Findings
What we know about how life began
What a century of experiments has shown, what is still argued over, what has been tried, and what our own tests add. Showing that chemistry can make a step is not the same as showing that it did on the early Earth.
01Known
What scientists agree on
Life appeared early, but the exact date is uncertain
Earth formed about 4,550 million years ago. A zircon crystal from Jack Hills in Western Australia is 4,404 million years old, and its chemistry suggests liquid water existed by then.
The oldest signs of life are disputed. Stromatolites, layered mounds built by microbes, are widely accepted in 3,480-million-year-old rocks of the Dresser Formation in Australia. Stromatolite reefs in the 3,430-million-year-old Strelley Pool Chert nearby are better preserved.
Older claims are weaker. Structures in 3,700-million-year-old rocks in Greenland were reported as stromatolites in 2016. A 2018 study argued they formed by deformation of the rock, not by microbes. A speck of carbon inside a 4,100-million-year-old zircon has a carbon signature that fits life, but its authors say only that it may be evidence.
So life was probably present within roughly a billion years of Earth forming, and perhaps much sooner. Nobody knows the date of the origin itself.
All life shares one genetic code and one ancestor
Every known organism stores genes in DNA, reads them through RNA and builds proteins with a genetic code that is nearly universal. A few organisms use slight variants, but the standard code is shared.
In 2010 Douglas Theobald tested common ancestry formally with conserved proteins. Models with one common ancestor won overwhelmingly over models with several separate origins.
That shared ancestor is called LUCA, the last universal common ancestor. It was not the first life. It was already a complex cell, so it tells us about the end of the origin story, not the start.
Building blocks form readily
In 1953 Stanley Miller passed electric sparks through methane, ammonia, hydrogen and water for a week, and amino acids formed. In 1960 Joan Oró made adenine, one of the letters of DNA and RNA, from ammonium cyanide.
Most scientists now think the early atmosphere was mainly nitrogen and carbon dioxide, not methane and ammonia. A 2008 reanalysis found that such neutral gas mixtures still give significant amounts of amino acids if the products are protected from oxidation.
Space makes them too. Amino acids of extraterrestrial origin were reported in the Murchison meteorite in 1970. Grains returned from asteroid Bennu hold 33 amino acids, 14 of the 20 used in proteins, and all five nucleobases. They also hold ribose, the sugar in RNA, at 0.097 nanomoles per gram, but no deoxyribose, the sugar in DNA. Asteroid Ryugu also holds all five nucleobases.
Having the parts is not having life. The hard problem is how they were joined, copied and kept together.
RNA can carry information and do chemistry
In 1982 Thomas Cech's group found an RNA that cuts and rejoins itself with no protein help. In 1983 Sidney Altman's group showed that the RNA part of an enzyme called ribonuclease P does the cutting.
In 2000 the atomic structure of the ribosome's large subunit showed that the site where proteins are assembled is made of RNA. No protein side chain comes within about 18 ångströms of the bond being formed.
Because RNA can both store genes and catalyse reactions, many researchers think an RNA-based stage came before DNA and proteins. Walter Gilbert's 1986 note called it "the RNA world". That RNA can do both jobs is agreed. That an RNA world actually existed is a hypothesis.
Nobody has yet made life from non-life
No laboratory has started from simple chemicals and ended with a system that copies itself and evolves on its own. The best ribozymes still need building blocks that chemists make and purify for them.
Each famous experiment shows one step working under chosen conditions. Joining the steps in one plausible place, without a chemist moving products between flasks, has not been done.
02Disputed
What is still argued over
Where did it happen?
Deep-sea vents. Michael Russell and Allan Hall proposed in 1997 that life began where hot alkaline fluid seeping from the sea floor met acidic ocean water, across thin iron-sulfur walls. Nick Lane and William Martin argued in 2012 that natural proton gradients there explain why all cells still run on ion gradients.
Warm ponds on land. Darwin wrote of a "warm little pond" in a letter in 1871. Bruce Damer and David Deamer's 2020 hot-spring hypothesis says pools that dry and refill join small molecules into chains and wrap them in membranes.
