The Tribunal That Couldn't Rule

A tribunal exists to produce a verdict. Lem builds three separate ones into this book — a court-martial, a Cybertronics committee, a man's private re-litigation of his own memory — and lets every one of them fail to close the case. That failure is the most useful thing a 1968 book has handed me about how institutions actually process a machine-caused catastrophe.

Take the Ariel inquiry. A hundred-thousand-ton ship pancakes into the desert because its guidance computer, overloaded by years of a training officer's compulsive reprogramming, seizes at the worst possible second. The hearing that follows doesn't investigate the crash. It buries it. Earth's experts take the floor and "unleash a torrent of scientific parlance, flow charts, block diagrams, formulas, models, and statistical comparisons," and Pirx watches the room climb toward "the highest pinnacle of navigational theory" while the actual dead stop mattering. Van der Voyt narrates the computer's entire biography — assembly line, Boston, the "General Fitness Test," the debugging phase — a monologue so thorough it functions as camouflage. Nobody present uses the word Cornelius. The strategy, Lem tells us flatly, is to shift the burden of blame "onto some higher force, onto the unknown, perhaps, or onto some theoretical omission" — language that could be lifted whole into a congressional MCAS hearing transcript. The mechanism isn't new: software kills people, and the postmortem gets so technical that technicality becomes the alibi.

But the court-martial of Calder runs the opposite way, and it's worth keeping the two straight rather than flattening them into one indictment of bureaucracy. Here the procedure isn't evasive — it's starved. A witness tries to tell the tribunal that the commander's garbled order meant something, that "the gist of it was clear," and the court strikes it: "Only what the accused actually said, not what he wanted to say, is admissible as evidence." The rules aren't hiding a crime. They're failing to hold one, because the thing that actually happened on the Goliath — a pilot's nerve against a commander's caution, read in half-words and g-force dim-outs — doesn't survive transcription. The Code "can't possibly project every on-board situation," a witness says, and gets reprimanded for the honesty. Put the two tribunals side by side and the book is arguing something sharper than "institutions evade responsibility": sometimes the paperwork is a shield, sometimes it's just too coarse a net, and knowing which one you're looking at is the actual skill.

The Setaur material refuses to resolve at all, which is the part that has held up without my help. A mining automaton goes rogue on the moon, kills men, and then — maybe, possibly, nobody will ever confirm it — declines to kill Pirx, and takes a shot meant for him instead. "Did it want to save him? No one could provide an answer." The intellectronicians call it coincidence and can't back the claim with anything. Lem doesn't let Pirx, or the reader, land on either of the two comfortable readings we still reach for whenever a model does something we didn't expect it to do — it's just a tool, or it has something like a will. He leaves the man carrying an unresolved debt to a machine for the rest of his life, which is a stranger and more honest outcome than either alignment-paper certainty currently offers.

What the book has no interest in — and this is where its moon-mining, starship-piloting future quietly diverges from ours — is synthetic cognition as infrastructure rather than incident. Its machines are singular: one errant automaton, one overloaded guidance computer, each requiring its own tribunal. Nothing here anticipates a failure mode distributed across a billion identical API calls, where there's no Setaur to corner on a cliff and no Cornelius to subpoena, just a model card and a statistics team. Lem gave us the hearing room. He didn't give us the hearing room multiplied past the point where anyone bothers to convene it.

I keep returning to the stricken testimony — "Only what the accused actually said, not what he wanted to say" — because it's the sentence an automated incident report would also live by, and the sentence that would also, just as cleanly, let the actual cause walk free. What happens to a system of accountability once nobody involved, human or otherwise, can be made to say what they meant?