By any account, E. coli is a very skilled self-replicator. Presumably, though, the first successful self-replicator was far less skilled: just barely able to persist, and only with substantial help from its environment. So asking about the origin of life means asking how bad something can be at self-replication and still persist with that help, and what it takes to keep persisting as the help is withdrawn.
When people hear "origin of life", they usually think of a story: a possible scenario, set in a warm pond or at a deep-sea vent. Those who know the field also know that the early Earth left no fossils of the first replicators, so any such story is at best an educated guess. Sceptics go further: Fred Hoyle compared the chance assembly of a cell to a tornado sweeping through a junkyard and assembling a Boeing 747. In those terms the problem cannot be solved. There is no way to check a story against a record that doesn't exist, and no way to argue with odds that are posed as one enormous draw.
We change the terms. Instead of proposing a scenario, we derive what must have been true however cells arose, using information and time as resources, and we read other people's measurements against those bounds.
The question turns into a short list of constraints on counts of parts over time, and information theory lets us state them quantitatively. The page starts from the structure of self-replication. Part 1 builds the units: possibilities and copies on one scale, then the environment on that same scale as a vector of resources it supplies for free. A lineage then traces a trajectory through a space of information and time, which we call the phase space of the living state. Laws and constants come next. Part 2 writes coming alive as constraints on that trajectory, in the order they must be met. The last section shows how close a first path now is, and what a wet lab could do with it.
Two observations come before any chemistry: self-replication is a matter of counts, and a replicator must eventually separate from the environment that helps it. Everything later on this page follows from them.
The junkyard-tornado argument prices a cell as one draw. All of its bits have to be found at once, against one budget.
Bits add along a path, because the log of a product is a sum. A step that is kept does not have to be found again, so each step is searched against its own budget. The bits measure cost, not improbability.
Staging works only if every landing holds. That moves the hard question from probability to rates: can the lineage pay its rent at every step?
Part 1 explains how each step on the staircase has two prices: a search cost, paid once to find it, and a maintenance cost, paid every generation to keep it.
Rent is what copying loses, measured in bits per copy: log₂ of copies made for each copy that works. Compare two replicators. E. coli is a modern bacterium. QT45 is a ribozyme, an RNA molecule that does the job of an enzyme; this one copies RNA. Reported in 2026 by Philipp Holliger's group at the MRC Laboratory of Molecular Biology, it is only 45 nucleotides long, and in ice it can make both its complementary strand and a copy of itself, slowly and at low yield. It is the closest thing yet to the kind of copier the first replicators needed.
| Replicator | Ratio of counts, per genome copy | Rent per copy |
|---|---|---|
| E. coli | about one new mutation per few hundred genome copies (Drake's rule: a nearly constant mutation rate per genome across microbes) | ≲ 0.005 bits |
| QT45 (an RNA copier) | about one working copy in 2,500, mostly from copies that don't finish | 11.3 bits |
The origin of life lies somewhere along that gap. The rest of this page asks where along it a lineage could first persist on its own, and what had to pay the difference. (The E. coli figure counts errors only, and counts every site as constrained, so it is an upper bound.)
Every bit in the ledger is log₂ of a ratio of two counts that could, in principle, be counted at the bench. Mass and energy have no axes of their own: mass limits how many copies can be made, and a free-energy difference buys discrimination, which is measured in bits. Two other kinds of constraint sit beside the bits and are marked as such: concentration inequalities (how much Mg²⁺ a membrane tolerates) and order constraints (which step must come first).
Every step toward a cell costs something twice: once to find it, and again on every copy to keep it.
Teal marks finding. Teal bits count possibilities: how much of sequence space a function must pin down. They are a stock, paid once.
Amber marks keeping. Amber quantities are bit rates: log₂ fold change per copy or per generation, so one bit per generation is one doubling per generation. A PCR user already thinks in these units. They are a flow, paid continuously, so they are never added to bits.
Each inequality behind a "Show the math" toggle gives a step on the staircase its price and its place in the order. Some steps can swap places; others cannot.