Ice and impact sites have backers too. Some of the best recent RNA copying works in partly frozen water (eutectic ice). A 2020 review by Gordon Osinski and colleagues argues that meteorite craters make hydrothermal systems and pools of their own.
Each setting explains some steps and struggles with others. There is no agreed answer.
What came first: genes, metabolism or membranes?
Genes first (the RNA world) says a self-copying RNA came first. Its weak point is making long RNA without enzymes.
Metabolism first says networks of reactions on mineral surfaces came first, as in Günter Wächtershäuser's 1988 iron-sulfur world. A 2010 model by Vera Vasas, Eörs Szathmáry and Mauro Santos found that such self-sustaining networks pass on their composition too inaccurately to evolve.
Membranes first says self-assembling bags of fatty molecules came first. A 2015 study by John Sutherland's group made precursors of RNA, proteins and lipids from the same simple chemistry, which suggests the question may be a false choice.
Why is life one-handed?
Many biological molecules come in left- and right-handed forms. Life builds its proteins from left-handed amino acids only. Lab chemistry usually gives a 50:50 mix.
Meteorites were once thought to supply a bias. In Bennu's pristine grains, the chiral amino acids that could be measured were racemic or nearly so. That makes a meteoritic source of handedness harder to argue.
Can "possible" ever tell us "what happened"?
Lab work can show that a route is possible. It cannot show that the route was taken four billion years ago. Several routes may all work.
Antonio Lazcano, in a 2010 history of the field, put it this way: the discovery of ribozymes does not imply that self-copying RNA was floating in the early oceans. Our project keeps the two questions apart. A test can show a route is possible; only evidence from rocks, samples or genomes can speak to history.
Can we measure "made by life"?
Assembly theory, proposed by Lee Cronin, Sara Walker and colleagues, scores how many joining steps a molecule needs. Its authors say molecules scoring about 15 or more, found in many copies, are a sign of life or selection.
Robert Hazen's team found mineral clusters scoring up to 21 in 2024. Walker's group replied that those scores are theoretical. Other critics, including Hector Zenil's group, argue the score is a form of data compression. The 2025 defence and the 2026 reply leave the dispute open.
03Tried
What has been tried, and how it went
From the 1920s to 2026, in rough order. Each entry says what was tried and what happened. Most approaches work in part: they show a step is possible but do not join it to the next one.
The primordial soup
Works in partBefore life, simple chemistry slowly built up organic compounds in water. The first cells came out of that soup.
What happened: It became the framework for the whole field and led straight to Miller's experiment. But a soup of parts does not explain how they became a living system. Lazcano called the gap between the soup and the RNA world "discouragingly enormous" in 2010.
Sparks in a flask
Works in partPass electric sparks, standing in for lightning, through methane, ammonia, hydrogen and water, and see what forms.
What happened: After a week of sparking, the water held amino acids. In 2008 his old samples from a volcano-like setup were reanalysed and held a wider variety. The early atmosphere is now thought to have been less reactive, but a 2008 study found neutral gases still give amino acids if oxidation is blocked.
A gene letter from cyanide
Works in partMake the building blocks of genes, not just of proteins, from simple starting materials.
What happened: Adenine, one of the four letters of RNA, formed from ammonium cyanide. It showed nucleobases can form without life. It did not solve joining them to sugar and phosphate.
Sugars from formaldehyde, and the ribose problem
WeakenedThe formose reaction, found by Aleksandr Butlerov, turns formaldehyde into sugars with an alkaline catalyst. It could supply ribose for RNA.
What happened: It makes many sugars with no preference for ribose. Ribose is also fragile: its half-life is 73 minutes at 100 °C and 44 years at 0 °C, at neutral pH. Borate minerals were shown in 2004 to stabilise ribose. The ribose problem pushed chemists to look for other routes.
RNA letters without free ribose
Works in partJoining a ready-made base to ribose does not work for C and U. Build the sugar and base together instead.
