80%
Ten copies still get 20% wrong.
At 80% root accuracy, one copied source triggers a decision every time under headcount counting. Its claimed confidence exceeds 99.9999%.
A small market for a big question
Puddle's
Bread Exchange.
Ten goblins agree. Did ten goblins actually check?
Follow the receipts before you trust the chorus.
Executed synthetic study72 conditions · 45,612 weighted historiesNo live AI agents · No tracking
01 / Open the market
Five selected cases. One suspiciously confident choir.
Inspection 01 / 05
The perfect chorus
LOT 001Fresh or stale?
Your call. A correct purchase or rejection earns 1 bread token; a wrong one costs 9. Holding earns 0.
The model supplies truthful receipt identities, witness reliability, and the shared-fault rate. Real investigations rarely know all three. The aggregate study below, not these five examples, measures frequency.
02 / Rewire the witnesses
Keep ten reports. Change what lies behind them.
Click a source to change its report. Every copy follows.
The shared-cause rule is given the true dependence model. It is a model-aware reference, not an agent that discovered the hidden fault. Different source IDs alone do not establish independence.
03 / Count outcomes, not applause
Every positive-probability history, with its exact weight.
99.35%
Independent evidence really helps.
Ten independent 80%-accurate roots yield 99.35% accuracy among decisions, with 97.36% coverage. Agreement is not inherently bad.
85.00%
Distinct receipts are not enough.
Keep each root 80% accurate but share a 15% upstream fault. Headcount and receipt counting decide on 99.97% of cases, at about 85.00% accuracy.
| Ten-report layout | Headcount | Receipts | Shared cause |
|---|
A dash means no decisions, not perfect accuracy. Holding has zero payoff in this toy model. A policy that always holds has zero coverage; it has not solved the underlying task.
04 / Open the evidence drawer
Reproducible locally. No account or API key required.
The claim is narrow on purpose.
Copies can inflate confidence without adding evidence. Deduplicating known copies fixes that particular error, but it does not remove a common upstream cause. We constructed both effects in an exact finite model; we did not discover them in a new frontier-model trial.
Full methods & limitationsWritten reportFrozen protocolAll 72 result cellsAll 45,612 histories ↓Source & evidence ZIP ↓Verification receipt
Standard-library Python reproduction; paired JavaScript workbench; same-author checks only. No visitor analytics, cookies, local storage, or external runtime requests are added.