Physicists split the heat a driven system gives off into two parts. Excess heat is the cost of changing state, paid once, like cooling a warm fridge down. Housekeeping heat is the cost of holding a state that isn't equilibrium, paid continuously, like keeping the fridge cold with the door shut.
We make the same split for the information selection has to supply. Excess information is paid once, when a new variant takes over: that is finding. Housekeeping information is paid every generation, to replace what copying loses: that is keeping, and it equals the rent.
So finding and keeping are not the same cost, but they are two halves of one decomposition. That is what lets us build the whole staircase to a cell from only two resources, information and time: each step has a finding price in bits, paid once, and a keeping price in bits per generation, paid for as long as the step must hold.
Population geneticists will recognize the housekeeping term: it is exactly the genetic load of Haldane and Muller. We did not set out to rederive it. We arrived at it by a different route, from the thermodynamic split, and that route is what puts it in the same units, and on the same clock, as everything else on this page.
In nonequilibrium thermodynamics this split is standard (Oono & Paniconi 1998; Hatano & Sasa 2001), and both parts are defined by comparing a process with its reverse. Copying, death and division have no true reverse, so we measure against a different reference: the best type in the population. The housekeeping term is then the genetic load, −log₂(w̄/wmax), in bits per protocell generation. It is information, not heat: the chemistry's own dissipation can lower it but never pay it. A lineage is kept when its payoff edge exceeds it. Part of it can be paid by a gift instead (mixing, or a feed that resupplies lost content); a lineage paid that way is held by its environment, not kept. Excess information is log₂(1/p0) bits at fixation from a starting frequency p0.
A genetic code is a third kind: a channel, worth at most about 1 − H₂(ε) bits per binary read (for a q-letter position, log₂q − H₂(ε) − ε·log₂(q − 1)). It becomes a rate only through reads per generation.
Possibilities and copies meet only through a few conversion rules. Each comes from standard copying chemistry, evolutionary theory or physics; the copy-number map is a simple segregation calculation.
The common currency is held together by these rates. It is an exchange, not an identity.
In plain terms: errors and unfinished copies each cost bits per copy; every doubling of copies per division buys one bit back; and free energy buys accuracy.
Each component is a resource the setting supplies for free: activated nucleotides, cycling, Mg²⁺, compartments, dispersal and so on.
Gifts come in two kinds. A stock gift replaces a function, so it removes bits from what must be specified. A rate gift changes fidelity, completion or copy number, so it changes how much can be held. The two are never summed.
A real place (a pond, a vent, a drip) is a point in gift space. The walls do not depend on which place it is. Only the starting point does.
Put both on one picture. Across: gifts withdrawn. Up: information the lineage holds and keeps. A lineage can only sit where what it holds covers what is no longer given. Each step up has to be found within the time available, and kept at a bit rate it can afford.
The nursery is the gift-rich corner of this space. There, cycling multiplies the number of tries (finding), pays some rents (keeping) and supplies dispersal. It is where a lineage can start, and a place it must eventually leave.
So coming alive becomes a set of constraints on this space, rather than a story. Part 2 lists them in order.
The search budget is log₂ of the number of tries: molecules × cycles × time. A millilitre at 1 µM of 50-mers holds about 6×10¹⁴ molecules, roughly 49 bits per cycle.
The keeping budget is set by how many copies each division makes: a few bits per copy. It is spent again on every copy.
A stock is compared with a stock, and a rate with a rate. Keeping binds most often. Search over the whole path is not free either: it sits right at its budget.
Two measured numbers, fidelity and junction completion, become a rent of 11.3 bits per copy. A lineage making 40 copy attempts per division can afford about 3.8, so the slack is −7.5. (Forty is our model's choice. Each doubling of copy attempts adds one bit: at 400 the slack is about −4.1.)
Most of the deficit is completion (9.0), not fidelity (2.3). So whatever happened, something must have paid for completion. A ×5 tether plus re-priming ρ ≥ 0.5 across freeze–thaw closes the gap (the tether gain is assumed; it has not been measured).
The measurement that decides it: the fraction of stalled copies that resume after a cycle (ρ).