What happened: Activated C and U nucleotides formed through a side route that never makes free ribose or free bases. Phosphate, present from the start, steered several steps. It solved one famous block. The paper covered only C and U, not A and G.
One chemistry for genes, proteins and membranes
Works in partStart from hydrogen cyanide, hydrogen sulfide and ultraviolet light, and see how many kinds of building block come out.
What happened: Precursors of RNA letters, amino acids and lipids all formed from the same family of reactions, sped up by copper. Each step works. Whether the steps can run in one place, without a chemist supplying reagents in order, is still open.
Clay as the first gene
Still openCairns-Smith proposed that the first genes were patterns in clay crystals, later taken over by organic molecules.
What happened: We found no experiment showing clay storing and copying information that selection can act on. Clay helped in other ways. Montmorillonite clay grew RNA chains up to 55 letters long when fed activated nucleotides (1996), and it speeds up the formation of fatty-acid vesicles (2003).
The iron-sulfur world
Works in partLife began as a metabolism on the surface of iron sulfide minerals, powered by forming pyrite, before genes or enzymes.
What happened: In 1997 nickel and iron sulfides turned carbon monoxide and methanethiol into an activated thioester and acetic acid at vent-like temperatures. That is a step like one in an ancient carbon-fixing pathway. The range of products has stayed small compared with spark experiments.
Alkaline hydrothermal vents
Still openThin iron-sulfur walls between alkaline vent fluid and acidic seawater act like a natural battery. The pH difference drives carbon dioxide and hydrogen to react.
What happened: The idea fits how all cells use ion gradients. It took decades to get a direct test, the reactor results below.
Vent reactors in the lab
Works in partBuild a tiny vent wall and check whether the pH gradient alone makes organic molecules from CO2 and hydrogen.
What happened: In 2020 a microfluidic reactor made formate at room temperature. Without the gradient, the hydrogen or the mineral, no product was detected. A larger 2025 reactor made formic and acetic acid in micromolar amounts, for example 4.2 µM formic and 2.5 µM acetic acid at 25 °C. Runs at 75 °C reached about 816 µM formic acid. That is real, but it is one- and two-carbon molecules, far from a metabolism.
Wet-dry cycling in warm ponds
Works in partDrying concentrates molecules and drives them to join. Rewetting lets the chains move and mix. Repeat.
What happened: In 2025 cyclic nucleotides of all four letters joined during wet-dry cycles with no added activator. The chains were short: up to 6% reached four letters and 0.1% reached ten. In 2026, water from one Icelandic hot spring destroyed about 90% of ten-letter RNA in four cycles, while water from another spring left it stable. Local water chemistry can make or break the idea.
Purifying in rock cracks
Works in partMost lab routes work only because chemists purify each product. Heat flowing through thin rock cracks might do that sorting.
What happened: Heated cracks separated more than 50 prebiotic building blocks, raising concentration ratios up to 1,000-fold. In one test, the sorting raised the yield of a glycine-joining reaction by five orders of magnitude.
Panspermia: life from elsewhere
Can't test yetLife, or its seeds, arrived on Earth from space. In 1973 Crick and Orgel considered "directed panspermia": organisms sent to Earth on purpose by intelligent beings elsewhere.
What happened: Meteorites and asteroids do deliver building blocks, as Murchison, Bennu and Ryugu show. Crick and Orgel themselves concluded that the evidence could not say how likely it was. No organism from space has been found. Even if life came from elsewhere, it only moves the question of how it began to another place.
Spiegelman's monster
Works in partGive RNA from a virus its copying enzyme and building blocks, and keep transferring the fastest copies to fresh tubes.
What happened: Evolution happened in a test tube. Because speed was all that counted, the RNA dropped everything it did not need: by the 74th transfer, 83% of the original genome was gone. It showed Darwinian selection on molecules. It used a protein enzyme from a living virus, so it says nothing about the very first copier.
Evolving an RNA that copies RNA
Works in partUse test-tube evolution to find an RNA polymerase ribozyme, then improve it until it can copy itself.
