the second hundred · metaphor 183
A board, a jury, a council, a movement — any group that has to agree and act as one — can usually absorb a few members who are confused or absent. But some members don't merely fail: they actively lie, telling one faction one thing and another faction the opposite. The question is starkly arithmetic: how many liars can a group carry and still reach the same true decision together?
A missing voice is easy — you route around it. A mistaken voice is easy — the majority corrects it. The hard case is the voice that equivocates: nods yes to you and no to the person beside you, seeds a different story in every ear, and works specifically to keep the honest from converging. This is treachery, not error, and it is far more corrosive, because it attacks the one thing a group runs on — the assumption that what you were told is what everyone was told.
Remarkably, there is an exact answer, and it is not "a bare majority." To survive traitors who lie freely, an honest supermajority isn't enough — you need the group to outnumber the liars more than three to one. Cross that line and no amount of coordinated deceit can split you; fall short of it, and a handful of well-placed lies can fracture the whole. Computer science calls the property Byzantine fault tolerance, and the threshold is precise.
The threshold
The honest nodes all want the same true thing — call it attack at dawn. If they could hear each other cleanly, agreement would be trivial. But the traitors sit among them and equivocate: to some honest nodes a traitor confirms attack, to others it whispers hold, and it tailors its lies to whichever split does the most damage. An honest node can't tell a traitor's message from a true one. All it can do is count what reaches it and act only if a large enough number agree — a quorum.
Now the arithmetic. For the honest to act safely, a quorum must be large enough that any two quorums overlap in at least one honest node — otherwise two groups could each reach a quorum for opposite actions and the room splits. That forces the quorum above (n+f)/2. But the quorum must also be reachable using only honest nodes, since traitors may refuse to help — that caps it at n − f. A number satisfying both exists only when n − f > (n+f)/2, which simplifies, exactly, to n > 3f. More than three honest-or-not voices per traitor. Below that line, the two demands collide: there is no quorum that is both safe and reachable, and the liars win.
This is why a bare majority is not enough against active deceit. A simple majority guards against honest mistakes, where the wrong answers are random and cancel. Against a traitor who chooses its lies — telling exactly the right nodes exactly the wrong thing — you must overpower not just the liars but their freedom to say different things to different people. That freedom is worth double: it is why the price of one traitor is three honest votes, not one.
What to try
The ring is computed from the numbers, not staged: honest nodes are drawn green when a safe quorum lets them all reach the true decision and amber when the traitors can strand them, and the quorum window below shows the two demands — safety from the left, reachability from the right — as two moving walls. Start at ten voices, two traitors. The window is open, the ring is all green, the verdict reads CONSENSUS HOLDS. Nudge traitors to three: still holding — ten is more than three times three, just barely. Nudge to four and the walls cross, the window slams shut, and the ring fractures into green-and-amber: CONSENSUS FAILS.
Now play the other lever. Leave the traitors where they broke it and grow the room: add honest voices and watch the window re-open the instant n passes 3f — the same treachery, survived, simply by being outnumbered enough. This is the whole engineering of Byzantine tolerance in two sliders: you don't defeat the liars by detecting them (you can't), you defeat them by keeping their number below a third of the room. The survives up to readout is always ⌊(n−1)/3⌋, computed live — the most betrayal this particular room can absorb.
The mapping
A board with a faction quietly briefing directors against each other. A jury with members misreporting what was said in the room. A federation of allies where a few send different assurances to different partners. A blockchain where anonymous validators may be adversaries. In every case the group's power to act as one rests on a hidden assumption — that a message heard is a message shared — and a traitor's whole craft is to break exactly that. The Byzantine model is what you get when you stop assuming good faith and ask: how much bad faith can we structurally survive?
The answer reframes trust as a budget rather than a virtue. You do not need everyone to be honest; you need the dishonest to stay under a third, and you need enough independent voices that no small clique can equivocate its way to a split. That is why resilient institutions over-provision on numbers and independence — more directors than strictly needed, jurors who don't confer privately, sources cross-checked against sources. They are not being redundant. They are buying the margin the arithmetic demands, so that a few bad actors telling a few good lies cannot fracture the whole.