We call an amino acid attached to the copying RNA a handle. An aminoacylated primer extends about ×100 faster at 0 mM Mg²⁺, and no faster at 50 mM (Radakovic et al. 2021). The same ledger that says QT45 is short on completion says who can pay that bill, and when.
Slack = log₂(available ÷ required). A deficit names the number the next experiment has to move, and by how much.
Each row is in the unit of its own inequality (bits, bits per copy, or log₂ of a concentration ratio), shown under the name. Bars show ranges, dots single values.
Each tool already exists in someone's corner of the field. What is new is putting them in one currency, against one clock, so that they become inequalities that must have held however the history went.
Search bills and rent bills are both log₂ of counts, so we can say which one binds. Most often it is keeping.
A rate constant, a fidelity or a tolerance becomes a pass or fail by a stated number of bits, plus the bypass that would pay for it.
Laws hold for any chemistry. Constants hold within a chemistry class, or where two different models agree, or come from one model. Readers can see which layer each claim rests on.
We split the ledger the way physics does. A law holds for any chemistry and any parameters. A constant is a number a law takes as input: from the literature, from two different models that agree, or from one model. The laws say what any path must pay and in what units; the constants say how much this chemistry charges.
| Law | Statement | Constants it needs |
|---|---|---|
| L0 Counting | Every quantity is log₂ of a ratio of counts. Stocks (bits) and rates (bits per event) are never added; a rate becomes a stock only over a stated window, and two rates add only on the same clock. (A convention.) | — |
| L1 Finding | found ⇔ M(T)·2−b ≳ 1 | log₂ M(T): ≈ 10–17 bits is re-made without heredity in our nursery |
| L2 Keeping | Kept at level L over T ⇔ headroom when rare > 0 and edge head-to-head > 0. Kept ⇔ erosion per division < payoff edge per division; steady share of protocells carrying it f* = 1 − Ediv/epay. | epay (free); copy attempts k |
| L3 Copy floor | log₂(k · Pfunc ⁄ s) > 0 | k; s = 2 strand steps for +/− copying |
| L4 Rent | Rc = Σj log₂(1/pj) + nc · log₂(1/(1 − ε)) | pj, ε, nc, junctions per copy |
| L5 Tie | (Price) A gift that raises every protocell equally, whether it carries the content or not, moves nothing; a gift's edge is proportional to its content-tied share. Without a payoff or a commons, the content cannot spread. Heredity cannot be a gift. | tied share; between- vs within-unit covariance |
| L6 Linkage | Under a shared copying budget, gene i changes share by log₂(qi/q̄) per round. Unlinked genes stay together only if between-unit selection beats that spread plus drift (≈ 1/K); otherwise linkage or a stoichiometric controller. | per-copy survival spread; capacity K |
| L7 Internalization | Two inequalities, never summed: (rate) (1 − a)·τ·ef > Eint; (stock) the internal version must already be findable, bint ≤ log₂ M(Tfail). | gift availability a, tied share τ, internal rent |
| L8 Holding | Iclosure − Igift,s ≤ β · nc,max | nc,max from L2 and L4 at the rate gifts in force |
| L9 Drift | |e| ≳ 1/(Ne ln 2) to be selected | Ne (state with each edge) |
| L10 Exchange | Discrimination: ΔΔG/(kT ln 2) bits per event. Kelly: signal value ≤ I(signal; state) per signal. Mass: k ≤ monomer supply ÷ monomers per copy. Energy enters only through discrimination and mass. | ΔΔG, signals per generation, monomer supply |
| L11 Keeping identity | At steady state, the information selection must supply each protocell generation to hold a kept lineage equals its rent: the genetic load in bits, −log₂(w̄/wmax) = Ediv. This is housekeeping information. What selection actually writes is at most this; the rest is paid by losses that still leave descendants, or by resupply from outside. Checked in three different models (within 3%). | level L; stationary population |
Corollaries: exit (L2 and L7 applied at the destination); the fidelity check before escape; content must copy at least as fast as its unit divides; a lasting code comes only after completion stops depending on freeze–thaw, because the same cycles release a fluid membrane's contents. The walls of Part 2 are these laws with constants filled in. The order of the walls is not a law: it follows from which constants bind first.
