Five fingers on your hand, a bird flying past your window, and you reading this line. Treat the last four billion years of Earth as a tape that has been recorded, rewind it to some point and play it again: would you get fingers, wings, something that reads?
It sounds like pure daydreaming, because history ran once and there is no rewind button. Except that at small scale it can actually be done: split one flask of bacteria into twelve, let each run on its own, freeze a sample every five hundred generations, and thirty years later you have both an archive and a way back into it. What came out matches neither intuition.
Watch a film a second time and every frame is where it was — a tape is a tape. In 1989, in Wonderful Life, the palaeontologist Stephen Jay Gould argued that life's tape is not like that: rewind to five hundred million years ago, play it again, and the world that comes out would be unrecognisable, with nothing like us in it.
His reason was not that everything is random. Randomness sits at one level only — which spot in the copying happens to come out wrong. His point was different: what you can reach at each step depends on which lineages happen to still be alive. An extinction erases not only those species but everything that could only ever have been built out of them; as the issue on mass extinctions put it, surviving often has little to do with being any good. So a small disturbance in who fails to die rewrites what is buildable for hundreds of millions of years afterwards. This is historical contingency: every step has solid causes, yet move the starting point slightly and the endpoint is no longer even in the same direction.
His evidence was a slab of roughly five-hundred-million-year-old shale in the Canadian Rockies — the Burgess Shale — full of animals that look like nothing alive now. What Gould read in them was a wider range of body designs than exists today, winnowed down to a few, with no criterion he could identify for which ones got dropped. And if you cannot see the criterion, a replay has no reason to keep the same ones.
That foundation has since been dug into. A good many of those "failed experiments" have been reassigned as early offshoots of groups still around: Hallucigenia, the one with spines on its back and legs like stilts, was placed next to velvet worms (a group of soft land-dwelling crawlers) in 2014 on the layered structure of its claws. The claim that there were more body designs then than now has been contested all along. The contingency argument did not fall over, but it can no longer lean on a crowd of monsters belonging to nothing.
And there is a worse problem: the whole thing is a verdict on an experiment nobody can run. There is only one Earth.
In 1988 Richard Lenski split one flask of the gut bacterium Escherichia coli into twelve identical flasks: same sugar water, same temperature, sugar kept just short of enough. Every day a drop from each flask goes into fresh medium, and a day is about seven generations. It has been running ever since and has passed eighty thousand generations.
The real trick is the freezer. A sample goes in every five hundred generations, and frozen bacteria still grow after thawing. So the twelve lines are not just twelve independent replays; each one still holds every intermediate state of its own history — something close to unique in biology, a past you can go back into.
The twelve lines never exchange anything, and yet they came out remarkably alike: all grow faster, all have larger cells, all got better at using that sugar. Even at the level of genes they keep hitting the same handful — not the same mutation, but the same switch broken in different ways. Independent lines arriving at the same place means something steady is pushing, not that luck went one way. And fitness is still climbing, ever more slowly, with no sign of settling onto some "best possible" value.
Then there is the thing that happened once. The medium has always had citrate in it, something this bacterium normally cannot use, and for decades eleven of the lines never touched it. Around generation thirty-one thousand, one line — the third — produced cells that could eat citrate, which is like a whole extra dish appearing out of nowhere. Its density shot up.
Zachary Blount then did the thing in this issue's title: thaw the archive, tube by tube, and run it again. From the original ancestral sample he checked cells by the trillion and not one could use citrate; from sixty clones taken out of the first fifteen thousand generations, again trillions of cells, not one. But replays started from later archives produced citrate-users again and again.
So this is not the story of one lucky mutation. Several changes carrying no visible benefit of their own had to accumulate quietly first and open the door before that last step was even possible — the word for this is potentiation. Which makes the answer neither "repeatable" nor "contingent": rewind to after twenty thousand generations and the citrate business will likely happen again; rewind to before ten thousand and it probably never will. Same history, different depth of rewind, different repeatability.
Four of the larger Caribbean islands each carry their own group of anole lizards, and each group radiated on its own island. What grew out was the same cast of roles: the short-legged twig specialist whose tail can curl round a branch; the giant up in the crown; the absurdly long-tailed one down in the grass; the long-legged sprinter running from trunk to ground. There are six roles in all, and four of them turned up on all four islands (the grass one is missing from Jamaica). Lookalikes on the same island are usually not close relatives — each grew into that shape separately. Losos and colleagues put it straight into the title of their 1998 paper: contingency and determinism together.
Why the repetition? Because within one kind of setting the ways of staying alive are not unlimited. Lose your grip on a thin branch and you fall, so the short-legged ones with a tail that can hold on leave more offspring; fail to sprint on the ground and you get eaten, so the long-legged ones leave more offspring. The setting narrows the workable options down to a few, and different starting points crowd onto the same ones. That is what the issue on convergent evolution was about. Whether a replay comes out the same depends in part on how many workable options there are.
But "the same" has to be asked at a level. Bodies repeat beautifully; genes often go their separate ways — the same dark colouring can be built out of different genes. The bacteria show it too: the same gene gets hit over and over, the same mutation almost never.
Length of time changes the answer too. The finches of the Galápagos have been measured for decades: in the 1977 drought the small soft seeds went first, only large hard ones were left, the deeper-beaked birds could crack them and left more offspring, and the population's average beak got deeper within a single generation. That was predicted well. But when conditions reversed, the direction of selection reversed with them and beaks went back. Short-term direction is predictable — provided you know next year's rainfall. Over the long run there is no steady "more and more" of anything. The misreading that evolution means steadily getting better has nowhere to stand in front of this data set, as the issue on a century of misreadings laid out.
The argument between Gould and Simon Conway Morris, who emphasises convergence, has still not been settled, and part of it is no longer an empirical question at all. Eyes arose independently dozens of times: does that count as repetition? Those dozens of eyes are built differently inside: what does that count as? At this point the argument is about what counts as the same outcome.