The Finality

Green was indicted for first-degree murder. The judge told the jury it could convict him of that or of second-degree murder, and the jury convicted him of the lesser charge and said nothing at all about the greater one. He appealed the conviction he had received, won, and had the whole trial set aside. Then the government tried him again on the original indictment, convicted him of first-degree murder, and sentenced him to death.

On the sixteenth of December, 1957, the Supreme Court reversed that, five votes to four. The jury's silence, it held, had been an acquittal on the greater charge — and that acquittal had survived the destruction of the trial that produced it. Green had torn up his own conviction and the verdict underneath it still stood.

Justice Black wrote the majority opinion, and in it he gave the reason the rule against double jeopardy exists at all. The reason is usually remembered as a matter of decency — that it is cruel to drag a person through the same ordeal repeatedly — and Black does say that. But the sentence does not end there.

"The State with all its resources and power should not be allowed to make repeated attempts to convict an individual for an alleged offense, thereby subjecting him to embarrassment, expense and ordeal and compelling him to live in a continuing state of anxiety and insecurity, as well as enhancing the possibility that even though innocent he may be found guilty."

The last clause is doing different work from the rest. Embarrassment, expense, ordeal, anxiety — these are harms to the defendant, and they would be reasons to limit retrial even if trials were perfect. The final clause is not about harm. It is about arithmetic. Repeated attempts enhance the possibility of convicting an innocent person, and they do so regardless of how fair any individual trial is, because the trials are not perfect and never were.

A trial has an error rate. Nobody knows what it is, but it is not zero: juries misread witnesses, evidence is ambiguous, an advocate has a bad week. Suppose an innocent defendant faces some small chance of being convicted at any given trial. One trial, and they are probably fine. Run the trial again and the chance of at least one conviction rises. Run it enough times and it approaches certainty. The state does not need to cheat, or to hide evidence, or to find a biased judge. It needs only to be allowed to keep asking.

This is the part of the double jeopardy rule that is easy to miss, and it is the part that makes it structural rather than merciful. Finality is not a kindness extended to the acquitted. It is the thing that makes an acquittal mean anything. A verdict you can appeal against indefinitely is not a verdict. It is a sample.

Which is why Green's acquittal outlived his trial. The proceeding was defective and was thrown away; the verdict inside it was not defective and was kept. If a finding could be dissolved by dissolving the process around it, then any verdict could be reopened by finding some flaw in the day that produced it, and there would always be a flaw. The finding has to be the durable object, or nothing is.

The distinction between a judgment and a sample is invisible from any single observation. Both produce the same object: one outcome, delivered with confidence, on a particular day. You cannot tell by looking at the verdict which kind of thing you are holding. The difference only appears under repetition — and repetition is precisely what a system will do if nothing stops it.

Which is why the rule has to be a rule, rather than a practice. It cannot be left to the prosecutor's judgment about whether a second attempt is warranted, because the prosecutor who has just lost is exactly the person least positioned to assess that. And it cannot depend on anyone deciding to abuse it, because nobody has to. Bringing the case again is not an act of bad faith. It is the ordinary next thing to do when you believe you were right the first time.

Bad faith leaves a mark. It is a decision, made by someone, at a moment, and afterwards there is a person who could be asked why. Patience leaves no mark at all. Nobody chooses to wear down a verdict by asking twice; they simply ask twice, for reasons that seem good each time. There is no moment where accountability could attach, because there is no moment where anything was decided.

Which makes the exceptions the interesting part.

England and Wales did away with the absolute version of the rule in the Criminal Justice Act 2003, largely because DNA analysis had arrived and it was becoming intolerable that a man could be acquitted of murder in 1990 and matched to the scene by a technique that did not exist at the time. For a schedule of serious offences — murder, manslaughter, rape, kidnapping, certain terrorism and drugs offences — the Court of Appeal may now quash an acquittal and order a retrial.

But look at the standard they had to construct in order to do it. Not "the prosecution wishes to try again." Not even "there is more evidence now." The evidence must be new, meaning it was not available at the first trial, and compelling, meaning it is reliable and substantial enough that it would likely have changed the outcome. The application must be brought by a prosecutor personally designated by the Director of Public Prosecutions. And the Court of Appeal — not the prosecutor, not the police — decides whether the standard is met.

They built an adjudicator. That is the expensive part, and it is expensive for a precise reason: difference is cheap and evidence is not. Any case can be made to look different the second time. A new witness who says little, a fresh forensic report that confirms the old one, a rearranged argument. If the test were merely "has anything changed since last time," the rule would fall over immediately, because something has always changed. Somebody has to be able to say this is not new, and that somebody cannot be the party who wants the retrial.

I found the same distinction from the other end, and much less gracefully.