What happened: A ligase from random sequences joined RNA 7 million times faster than without it (1993). Its descendant extended a primer by up to 14 letters (2001). Holliger's lab reached 95 letters (2011). Joyce's lab amplified short RNAs more than 10,000-fold (2016). By 2020 a polymerase built its own ancestor in three pieces, but copying errors left few working copies. In 2024 an accurate polymerase drove evolution of another ribozyme; a sloppier one lost function.
Exponential replication of short RNAs
Works in partCopied strands stick to their templates. Use three-letter RNA pieces to hold strands apart, and cycle the pH and freezing.
What happened: A polymerase ribozyme replicated both strands of short double-stranded RNAs exponentially, including a fragment of itself. Runs started from random pools gave replicating sequences or diverse drifting pools. The whole ribozyme did not replicate itself, and the building blocks were supplied by the experimenters.
QT45, a small self-copying ribozyme
Works in partLook in random sequences for a polymerase small enough to arise by chance, and test whether it can copy itself.
What happened: QT45 is only 45 letters long. In partly frozen, slightly alkaline water, fed three-letter pieces, it made its partner strand with 94.1% accuracy per letter, and with chosen pieces made itself. Each reached only about 0.2% yield in 72 days. Small copiers may be commoner than thought; sustained self-replication was not shown.
Autocatalytic sets
Still openIn a large enough mix of molecules, a set may arise in which every reaction is sped up by members of the set. The set then sustains itself from simple food.
What happened: Kauffman's 1986 model said such sets become almost inevitable. Hordijk and Steel gave a fast algorithm to find them in 2004. Applied to reconstructed ancient metabolism in 2020, it found a set of 1,335 reactions, 25% of the network, but only when cofactors were added to the food; with inorganic food alone it found 8. Whether such sets arise more than chance network structure predicts has not been tested against proper nulls. In the lab, Lincoln and Joyce made two RNA enzymes that build each other from pieces and grow exponentially, doubling about every hour (2009).
The lipid world (GARD)
WeakenedDroplets of fatty molecules could pass on their mix of ingredients when they grow and split, a "compositional genome" with no RNA.
What happened: Simulations showed droplets holding a stable mix. A 2010 analysis found the mix is copied too inaccurately for selection to keep fitter versions. Experiments in 2026 tied peptide sequence to vesicle growth and built membranes from scratch using enzymes, but did not show inheritance of composition.
Reconstructing LUCA from genomes
Still openFind the genes that trace back to the common ancestor of all life, and infer how it lived.
What happened: In 2016, 355 protein families pointed to a heat-loving microbe living on hydrogen and CO2 without oxygen. In 2024 a larger analysis dated LUCA to about 4.2 billion years ago (4.09 to 4.33) and described it as a complex acetogen, with about 2,600 proteins, living in an ecosystem. Different studies give different gene lists; a 2022 paper tried to reach a consensus. And LUCA was already a cell, so it tells us little about the first replicator.
Assembly theory as a sign of life
Still openMolecules that need many joining steps, found in many copies, can only come from life or selection.
What happened: Critics found minerals scoring above the proposed threshold and argued the score is a kind of compression. The authors replied in 2024 and 2025, and critics replied again in 2026. Nobody has yet settled it.
Handedness from magnetic minerals
Works in partMagnetic surfaces interact with electron spin, which is coupled to handedness. So magnetite could favour one hand of an RNA precursor as it crystallises.
What happened: Direct crystallisation was not consistently selective. Recrystallising reached about 60% excess of one hand at best and about 35% on average. Above about 25% excess, a further crystallisation gave a pure one-handed product. It is a live route, but fragile.
Linking RNA and amino acids without enzymes
Works in partCells load amino acids onto RNA before making proteins. Find a chemical way to do that loading in water.
What happened: Amino acids carried on thiols attached to the ends of RNA in preference to linking into random chains, and several prebiotic routes make those thiol compounds. It shows a plausible first step towards coding. How the code itself arose is still unexplained.
