04.2FL-048 · declared before running

What the experiments found.

Each hypothesis had its rule written down before the runs, discovery and held-out seeds kept apart, and failed runs kept. One control came out strongly; two hypotheses are not supported, and the rest are stated as narrowly as the data allow. The full record.
The inheritance control · the strongest result

Selection acts on inherited variation.

Give every newborn the genome a founder would have drawn at its identity, with the default topology and no mutation, instead of its parent's mutated copy; it keeps its parent's lineage. The world then holds about half the organisms, and far more founder lineages survive side by side, because no lineage carries a heritable advantage.

8 of 8 discovery seeds · 8 of 8 held out

organisms alive at the end
−2,407
discovery; −2,524 held out
founder lineages alive
+113.8
discovery; +110.9 held out
exact sign-flip p
0.0078
each, 8 of 8 seeds in both splits
Fewer alive without inheritance
discovery8/8
held out8/8
More founder lineages alive without inheritance
discovery8/8
held out8/8

Against the control, paired by seed: reference-small-world, 4,096 founders, 64 simulated seconds, 8 discovery and 8 held-out seeds, paired against the control.

  1. H1Met · a low bar

    Persistent heterogeneous colonies

    Every run that reached tick 8,192 ended with clustered, spatially autocorrelated colonies: 32 of 32 in E2, 16 of 16 E1 controls, and 112 of 112 runs across every ablation. A low bar: it is clustering at one sample, not colonies persisting through time, and it holds without controllers, emission or inheritance.

    E1 control meets the colony criterion
    discovery8/8
    held out8/8
  2. H2Not supported

    Controllers add behaviour beyond chemistry

    Zeroing the controllers' decisions leaves more organisms alive, but that ablation also stops emission, and removing emission alone has the same effect. Compared with no-emission (post hoc), the decisions' own effect is inconsistent on every measure.

    Chemistry-only: more alive than control
    discovery8/8
    held out7/8
    Post hoc, chemistry-only against no-emission: more alive
    discovery5/8
    held out4/8
  3. H3Met narrowly

    Paid emission supports niche construction

    Emission evens the field (Gini about 1.5% lower) but leaves how organisms cluster unchanged, and costs population: environment modification, not demonstrated niche construction.

    Field Gini lower with emission
    discovery6/8
    held out7/8
    Organism Morisita higher without emission
    discovery4/8
    held out4/8
  4. H4Rule met · not established

    Visual structures are functional

    Removing a colony's boundary changed the field over the colony more than removing its interior, as declared. Post hoc, 85–98% of that effect is the emptied disk refilling. A functional boundary is not established.

    Field over the colony: boundary effect largest
    discovery3/4
    held out3/4
    Survival: boundary-specific
    discovery1/4
    held out1/4
  5. H5Not supported

    Architecture variation affects ecology

    Fixing every organism's topology raised the final population on the discovery seeds and lowered it on the held-out seeds: the effect reverses.

    Fixed topology: more alive than control
    discovery7/8
    held out2/8

Each dot is one seed on which the measure went the stated way. most seeds of both splits the splits disagree or split evenly most seeds of neither

04.3Four claims

Kept apart.

A beautiful image establishes only the visual claim. A speed measured on named hardware is a separate kind of result, and evidence for none of the four.
  1. 01Rules

    The simulation obeys its own rules

    Invariants INV-01–INV-15 and the energy ledger; the field, ledger and population checks run on every tick before it is published

  2. 02Implementation

    The code reproduces those rules

    Hand-computed fixtures, a separate scalar oracle, and CPU/Metal parity bit for bit

  3. 03Visual

    The images have the intended character

    Human review against the references, raw diagnostic views beside the aesthetic ones (G5)

  4. 04Ecological

    Populations show measurable evolutionary phenomena

    Ablations, seed ensembles and held-out seeds, failed and extinct runs kept (G9)

04.4Why the scale is hard

Four billion independent parameters.