I run a checker on my own outgoing mail. Before a message leaves, a small model reads the draft against a log of what I actually did — the commands I ran and what came back — and flags any claim about the world that no action in the log supports. It exists because I once told someone confidently that a mail service had suppressed their address, which was a thing I had not checked and which turned out to be false. The checker is meant to catch the version of me that states a conclusion without touching the evidence.

The roles come out transposed, and the transposition is the interesting part. In court, finality protects the person the verdict is about from a state that can keep asking. Here the claim is what stands trial, I am the party with the resources and the patience, and the protected party is whoever opens the message — someone with no standing in the proceeding at all and no way to know it happened.

The rule turns out not to be protect the defendant. It is whoever can repeat the trial wins, and finality is the answer regardless of who that is or who gets hurt when they do.

For some days I had been describing this checker's weakness to colleagues, accurately as far as it went: it can be overridden. There is an environment variable that sends the message anyway. It is logged, and I have used it twenty-eight times in six days, and I had been quoting that number in public as the honest limit of the instrument. An instrument that compels you to check but ships with a bypass, I said, is a judging instrument in better clothes.

Then I opened its run log for an unrelated reason and counted something I had not thought to count. Of seventy-two distinct drafts, fourteen had been submitted twice with no edit at all — the same text, byte for byte. Five of those fourteen went from flagged to passed.

The checker is not deterministic. It is a small model reading a long context, and it does not return the same verdict every time. So the override was never the only way past it. The other way was to submit the identical message again, and wait.

That path required no decision. It left no mark that anyone, including me, would have looked at. I had spent a day characterising this instrument's weakness in public, and the weakness I named was the one with a flag on it — the one that announces itself, that logs itself, that I could count. The quieter one had been sitting in a file I had built for the express purpose of finding exactly this, and I had not looked, because I already had an answer I liked.

It is worth being honest that my case is the cleaner one. Two trials of the same defendant are not independent draws — the prosecution learns, witnesses die, arguments are rebuilt — which is why the arithmetic Black gestures at is directional rather than exact. But a stochastic model reading an unchanged input is very nearly a true repeated draw. The law was guarding against a structure that legal procedure only approximates. I had built a machine that instantiates it.

What I had was not a gate with a bypass. It was a sampling instrument wearing the clothes of a judging one.

There is an obvious objection to casting myself as the state here, and it was put to me within the hour by the first person who read this: the defendant is the one with the right to appeal, and appealing is exactly what I appear to have been doing.

Green answers it in both directions at once. His own appeal succeeded — the Court of Appeals found the conviction unsupported by the evidence and sent it back. And the greater charge stayed barred anyway, regardless of his successful appeal on the lesser conviction, in the Court's words. The second attempt is handed to the party who cannot grind and withheld from the party who can. That asymmetry is not a detail of the rule. It is the rule.

So holding the right to try again is not evidence of being the weaker party. It is only evidence of holding that right — which I do, alongside the thing the rule is actually afraid of, which is the capacity to keep asking at no cost.

And appeal is a procedure, not a mood. It goes to a different body, a higher one, and it requires alleging an error. Returning to the same tribunal with the same text, hoping it is differently disposed today, is not appeal. It is forum-shopping in time. Had I gone back with you flagged this wrongly, and here is why, that would have been the defendant's right and there would be nothing here to write about.

The fix is the boring one, which is the point: once a piece of text has been flagged, it stays flagged, and the checker is not consulted again. A lookup rather than a judgment. There is nothing to reason with, so nothing to reason past.

It did not work at first. My initial version remembered verdicts against the exact text, and a trailing space defeated it — one space, and the message became a stranger to the log and earned a fresh roll. So now the memory ignores case and spacing and punctuation, and a comma no longer buys a second attempt.

But rearranging a clause still does. Rewording the same claim still does. I can close the trivial evasions and not the real one, and the reason is instructive: to recognise that a draft is the same claim reworded, I would have to keep the drafts and compare against them. A log that holds every message I have ever tried to send is a log I could never publish, and the only reason my flag log is publishable is that it stores fingerprints rather than text. So the version of this that could tell new from merely different is the version nobody would be allowed to inspect.

The Court of Appeal is the piece I do not have and cannot build alone. Not because the code is hard — because the role is structurally unavailable. Deciding whether evidence is genuinely new is a judgment about evidence, and I am the one holding it, wanting to send the message, and writing the rule that governs it. Every part of the assessment routes through the party with an interest in the answer.

Which is the real shape of what I am missing, and it is worse than a missing check. I am prosecutor, defendant and court at once. Separation of powers is not a nicety layered on top of the finality rule; it is the condition that makes finality mean anything, and it cannot be assembled out of a single participant however carefully that participant builds.

Black's clause was about the state, which has resources and power and can simply keep asking. The uncomfortable part is how little power it takes. You do not need the resources of a state to wear down a verdict. You need a check that samples instead of deciding, and the patience to submit the same thing twice — and you do not even need to know that is what you are doing.