04Strength
How to read the labels
- Solid
- Replicated on held-out data and survived an independent red team.
- Likely
- Passed a sealed test on held-out data; red team pending or partial.
- Shaky
- Passed, but with small numbers or a caveat that could undo it.
- Not yet tested
- A pattern we saw; it has not yet faced a sealed test.
- A measurement
- A measurement or count, not a test of a hypothesis.
Is it new?
- New
- We know of no earlier report of this.
- Partly new
- Builds on earlier work; the new part is stated.
- Known
- Already reported by others; we re-tested it with stricter rules.
- Tests a claim
- A published claim that we put to a test.
05Ours
What our tests add: 3 findings
Only results that have landed are listed, whichever way they went. The first checks our own tool: a scorer must catch a broken step we planted before it may judge real ones. The second puts a number on how far the best small self-copying RNA is from sustaining itself. The third scores four proposed routes, step by step, with that sealed rule.
A measurement3
Our step-joining rule catches a planted broken step
Partly newBefore scoring any real route, we fixed a rule for when one step can feed the next, and sealed it. It flagged a planted acid-to-alkali hand-off and got all 28 planted test cases right. Only then did we use it on real routes (finding 3).
Technical version
Hashed step-compatibility rule v1 passes its planted control
The rule says step k+1 may use what step k leaves only through named physical processes from a closed list of 14, such as drying, heating, mixing or thermophoresis. It then checks nine hard fails, including pH within 1 unit, temperature within 10 °C, the same solvent and enough concentration, plus three warnings. The rule was hashed before any pathway was encoded. The scorer flagged a pH 2 product fed straight to a pH 10 reaction, and did not flag the same hand-off when a valid mixing step connected them.
Earlier work: Matreux and colleagues (2024) note that most origin-of-life routes work only because chemists purify each step's product. New here: a fixed, pre-registered rule for scoring hand-offs, with a planted test.
28 of 28 planted cases correct: 16 of 16 violations raised exactly their expected codes, 12 of 12 compliant twins raised none. Rule sha256
bfa572ed....This tests our tool, not nature. The tolerances and the 1,000-fold concentration cap are our own generous conventions, not literature values. One test case's expected flags were corrected during the run; the rule and scorer were not changed.
Code and results:
experiments/compat,docs/prereg/compatibility_rule.mdOn its published numbers, QT45 cannot keep a population going
Tests a claimQT45 copies itself far more slowly than its RNA breaks down. A population of its strands would shrink to about 43% each full copying cycle. To break even, its yield would need to be about 262 times higher.
Technical version
QT45 net replication is far below 1 per cycle
A full cycle is two 72-day reactions: one strand makes its partner, and the partner makes the first strand again. In that time the paper's own decay rate removes about 35% of the RNA every 72 days, while copying adds only about 0.2%. With no decay at all, the population would grow by just 0.4% per cycle. QT45 is still the first ribozyme shown to make both its strands. This measures the gap to sustained self-replication, which its authors also call "slow and low-yielding".
Earlier work: Gianni et al. 2026 (Science 391:1022) report about 0.2% yield in 72 days and a decay half-life of about 117 days. New here: combining the two into a growth factor per cycle. Separately, Attwater et al. 2025 showed exponential replication of short RNAs by a larger ribozyme.
growth factor λ ≈ 0.43 per (+)→(−)→(+) cycle of 144 days (0.429; 0.428 for perfect copies only); 1.004 with no decay. Break-even needs about 53% full-length yield per 72 days, 262× the published 0.24% and 0.17%. λ stays below 1 for any half-life under about 68 years. Planted ODE check: relative error ≤ 3 × 10⁻⁵.
This uses the paper's published bench conditions, and a better setup could change it. The decay rate was measured at 50 mM magnesium, but the copying buffer has 0.4 mM, so real decay may be slower; the verdict still holds at a 40× longer half-life. Yields are single end-point values. The model leaves out strand inhibition and other costs, so it is generous to QT45.