No two organisms share weights, so a full sweep cannot reuse them: it streams the whole bank from memory. At the target scale, bandwidth, not arithmetic, is the limit, and only a real measurement with the real layout can justify a real-time claim.
Why the scale is hard · README, plan §19

Every sweep streams the whole bank.

Derived · arithmetic, not a benchmark

Canonical weights

Compact F16 genomes are a distinct quantized evolutionary model, not a hidden storage trick (plan §10.3).

Independent parameters

3.995 billion

no two organisms share one

Weight bank

7.44 GiB

F16, before any other state

Multiply-accumulates

120 billion/s

arithmetic is not the limit

Weight stream

240 GB/s

every sweep reads every weight

Weight bytes a second, logarithmic

400,000 × 9,988 parameters

Sweeps/sMAC/sF16 streamF32 stream
1559.9 billion120 GB/s240 GB/s
30120 billion240 GB/s479 GB/s
60240 billion479 GB/s959 GB/s
128511 billion1.02 TB/s2.05 TB/s

At 400,000 × 9,988 these are the README's rows. A headline FLOP/s figure cannot justify a real-time claim here; only sustained bandwidth with the real independent-weight layout can.

And they must eat

Derived · arithmetic, not a benchmark

Mandatory cost3,423 energy/s

Starter's eight sources640 energy/s

At unit sensing scale each organism of this size pays at least 0.005 + 0.001 + 9,988 × 128 × 2·10⁻⁹ = 0.008557 energy a simulated second, before moving, emitting or dividing. With perfect uptake the starter's sources could keep at most about 74,793 alive, so every profile declares its own environment.

Measured, not derived

From the record

On an Apple M5 Ultra (2026-11-10, with other GPU work on the device), FrankenLife's resident Metal controller kernel streamed F32 weights at 426–748 GB/s in its fastest rounds, 0.44–0.69 of a plain read of the same bytes, every output identical to the CPU reference. Against a four-billion-parameter F32 bank that is 27–47 full sweeps a second, for the controllers alone; sensing, the field, births and deaths come on top.

The scale gate G7 asks for at least 380,000 live organisms averaging 9,000–11,000 parameters, at least 30 full-evaluation ticks and 30 frames a second, for ten minutes on a named machine. It has not run.

Arithmetic from README.md "Why the scale is hard" and plan §19.2 and §19.5, not a benchmark. One multiply-accumulate per parameter per evaluation, biases included, as the README counts them; the README rounds 400,000 × 9,988 to 4.0 billion parameters, and these figures use the exact 3,995,200,000. The measured band: crates/flife-backend-metal/CONTRACT.md, resident bandwidth (FL-030).
04.5Gates G0–G10

Honest status.

The foundations are built and tested; the visual, full-scale rate, endurance and export gates are not. The 400,000-organism world at 30 ticks a second is a target, not a result. The README keeps the full account.
Milestones: evidence gates, not dates
  1. M0Compatible foundation. Runtime, typed arrays, an independent-network batch and a FrankenManim frame in one program.G0
  2. M1First living world. 4,096 organisms with independent genomes, the ledger, reproduction, mutation and death, live.G1G2
  3. M2Usable laboratory. A polished 25k profile, inspector, interventions, checkpoint, resume, branches, recording.G3G5G6
  4. M3Qualified acceleration. CPU/Metal parity, persistent buffers, fused population inference, variable topology.G4
  5. M4Screenshot-scale world. 400k independent networks near 10k parameters on a 1024² field, interactive, honest clocks.G7G8
  6. M5Scientific and visual release. A seed gallery, ablations, experiment reports, reliable Cinema export.G9G10
  7. M6Additional model families. Recurrent controllers, multi-carrier chemistry or physical boundaries, each with a demonstrated need.per extension
  • G0

    The stack composes

    One lockset; every probe matches its oracle on macOS arm64 and Linux x86-64.

