Abstract
Is it possible for a large sequence of measurements or observations, which support a hypothesis, to counterintuitively decrease our confidence? Can unanimous support be too good to be true? The assumption of independence is often made in good faith; however, rarely is consideration given to whether a systemic failure has occurred. Taking this into account can cause certainty in a hypothesis to decrease as the evidence for it becomes apparently stronger. We perform a probabilistic Bayesian analysis of this effect with examples based on (i) archaeological evidence, (ii) weighing of legal evidence and (iii) cryptographic primality testing. In this paper, we investigate the effects of small error rates in a set of measurements or observations. We find that even with very low systemic failure rates, high confidence is surprisingly difficult to achieve; in particular, we find that certain analyses of cryptographically important numerical tests are highly optimistic, underestimating their false-negative rate by as much as a factor of 2<sup>80</sup>.
Introduction
In classical statistics and standard engineering practice, we often assume that measurements are independent. Under this assumption, every new piece of confirmatory evidence—a witness identification, a positive lab test, or a successful software check—increases our confidence in a hypothesis.
However, the real world is messy. Systems fail, labs get contaminated, and human processes become biased. This paper addresses the "paradox of unanimity": the phenomenon where a large sequence of unanimous observations actually makes a conclusion less likely because the probability of a systemic failure (the test itself being broken) begins to outweigh the probability that the hypothesis is true. The authors formalize why "too good to be true" is a mathematically sound skepticism.
Methods: Modeling the Failure State
The core innovation of this research is the introduction of a hidden "failure state" variable ($F$) into the standard Bayesian framework. Instead of just calculating the probability of a hypothesis $H$ given evidence $X$, the model is expanded to account for the possibility that the measurement process itself has failed.
This framework is applied to three distinct scenarios:
- Archaeological Evidence: Determining the origin of a Roman pot where a small percentage of factories might "contaminate" the results.
- Legal Proceedings: Analyzing police lineups and the ancient Jewish Sanhedrin law, where unanimous conviction led to acquittal.
- Cryptography: Evaluating the Rabin-Miller primality test, specifically looking at how physical hardware errors (like cosmic ray bit-flips in memory) affect security guarantees.
Key Insights and the Paradox of Unanimity
Non-monotonic Confidence
In systems with a failure mode, confidence does not rise indefinitely. It often peaks and then declines as more unanimous evidence arrives. For example, in a model determining the origin of a pot with a 1% contamination rate, confidence in the origin peaks at just five positive tests and then drops back toward 50/50 as more unanimous results arrive.
The "Failure State" Takeover
As the number of unanimous results grows, the likelihood that you are seeing a "true" result increases linearly in the exponent, but the likelihood that the entire process is biased eventually becomes the more probable explanation for the lack of any dissenting data. This explains why ancient Jewish law (the Sanhedrin) required acquittal in the face of a unanimous guilty verdict; it was a sophisticated intuitive grasp of systemic bias.
Key Results
- The 95% Ceiling: In a police lineup with a 1% systemic failure rate (e.g., a biased officer or flawed procedure), it is mathematically impossible to reach 95% certainty of guilt, no matter how many witnesses provide unanimous identifications.
- Cryptographic Optimism: Current analyses of primality tests are "highly optimistic." By ignoring the reliability of the underlying hardware (e.g., bit-flips occurring at approx. $10^{-19}$ per bit-second), they underestimate false-negative rates by factors as high as $2^{80}$.
- The Dissent Requirement: For a conclusion to be truly robust, we should actually expect and require a small amount of "noise" or dissent. A perfectly clean, unanimous signal in a noisy world is a red flag for systemic failure.
Conclusion
Accounting for even tiny rates of systemic failure sets a hard ceiling on achievable certainty. Future work will explore continuous bias models and the development of practical "upper thresholds" for automated decision-making systems to trigger alerts when results appear too consistent to be real.
Summary updated