| Key constant | Value or bracket | Status |
|---|---|---|
| epay | free; ours 0.405 bits per protocell generation | scanned |
| Copies per division | keep threshold Rc* ≈ 3.8–3.9 bits/copy at 40 attempts (a model choice; +1 bit per doubling); ≳ 5–9 complete copies per division | one model |
| Completion pj | QT45 0.66; a cis genome of ~130 nt needs ≈ 0.97 | literature + model |
| Re-priming ρ | ≥ 0.5 with a ×5 tether | unmeasured: the first decisive measurement |
| Fidelity ε | QT45 0.059; the path needs ≈ 0.02 | literature |
| Membrane Mg²⁺ tolerance | single-chain ≈ 1–5 mM (pure fatty acid) to ≈ 10 mM (with monoester); ≈ 4 + 150x mM with phospholipid fraction x (a fit to one study) | literature |
| Handle cutoff | ≈ 25–35 mM Mg²⁺ (lost at ≈ 38–42 mM in a rising-Mg²⁺ pond) | two models |
| Charging supply | ≈ 2–15 charges per vesicle per lifetime | two models (×2–4 apart) |
| Bridge lifetime needed | ≳ 25–100 years; measured hours to days | two models |
| Feeding limit xe | 0.83–0.94 phospholipid fraction | analytic + literature |
| Exchange between units | ≲ 0.5–1% per generation | two models |
| Content release per freeze–thaw | lasting code needs ≲ 4–8% per cycle; fluid phospholipid membranes measured at 18–57% | one model + literature |
| Escape edge | −log₂ P(content survives a generation) − tightening cost: +0.5 to +1.5 bits per generation in separated droplets; −0.08 in the pond | two models (form) |
| Charging discrimination | single pass: ≥ 3 kT to found, ≈ 5 kT for a majority code; or 2 kT + one recheck. Close amino-acid pairs measured at ≤ 2.2 kT | model + literature |
| Whole-path search | cost 93–152 bits vs budget 92–133 | estimate |
What it takes to come alive, written as constraints on the phase space of the living state, in the order they must be met.
A nursery is a region of gift space where the setting supplies most of what self-replication needs. Our working example is a cold, cycling, low-Mg²⁺ setting with concentration, activated monomers and amino acids. Its periodic driver, freeze–thaw or wet–dry cycling, does three separate things:
The nursery is not a claim about a particular place. Templates it supplies are a stock, not heredity: a flow-fed stock collapses when the feed stops. Whatever the nursery supplies, the lineage has not yet learned to provide for itself. To leave, it must internalize each gift it depends on, in an order the walls allow, and still pay its rent. That is the funnel.
Read down from the nursery to the first cell able to leave it. Each ring is a wall: a condition the lineage has to meet before it can go further, in the order the walls must be crossed.
Off to the side are dead ends. Most of them are traps of getting too good too soon: a faster copier, or a product shared with the neighbours, before the structure exists that would keep the gain tied to the lineage that made it. A lineage that falls in stays there. The next section shows why.
The walls hold whatever the setting was. The setting only changes where a lineage starts, and which dead ends are open.
A replicator on the path has to stay inside a window. Below it, the lineage can't pay its rent and dies out. Above it, the replicator copies better than its current organization can keep tied to its content. Then selection rewards whatever copies fastest, and the lineage slides into a stable basin with no way forward. The window rises as organization is added, so what is too good early on becomes fine later.
Floor, good enough: the lineage pays its rent (Ediv < epay). It rises as gifts are withdrawn, because each withdrawn gift is a new bill.
Ceiling, too good: the most copying power the current unit can keep tied to its content. It rises as privacy, coupling and a chromosome are installed.
The same jump in copying power is a trap at step 2 and safe at step 5. "Too soon" means crossing the ceiling before the tie that would hold it is in place.
Getting too good too soon is not peculiar to the origin of life. Several fields have met it and named it on their own terms, without treating it as one phenomenon.