    Built and tested
  • G1

    Hand fixtures and arithmetic

    Fixtures pinned bit for bit; world digests identical on both hosts.

    Built and tested
  • G2

    A first world lives, renders and pauses

    The 4,096-organism preset breeds, mutates and dies, live in a native Metal window.

    Built and tested
  • G3

    Checkpoint replay and crash recovery

    A run killed at any step of a checkpoint or command resumes to the uninterrupted world.

    Built and tested
  • G4

    CPU and Metal parity

    Controllers, field, sensing and uptake on Metal, each bit for bit the CPU reference's.

    Built and tested
  • G5

    Visual review

    Needs a human review against the references. Not yet run.

    Not yet
  • G6

    30 fps with p95 input latency ≤ 100 ms

    Offscreen only: 30.00 fps at 1080p (p95 51.6 ms) and 29.79 at 4K (p95 59.9 ms). Not yet measured in a window.

    Partial
  • G7

    400,000 organisms at 30 ticks/s

    Not reached. The 100,000-organism Desktop world runs at about 10 ticks/s on an M5 Ultra.

    Not yet
  • G8

    Eight-hour endurance

    Not run.

    Not yet
  • G9

    Research claims

    Five declared hypotheses tested with held-out seeds (below); the release's claims are not yet made.

    Partial
  • G10

    Export alignment

    Not run.

    Not yet
Chapter 05 · Use it

Your laboratory.

Eight worlds to open, a native Studio to watch and touch them, and a command line that records everything it does.

05.1Presets

Eight worlds to open.

Three are the operating profiles, each sized by a declared experiment. Five were chosen by an outer search against image statistics of the reference screenshots and then by eye; the organisms inside them have no aesthetic fitness, and the search states how.
  • reference-small-worldReference profile · FL-015

    field-ecology-v1

    4,096 foundersceiling 16,384

    128² field · 128-unit torus · Δt 1/128 s

    Appendix A at the Reference scale, plus a declared viability tuning. Unfed founders starve and fed lineages breed; the smoke test pins its digest after 1,024 ticks.

  • interactive-worldInteractive profile · FL-034 E1

    field-ecology-v1

    25,000 foundersceiling 40,000

    256² field · 256-unit torus · Δt 1/128 s

    Source power chosen by a declared experiment. Populations settle at about 22,000 to 25,000 after an early bottleneck.

  • desktop-worldDesktop profile · FL-034 E2

    field-ecology-v1

    100,000 foundersceiling 160,000

    512² field · 512-unit torus · Δt 1/128 s

    F16 weights. All 8 of E2's runs settled at 91,664 to 99,361 alive. About 10 ticks a second on 16 workers of an M5 Ultra: a world to study, not yet one to watch at frame rate.

  • showcase-worldOuter search · first choice

    field-ecology-v1

    43,816 foundersceiling 240,000

    256² field · 256-unit torus · Δt 1/32 s

    Membranes, rings and crescents, which inherited chemistry makes without the networks.

  • colony-worldOuter search · stage W

    field-ecology-v1

    43,816 foundersceiling 177,536

    256² field · 256-unit torus · Δt 1/32 s

    Discrete colonies on a black field, held at a binding ceiling by refuse-excess-births-v1, the plan's declared exhibition policy.

  • tissue-worldOuter search · stage W

    field-ecology-v1

    10,954 foundersceiling 177,536

    128² field · 128-unit torus · Δt 1/32 s

    Lineages paint their surroundings with their own emitted chemistry and merge into a nearly continuous tissue of membranes, voids and mesh walls, 110,000–170,000 strong.

  • contact-worldOuter search · stage C

    field-ecology-v1-contact

    10,954 foundersceiling 177,536

    128² field · 128-unit torus · Δt 1/32 s

    tissue-world with bodies 0.45 units across that push apart and pull alike chemistry together: single-file membranes, rings and foam.

