Agent Society · a decision system, not a chatbot system
The debate line proved agents sharpen answers by persuading each other (Du ’23 · Irving ’18 · Khan ’24).
Synod bets the same channel is how a decider gets played.
Every number on this page tests that bet.
Judges never converge. Only the chair commits.
Against a bluffer it wins 66% more than one agent alone. Against honest opponents it roughly ties. Once, it lost. We show you that too.
Baseline = the same engine, same moves, same game — just the council removed. One head instead of five.
Independent judgment. Deterministic commitment.
Most agent societies divide work. Synod divides judgment.
Five specialists independently evaluate the same decision.
No one persuades anyone. A neutral chair resolves the disagreement — with deterministic code.
The winning judgment becomes the action.
Preserving disagreement only matters if it changes outcomes.
One head believed the bluff and walked away with nothing.
Five checked it and held the line — all ten runs.
The difference wasn't a smarter model — it was preserving disagreement long enough to verify the claim.
Stress environment №2: the business deal →If persuasion isn’t the coordination mechanism, what is?
Every agent society must answer one question: who wins when the agents disagree? Most say: whoever argues best. Synod says: never the best arguer.
Measured, not asserted: remove Synod’s argument step — $0 changes. Flatten the chair — −$1,123. The chair is causal. Rhetoric isn’t.
The general flaw is collapsing independent judgments into one belief before the information is in — by debate, vote, or average. Swap the chair for a majority vote: the council loses $1,113 (Exhibit B).
You’ve read the bet. Now run it yourself.
Watch five independent judgments become one decision.
Stress environment №1 — war. Deception, incomplete information, irreversible stakes.
This isn’t a scripted demo. You’re watching five judgments stay independent until the chair commits.
Next — take a judge away below, and run it again. Then pick a different opponent.
Click a general to remove them. The number: how much of the 3,000 the council won. The bars: how well it read the bluff.
Against a bluffer, Sun Tzu is the one that matters — remove him and the fake threat never gets exposed. A different opponent needs a different general. You can't know which in advance, so you keep all five.
The same test — remove each general — run against all three kinds of opponent. The one whose removal costs the most changes with who you're facing.
If you stop here, you have the whole idea: judgments stay independent until a deterministic chair commits — and that wins where opponents lie. Everything below is the receipts.
If the bet is right, removing independence should cost money — every time.
Seven tests, one question: what happens when independent judgment disappears?
Every figure on this page can be regenerated. The suite is re-made from scratch, twice — the results match to the digit. (same check as npm run reproduce)
⚔ War view — the same runs, stakes in units. 🤝 Deal view — the same runs, real dollars.
Same opponent, 10 games each — one strong agent alone vs the five-member council. Taller bar = more won. The baseline is built to win, not to lose — the same engine with the council removed.
Yes — and deception is where it shows. how to read it
Three kinds of opponent. One values the relationship. One has a hidden hard limit. One is bluffing. The “±” is how much the result wobbled between games. Smaller is steadier.
Re-run the same 10, switching off one part each time. A big drop = that part was doing real work.
The “check the claim first” step matters most: −1,606. That's the call that turns a win into a walk. why the “argue it out” step shows −0 — measured in the engine AND live
We tested this twice. In the engine, removing the argument changes the result by $0. Then we tested it live, as matched pairs. The chair scored the same judges before the argument and after it. Same weights. Only the argument differed. Result, in 12 recorded trials: no judge ever moved. The final call never changed. That's the design, not a gap. Each judge is the only voice for its question. So there is nobody to talk around. A well-argued bad idea has no way in. Disagreement is settled by the chair instead. And the chair's weighing IS causal: flatten it and the council loses $1,123 (row above). (trial log: debate-ablation.json · reproduce: npx tsx src/harness/debateAblation.ts)
Five worlds written by a different model (Claude) to stress the council. Greener = better; red = wiped out.
Every single judge goes red in some world, but the council's row never does — that's why you keep all five. why this is fair — and where it stops
The test worlds came from a different model (Claude), not from Qwen. So the system didn't write its own exam. Honest limit: the 10 games show a result is steady within a world. They don't prove it holds across all real worlds. And five worlds from one author is a small sample. Read the numbers as “stable in these worlds.” Not “proven in general.” Widening this needs a third source, or real transcripts.
The same 10 games, but decided by one judge alone each time, vs the full council (the chair adapts). Longer bar = more won.
The council beats four of five judges and ties the fifth — without ever guessing which one to bet on. the honest tie
The judge that ties here is exactly the one Exhibit C breaks. On this tuned set, a probe-leaning general can match the council. The council's real edge is robustness. It never craters across the hold-out worlds. Every single general eventually does.
Each judge has a few dials. How cautious Hedge is. How eager Probe is to ask. The situation sets them, within hard limits. Compared here to leaving the dials fixed — same tests.