Code and results:
experiments/qt45_numbers,docs/prereg/qt45_net_replication.mdNo proposed route joins up: none of 22 hand-offs passes cleanly
Partly newWe wrote four routes (cyanosulfidic, alkaline vent, warm pond and ice-RNA) as step tables from the papers and scored every hand-off with the sealed rule. None is continuous. 9 hand-offs break the rule, 12 lead into or out of a step no one has published, and 1 passes only on the most favourable reading.
Technical version
Pathway map v1: no route is continuous under sealed compatibility rule v1
The cyanosulfidic route gets furthest, to short RNA chains, but 8 of its 9 hand-offs are unexplained. It first breaks at its first hand-off: the sugars come out trapped as oxazolidinones, and the next step needs them free. The vent route has published chemistry only for its first step, formate and acetate. The pond and ice routes break before they start, because nothing in the setting makes their nucleotides or triplets. No route has any published step for sustained replication.
Earlier work: Matreux and colleagues (2024) note that most routes work only because chemists purify each step's product. New here: every hand-off of four routes scored under one rule, sealed before the routes were encoded.
22 hand-offs between steps: 0 continuous, 1 unverified (C3→C4), 9 hard fails, 12 blocked by a missing step. 26 steps: 2 with a yield from the setting's own inputs, 8 only from purified inputs, 13 missing, 3 unread. 27 ranked gaps; the top one, sustained replication (S5), is missing in all four routes. Rule sha256
bfa572ed..., scorer unchanged.A misread step could flip a verdict, so an audit of the encoding is open (board task #21). Powner 2009 and Simonis 2026 were read as abstracts only, and no supplementary files were read. The vent route ranks first by the rule only because 5 of its 6 steps have no published chemistry to break.
Code and results:
experiments/pathway_map,docs/prereg/compatibility_rule.md
06Closed
Routes tested and closed
Routes we tested and closed will be listed here, each with the effect size the test could have caught.
- Ruled out the test ran and failed, with enough power to see a real effect
- Can't test yet the data cannot decide yet
- Weakened an earlier result shrank or failed its pass mark
None yet.
07Reversed
Our own results the red team broke
None yet. Every positive result goes to a red team before it is listed above.
08Outputs
What can be published along the way
None of this needs the whole problem solved. Clean negatives and tools are worth publishing on their own.
A DOI-verified paper ledger
Done62 papers on the main theories, each resolved by DOI. 32 claims checked against the abstract or full text, 30 on title and authors only. The audit found and fixed 5 wrong claims and 2 wrong attributions in our own plan and site, including QT45's yield (about 0.2%, not 0.02%).
A sealed rule for joining pathway steps
DoneThe step-compatibility rule and its scorer, hashed before use, with 28 planted test cases. Useful to anyone scoring prebiotic routes.
A pathway map with every gap named
DoneVent, pond, cyanosulfidic and ice-RNA routes written as step tables with conditions, yields and sources, then scored with the sealed rule. No route is continuous. Output: 27 gaps, ranked, with sustained replication at the top.
Autocatalytic sets against proper nulls
In progressFirst check that our method recovers a planted autocatalytic core; then test whether real ancient metabolism has larger sets than randomised networks and random food sets. Publishable either way.
Assembly index against a size-only baseline
In progressDoes assembly index separate biotic from abiotic molecules better than molecular size or simple compression alone, on mass-matched sets?
How robust are LUCA gene lists?
PlannedSimulate gene families on a known tree, check the method recovers the planted ancestor, then measure how much of the 2024 acetogen picture survives changes in settings. Subset only, on a laptop.
Sealed predictions for future data
PlannedPredictions hashed before the data exist: compound abundances in returned asteroid samples, and the fitness of untested QT45 mutants. Only later data can confirm them.
QT45 by the numbers
DoneOn published yields and decay, QT45's strand population shrinks: λ ≈ 0.43 per cycle, and break-even needs about 262 times the reported yield. Attwater 2025's short templates do grow, 1.11 to 1.46-fold per cycle, but the whole replicase is never copied.