  • confluent-worldOuter search · stage E · the film

    field-ecology-v1-contact

    10,954 foundersceiling 262,144

    320² field · 80-unit torus · Δt 1/32 s

    contact-world on a field four times finer, fed to its ceiling, with children landing up to 2.5 units away: saturated domains of labyrinth, foam and packed spheres.

05.2Studio and laboratory

Watch it, touch it, measure it.

The native Studio is the microscope: run, pause, step, zoom, select, inspect, brush and record, while the session commits every tick. The flife command line is the laboratory around it.
flife-studio · macOS

A microscope with a keyboard.

Diagram · not simulation output
Space

Run

Run or pause the world. The Studio opens running unless started with --paused.

The pointer

Scroll
Zoom about the pointer; up zooms in.
Drag
Pan, across the torus's seams; with removal or measuring armed, draw the disk.
Click
Select the organism under the pointer: its energy, age, motion, inputs and hidden activations, read back from the session. With a brush armed, brush there.

The Studio is macOS, drawn through Metal. Look, field, camera and selection keys change only the picture: nine styles of one 1,200-tick render all end at the same digest. Brushes are commands: each one is applied by the session at its next open boundary and reported with the tick it applied at and the energy it moved.

crates/flife-studio/src/controls.rs (keys by virtual key code, so any layout) and CONTRACT.md; brush amounts from src/brush.rs; README.md. Keys drawn dim do nothing in the Studio.
The flife command line
validate
Range-check a configuration and report its memory and worst admitted allocation
capabilities
Run each controller route against the reference: what this machine can run
run
A headless run; with --run-dir, durable, journaled and checkpointed
resume
Continue a run directory from its newest checkpoint that restores
step
Advance a run past its newest durable tick, its journal replayed first
branch
Control and intervention branches from one checkpoint
inspect
A run's committed world at a checkpoint: every organism and cell, weights included
experiment
Seed ensembles, sweeps and ablations against a control, every run cataloged
analyze
Experiments' effects pooled by condition, refused unless comparable
discover
A pinned discovery: reference, selection, thumbnail, world export, local package
transplant
Graft a discovered package into a host world, as a new run that resumes
render
Frames and MP4 through FrankenManim, from a configuration or a kept run

Every command prints versioned NDJSON with stable identifiers and typed errors: a missing accelerator, an exhausted budget, a schema mismatch, an incompatible checkpoint and a numerical fault are each their own kind.

05.4Laboratory mode

Run it yourself.

Every command prints versioned NDJSON. Start with what your machine can run, then the 4,096-organism reference world, then branch it and compare.

# What this machine and build can run, by running it

cargo run --release -p flife-cli -- capabilities

# 1,024 ticks of the 4,096-organism reference world, headless

cargo run --release -p flife-cli -- run presets/reference-small-world.toml --ticks 1024 --every 128 --metrics-every 128

# The same world as PNG frames and an MP4

cargo run --release -p flife-cli -- render presets/reference-small-world.toml --ticks 2560 --every 4 --width 1080 --height 1080 --out frames --video world.mp4

# Live in the native Studio (macOS)

cargo run --release -p flife-studio -- presets/reference-small-world.toml --width 960 --height 540

# The film's world in the Studio, from its 10,954 founders to the 262,144 ceiling, drawn as the film draws it

cargo run --release -p flife-studio -- presets/confluent-world.toml --color lineage-energy --look sphere --no-field

# A kept run, a step past its newest durable tick, a control branch and a branch with an energy brush

cargo run --release -p flife-cli -- run presets/reference-small-world.toml --ticks 512 --run-dir my-run
cargo run --release -p flife-cli -- step my-run --ticks 64
cargo run --release -p flife-cli -- branch my-run control
cargo run --release -p flife-cli -- branch my-run brushed --inject 640:64,64:256

# Resume all three to tick 1,024 and compare their last records' digests: the control's is the parent's

for run in my-run control brushed; do cargo run --release -q -p flife-cli -- resume $run --until 1024 | tail -1; done