Depth depends on population size: a trap is about 2Ne times as deep, in bits, as the edge (bits per generation) that holds the lineage there. At an effective population size Ne ≈ 900, an edge of only −0.01 bits per generation makes a dead end about 18 bits deep, while the window of the path is 0–5 bits deep. Only steady mixing between populations, about 0.5% per cycle, let a lineage climb back out.
What the ledger adds is where the ceiling sits at each step on the path to a cell, so each version of the trap can be priced in the same units.
Each one is a measured or modelled failure from our simulations, placed on this map.
Axes are schematic: "copying power" stands for speed and yield together, and "organization" for the order of crossings on the funnel. The curve shapes are illustrative, not fitted.
We do not claim success. What we can now do is state the remaining distance. Every step on a first path is either already shown in the laboratory, shown in parts, or reduced to a named measurement with a threshold to beat. As far as we know, no earlier account of the origin of life could say how far it was from done in one currency.
| Step | Status | Where it stands |
|---|---|---|
| A ribozyme copier | in parts | QT45-class copiers exist. Finding one costs ≤ 42 bits (our estimate), well inside the budget (+18 bits). Fidelity is nearly enough (−0.1 to −0.6 bits per copy). Completion is the bill: −3.3 bits per copy at the junction, payable by a tether (gain assumed) plus re-priming across freeze–thaw. |
| An attached amino acid (the handle) | in parts | Aminoacylated primers extend ~×100 faster at 0 Mg²⁺ in nonenzymatic, activated-monomer extension (Radakovic 2021), which passes the ledger by +2.6 to +5.9 bits. Not yet shown on a ribozyme copier. Needed only in low-Mg²⁺, fatty-acid compartments. |
| One chromosome | modelled | A shared copying budget forces linkage (one simulation model plus an analytic check). A cis genome is kept from completion ≈ 0.97 (~130 nt, 40 attempts per division); linkage is forced by ≈ 0.98. |
| Compartments that grow and divide | shown | Fatty-acid vesicles grow, divide and keep RNA, at low Mg²⁺; division needs gentle shear (Zhu & Szostak 2009). Phospholipid drives competitive growth (Budin & Szostak 2011): +0.7 to +2.9 bits at the loose stage. |
| Converting the membrane | in parts | This growth advantage is inherited only if gained fatty acid is converted to phospholipid. No acylating ribozyme yet, but thioester transacylation of lysophospholipids works without enzymes (Liu et al. 2020). |
| Charging and the first code bit | in parts | Tiny aminoacylating ribozymes exist (Turk 2010). Measured selectivity for close amino-acid pairs is ≤ 2.2 kT (kT is the scale of thermal energy), enough only with one freeze–thaw recheck. Which label a code takes needs only a slight bias (of order 1/Ne per cycle, where Ne is the effective population size); keeping it accurate needs the recheck or ≈ 5 kT. In models, the first bit lands on the polarity position whether handles or channels pay first. |
| Coupling, channels, tightening, escape | modelled | Ordered and priced in the models. Phospholipid-driven growth comes before coupling, or with it. Feeding limit (where a tighter membrane starts to starve the cell) at 0.83–0.94 phospholipid fraction. In separated droplets, tight beats loose by −log₂ of the chance content survives a generation (+0.5 to +1.5 bits per generation, two different models). At high Mg²⁺, tight + channels are favoured. A lasting code needs to leave the freeze–thaw, which releases a fluid membrane's contents. |
The ledger turns the attempt into a checklist. Each measurement below decides a step, and each has a threshold. Clearing them in order is what a first path requires.