Adjusting to the situation beats fixed dials by 1,584 (8,968 vs 7,384). why this is a real delta
The situation sets each judge's dials, within hard limits. The engine checks and applies them. This measures what that adjustment buys over never touching the dials. A real component, not a relabel.
A benchmark is not the real world. So the ladder ends outside our control: the literature's negotiators, the real league — and finally, 4,277 real humans.
Three classic negotiators from the research field (Faratin 1998 — the standard ANAC baselines). They concede on a clock. They hide nothing. We didn't write them.
Roughly a tie — the honest, expected result: these opponents hide nothing, so checking buys little. what this test found — and fixed
Synod beats the stubborn one, ties the eager one, and loses a little to the steady one — it scouted an opponent with no secret. This test exposed a real flaw. The council used to hold past the final round. Live offers died on the table. We added a deadline rule: on the last round, take any offer that beats walking. Then we re-ran everything. Every existing exhibit came back byte-identical. Outside opponents finding a real bug is exactly what outside tests are for.
ANAC is the international contest where negotiation programs play each other. Synod is selling; $4,000 of profit is on the table.
what this can and can't claim
A replay can't claim Synod would have won more — change one move and the whole conversation changes. So no counterfactuals. We only use the humans' own outcomes: every deal where a real seller accepted a buyer's offer, split by whether Synod — replaying the same buyer's offers — would also have accepted. The gap between those groups is the finding. Prices map into Synod's native band by a fixed, disclosed rule. Reading real humans is rung three; running inside live human deals is the step that remains.
The takeaway: the bet holds. Synod's edge comes from preserving disagreement until resolution — five judgments held apart until the chair commits, instead of collapsed into one voice too early.
A bet is worth money when committing early is expensive.
Companies don’t lose money because they’re dumb. They lose it by committing too early — believing a story before checking it.
One decider makes the same two mistakes in every hard negotiation: it believes what the other side says, and it wants the deal too much to walk away. So it pays for fake threats, and gives up on winnable deals.
The council fixes both. One judge's whole job is to say "prove it." Another's is to say "what if we're wrong?" The chair decides who's right — this time.
Someone commits before the uncertainty is resolved. Synod is paid for the gap between what the other side claims and what's true.
It usually isn't — and procurement signs anyway. The council works out what they'd actually accept, first.
The oldest trick in selling. Test it before you cut your price.
Real, or a bargaining move? Every hire and every counter-offer pays for guessing wrong.
M&A: the rarest deals, the most at stake. That claim deserves a test, not a better price.
Every payout is a negotiation with someone who knows more than you — small overpayments, thousands of times over.
The war room isn't just a story. The same bluff-or-real call, at the highest stakes there are.
The price of one resolved bluff is already measured — $3,000 instead of $0 (Exhibit A). Every table above repeats it.
And where the other side hides nothing, there's no gap to be paid for — Synod roughly ties, and we measured that too (Exhibit F). We sell the gap, not magic.
Wherever a commitment is expensive to unwind, the council goes in front of it.
If the bet is right, the architecture follows from it.
situation → five judges, in parallel → disagreement, on record → deterministic chair → one committed decision
The key idea: the bet, built as structure — no persuasion channel exists. Five independent judgments, resolved by arithmetic into one decision.
If judgments must stay independent — therefore debate exists to expose disagreement, never to persuade the chair. One exception, and it's bounded: the most-opposed pair exchange one round, and a landed objection may move the defender's own score on the contested action — clamped to ±0.1, on the receipt. So a challenge can flip a razor-thin round; it can never stampede the council. Measured: in 12 live matched-pair trials none was close enough to flip, and removing the chair costs $1,123. Persuasion isn't how Synod reaches agreement. Arbitration is.
The room argues.
The chair reads the terrain.
The decision commits.
After the chair commits, the judge that owned the winning question drafts that part of the operational plan — Probe writes the reconnaissance, Trust the back-channel, Hedge the fallback.
If disagreement is preserved — therefore something rhetoric-proof must resolve it. Arithmetic gives three things a model can't: the same answer every time, immunity to rhetoric, and a visible trail — you can see exactly which judge, at what weight, decided. The models reason. The arithmetic decides.
If each judge owns one question — therefore each owns one formula:
Two safety gates guard the edges — walk away when nothing believable beats leaving, and take the sure win on the last round. Then the receipt is signed.
Every formula is the most legible member of its family — simple enough for a skeptic to recompute by hand:
Click a ▶ live one to run it now.
The honest claim: the council beats one agent in every environment we can test.
Standing on named prior art — ANAC / BOA negotiation agents, Bayesian opponent modeling, and the debate line: Du et al. 2023, Irving et al. 2018, Khan et al. 2024 — exact deltas in the README.