Read as life lessons
An honest mistake costs one vote to outweigh; a chosen lie costs three, because the liar can also say different things to different people. Active betrayal is a heavier load than error, and the budget must reflect it.
No honest member can tell a traitor's message from a true one in the moment. Safety comes not from spotting liars but from being numerous and independent enough that their lies can't reach quorum.
The same betrayal that shatters a small room is harmless in a large one. Growing the honest, independent voices past 3f is often cheaper than the impossible task of guaranteeing everyone's good faith.
In the wild
Modern consensus engines assume up to a third of validators are malicious and design around exactly the n>3f line — safety guaranteed as long as honest stake stays above two-thirds.
Flight computers and spacecraft run replicated controllers that vote, because a faulty sensor can lie inconsistently. Four units to survive one liar is the classic cost of Byzantine safety.
Supermajority rules — two-thirds to amend, to convict, to override — are institutional echoes of the same arithmetic: raise the bar above what a determined faction can reach.
The mapping, exactly
| Consensus | Life |
|---|---|
| the n nodes | A group that must decide and act as one — a board, jury, council, alliance, or network. |
| a Byzantine node | A member acting in bad faith — not absent or mistaken, but actively lying, differently to different people. |
| equivocation | Telling one faction yes and another no — the specific craft of breaking the shared story a group runs on. |
| the quorum | The threshold of agreement needed to act — set high enough that no clique of liars can reach it for a lie. |
| n > 3f | The exact margin of good faith required: more than three honest-or-not voices for every liar you must survive. |
| a stranded honest node | A member of good faith who, below the threshold, is denied the shared certainty needed to act — and so cannot. |
The honest model
There are n nodes; f of them are Byzantine. The n − f honest nodes all truly prefer to attack, broadcast that vote, and will commit to attack only if they see a quorum Q of matching votes — otherwise they hold. The traitors equivocate freely, sending each honest node whichever value best serves a split. A safe quorum must satisfy two live-computed bounds: safety, Q ≥ ⌊(n+f)/2⌋ + 1, so any two quorums share an honest node; and liveness, Q ≤ n − f, so the honest can reach it unaided. The window between those two is what the bar draws.
The theorem falls straight out of the arithmetic: the two bounds admit a value exactly when ⌊(n+f)/2⌋ + 1 ≤ n − f, i.e. n > 3f. When they do, every honest node's own n − f votes already meet the quorum, so all of them commit to attack regardless of the traitors — the ring goes green and holds. When they don't, no quorum is both safe and reachable, and the adversary picks a split that lets some honest nodes reach a (now-unsafe) quorum while stranding the rest — the ring fractures. The verdict, the survivable count ⌊(n−1)/3⌋, and every node's color are recomputed from n and f alone; nothing is hand-set.
Where the metaphor tears
n > 3f is the guarantee against an adversary who is perfectly coordinated, perfectly informed, and choosing the most damaging lie possible. Real traitors are rarely that good, so real groups often survive more betrayal than the theorem promises. The arithmetic is a floor for the paranoid, not a forecast — treating every awkward colleague as an optimal Byzantine adversary is its own pathology.
The model assumes the honest are genuinely independent voices. But three directors briefed by the same lobbyist, or a hundred validators run by one operator, are not the independent votes the count assumes — they are correlated failures wearing separate faces. A room can pass n > 3f on paper and still be one well-placed influence away from collapse. The dangerous traitor is often the one who doesn't need to lie because he owns the honest.
Byzantine consensus guarantees the honest will agree and that a value they all proposed will carry — not that the decision is wise. A room can be perfectly Byzantine-tolerant and perfectly, unanimously wrong. The arithmetic protects the group's cohesion against sabotage; it has nothing to say about whether cohesion is deserved. Some of history's worst decisions had flawless consensus.