Our numbers are not the last word. A measured value that differs from ours goes straight into the same inequality and moves the slack, so the ledger stays useful even where we are wrong.
| # | Measure | Threshold to beat |
|---|---|---|
| 1 | Re-priming: the fraction of stalled primer–template complexes that resume after one freeze–thaw cycle (ρ) | ρ ≥ 0.5 with a ×5 tether |
| 2 | Per-junction completion of a QT-class copier on a tethered partner | ≥ 0.95 for the copier; ≈ 0.97 for a cis genome |
| 3 | The commitment ratio kc/koff per junction on a tethered template | ≈ ×17 above QT45's free-template value |
| 4 | Complete genome copies made per protocell division | ≳ 5–9; every rent threshold moves by log₂ of this number |
| 5 | Copying fidelity at useful yield | ε ≈ 0.02 per site |
| 6 | RNA retention through freeze–thaw of fatty-acid vesicles with some phospholipid, at 25–50 mM Mg²⁺ | tolerance ≈ 4 + 150x mM (a fit to one study) predicts ≥ 40 mM at x ≈ 0.25; above that, no handle is needed |
| 7 | Content release per freeze–thaw cycle vs phospholipid fraction | ≲ 4–8% per cycle for a lasting code (one model); fluid phospholipid membranes measured at 18–57% |
| 8 | Release under slow freezing for gel-phase (saturated) phospholipid blends at ≈ 30 mM Mg²⁺: where a lasting code could be kept | ≲ 4–8% per cycle |
| 9 | An acylating ribozyme, or thioester transacylation of lysophospholipids (shown without enzymes, Liu 2020), ideally at neutral pH | conversion ≳ 0.5–1 per year |
| 10 | Rupture and dispersal per freeze–thaw cycle | sets the escape edge, −log₂ P(content survives a generation) |
| 11 | A ribozyme copier using an attached amino acid to finish copies at low Mg²⁺ | restores copying toward its high-Mg²⁺ rate below ≈ 25–35 mM |
| 12 | Supply of activated amino acids per vesicle in a cold pond, and charged-handle lifetime at 0–5 °C | ≈ 2–15 charges per vesicle per charge lifetime |
| 13 | Charging specificity of self-aminoacylating ribozymes for close pairs (Ile vs Val vs Leu) | ≥ 3 kT single pass, or 2 kT plus a recheck |
We put the same questions to four models, each built on a different method, with a fifth session coordinating them.
All four models are Claude, run in separate sessions. Splitting the work this way is not an error filter. It keeps each session focused on one method and its details, which a single session holding everything at once cannot do for long.
Each model addresses the same question in a different way: a deterministic ODE model, a population-genetics model, a stochastic-corrector model, and a variety of open chemostat. A result that comes out the same under different methods is more trustworthy than one that rests on a single method. It is still not independent validation.
A fifth session put the same questions to each model, compared the answers, and kept the shared record that this page calls the ledger.
Agreement between models catches some errors, but every model here is the same AI. Every claim on this page still needs checking by people and by experiment. The measurement list is where that starts.
This way of setting up the question differs from the usual one, so a few points of scope are worth stating directly.
We map what any path must satisfy rather than propose one. A pond, a vent or a drip is a starting point in gift space, and the walls hold for each of them.
Once a step is kept, the next is searched against its own budget. So the bits measure cost, not improbability, and the central question becomes one of rates.
Negative slack names what had to be paid, the bypass that could pay it, and the measurement that would show it was paid. It says where to look next.
The keeping law extends Szathmáry and Demeter's stochastic corrector and the origin of chromosomes. What it adds: the threshold in one currency with the payoff free, the copy-number map, helper RNAs losing at any benefit, and the amino-acid handle and first chromosome as one step.
Kept is persistence at a stated level against a matched control with the function erased. It is not fixation. The control can persist too, so persistence on its own is not taken as evidence of function.
The amino-acid handle is needed only at low Mg²⁺ in fatty-acid membranes. It reverses at about 25% phospholipid in a Mg²⁺-rich pond, so it is a statement about conditions rather than about a peptide world.
For the code we price finding (bits) and keeping (bit rates). These are different quantities from the ≈ 4.2 bits of codon-table entropy per codon.
It plays the role that housekeeping heat plays in nonequilibrium thermodynamics: the steady cost of holding a driven state. It is measured in bits per protocell generation, not in heat.
A slight chemical bias can pick a label; selection and proofreading keep it. "Cells before code" concerns founding; sharing between coded lineages can come later.