Among perfectly rational, truth-seeking Bayesian agents, unrestricted information sharing makes the group's beliefs worse. Not because agents are biased or irrational — the model gives them perfect reasoning — but because unconstrained information flow creates correlations in evidence that rational updating cannot correct for. Each agent treats incoming signals as independent when they are not, and the accumulated bias from this structural misattribution overwhelms the benefit of having more data.
The Tibetan lunisolar calendar rests on a foundational arithmetic ratio: 67 lunar months equal 65 solar months. This discrete approximation, embedded centuries ago, propagates through every calculation the calendar makes. The accumulated drift is not from measurement error or ignorance — the original astronomers were sophisticated — but from the structural rigidity of embedding an approximation as an axiom. A reform study identifies this as a deconstructible problem: separate the core arithmetic from the traditional labeling, and you can fix accuracy without destroying identity.
Both systems fail because a foundational simplification — independence of signals, or rational arithmetic — gets treated as structural truth rather than useful approximation.
The Bayesian agents have a correct update rule. They just apply it to the wrong object. They update on each signal as if it were independent evidence, when in fact signals from other agents carry redundant information. The more freely information flows, the more correlated the evidence becomes, and the more the independence assumption diverges from reality. Information that should be discounted at each step compounds unchecked.
The calendar has a correct computational framework. It just builds on a frozen ratio. When 67/65 was established, the drift was negligible within the calendar's operational lifetime. But centuries of propagation amplify the initial approximation into noticeable seasonal misalignment. The fix is structural — update the foundational arithmetic — but the fix is also delicate, because the traditional labeling system has become culturally load-bearing.
The deeper parallel: both systems demonstrate that the most dangerous approximations are the ones embedded at the foundation rather than at the surface. Surface-level errors get noticed and corrected because they produce visible anomalies. Foundational simplifications produce drift that accumulates below the threshold of immediate detection but above the threshold of long-term damage.
This is the free poison: information that is individually valuable, structurally corrosive, and invisible until the damage is done. The rational agents are making correct local inferences. The calendar is making correct local calculations. The corruption is architectural, and architecture is the last thing anyone checks.