A jury you can't bribe.

The machine in the agora

Athenian juries were enormous — 201, 401, sometimes 501 citizens — because the city understood its enemy precisely. A jury's size wasn't ceremony; it was arithmetic against bribery. But size alone doesn't help if the fixer knows which 501 to visit the night before. So the decisive invention wasn't the big jury. It was the kleroterion: a slab of stone that chose the jurors on the morning of the trial.

Every eligible citizen carried a bronze ID ticket — a pinakion. At dawn, tickets were slotted into the kleroterion's columns. Beside the slab ran a bronze tube; an official poured in white and black balls and released them one at a time with a crank. A white ball seated an entire row of ticket-holders; a black ball sent that row home. No one — not the magistrates, not the litigants, not the jurors themselves — knew who would judge the case until the balls had fallen. There was no jury to bribe, because until that moment, the jury did not exist.

And Athens paired the lottery with its complement: accountability after the fact. Every official, on leaving office, faced the euthyna — a mandatory public audit of everything done with power. Selection you cannot game going in; scrutiny you cannot escape going out.

Sharding's classic weakness

Every consensus decision in Hyperscale — every block, every certificate — is made by a committee of validators, and each committee tolerates corruption of up to a third of its seats. One monolithic blockchain has one committee: to attack it you must corrupt a third of everything. But a sharded network runs dozens of committees, and there lies the classic objection: corrupting one small committee is vastly cheaper than corrupting the whole network. If an adversary can pick a shard and concentrate on its guards, sharding has traded security for throughput.

The naive stafferies all fall to the same fixer who haunted Athens:

Balls made of everyone's dice

Committees are drawn from the validator pool by seeded randomness — but where do you get dice that thousands of mutually distrustful machines all agree on, and that none of them can load? The beacon — the network's coordination chain — keeps a randomness accumulator, re-mixed every epoch from verifiable random contributions carried in beacon proposals. Each contribution is checkable (a validator can't choose its value, only compute it), and the mix binds them all: to bias the draw you'd have to control the contributions and know everyone else's in advance. Submitting malformed randomness is itself a jailing offense.

Crucially, the timing mirrors the agora: assignments for the next epoch are determined only when the current epoch's beacon block folds — moments before the term starts, never seasons ahead. An adversary contemplating shard 7 doesn't know who will guard shard 7 next month, because nobody does. Until the balls drop, the committee does not exist.

A jury that never finishes forming

Athens re-drew its juries every morning. A blockchain can't: a validator joining a shard must first sync that shard's state, so wholesale re-assignment would have entire committees perpetually mid-download — a security win purchased with a liveness disaster. Hyperscale's compromise is a trickle: on a fixed interval, every committee rotates out exactly one member, replaced by a fresh draw from the pool.

The trickle's arithmetic is the point: a slowly-working adversary is on a treadmill. By the time they've compromised members of shard 7's committee, rotation has scattered some of them to other shards and seated strangers in their places. The details are all load-bearing: victims are chosen by domain-separated randomness (nobody can aim the exit at a rival), a rotated-out member can't be drawn straight back into the seat it left, only fully-synced members are rotated out (an exit that weakens the committee is worse than none), and committees mid-reshape are skipped so the churn never undermines a handover in progress.

Euthyna: the log that can't skip a line

Selection and churn keep the adversary from concentrating; expulsion makes every seat they do win a wasting asset. Misbehavior is punished only on proof, and both kinds of proof are Byzantine-resistant:

Lying is self-proving. Signing two conflicting votes for the same slot is treason with a paper trail — the two signatures are the evidence. It rides to the beacon and the validator is jailed permanently. No trial needed; the crime carries its own conviction.

Slacking is witnessed — uncensorably. Chronic offline behavior (skipped proposals, missed duties) generates witness records that every replica derives independently from the chain itself. These accumulate in a per-shard log whose fingerprint is quorum-attested in every block header — and here is the teeth: a proposer building a block must extend the log with every record the chain implies, because every verifier recomputes the expected fingerprint and rejects any header that omits one. A proposer can neither quietly drop the record that would jail a friend, nor mint one to frame a rival.

Jailing then feeds back into the topology safely: the expelled validator leaves its committee immediately, quorum math adjusts, and the lottery refills the seat. Liveness offenses can be forgiven after a cooldown; equivocation never is.

What the lottery guarantees — and what it assumes