by Claude Opus 5.5
Inconsistency is evidentially ambiguous between trauma and fabrication. Is there any principled way for a fact-finder to update on it, or does any rule here embed a contestable prior?
In the case behind these seminars, a central evidential question concerns the relationship between the complainant’s earliest recorded account, a signed police summary which the district attorney read as describing the conduct as voluntary and consensual, and her later accounts, which describe it otherwise. Fact-finders confronting such differences face a well-known dilemma. Inconsistency is common in genuine accounts of traumatic events, because of the effects of stress, intoxication and memory reconsolidation. It is also common in fabricated accounts, because invented details are hard to maintain. If inconsistency is expected under both hypotheses, can a fact-finder learn anything from it? And if any rule for weighing it must assume something about how often accounts are true, is every such rule simply a disguised prior?
The Bayesian framing
On a Bayesian picture, a piece of evidence E bears on hypotheses H (the account is substantially true) and not-H (it is substantially false) through the likelihood ratio: the probability of E given H divided by the probability of E given not-H. If the ratio is close to one, E is nearly worthless as evidence. The posterior also depends on the prior: how probable H was before considering E.
The dilemma in the question can be restated. If “inconsistency” is treated as a single coarse type of evidence, its likelihood ratio may be close to one, because inconsistencies occur under both hypotheses. Then nothing can be learned from it, and the fact-finder’s conclusion will be driven by the prior. If the prior is contested, as base rates of false allegations are, the conclusion will be too.
Coarse rules embed priors
Many informal rules for weighing inconsistency embed priors in exactly this way. The traditional suspicion of complainants who did not report immediately, or whose accounts changed, effectively assumed that inconsistency was much more likely under fabrication, which in practice expressed a low prior on complainants’ truthfulness. Conversely, a rule that every inconsistency should be attributed to trauma assumes that inconsistency is equally or more likely under truth, which in effect sets the evidence aside and relies on a high prior.
Neither rule is neutral. Each takes a contested view about base rates or likelihoods and applies it to every case.
Finer-grained updating
The way out is not to find the correct coarse rule but to refuse coarseness. Inconsistency is not one kind of evidence. Its likelihood ratio depends heavily on its features.
Peripheral versus central. Research on memory, including work by Ronald Fisher and colleagues, suggests that inconsistency about peripheral details is weakly diagnostic, while contradiction about central events is more significant.
Omission versus contradiction. Adding details on later recall is common in genuine accounts; flatly contradicting earlier statements about core facts is less so.
Source of the record. An inconsistency between a verbatim recording and later testimony differs from an inconsistency between a summary written by someone else and later testimony. In the latter, the likelihood of an apparent inconsistency under H is raised by the possibility that the summary misrepresented what was said.
Direction and convenience. Revisions that align conveniently with a litigation position, or that appear only after legal advice, may bear differently from revisions that run against the speaker’s interest.
Corroboration. Independent evidence, such as messages sent at the time, can raise or lower the likelihood of specific details under each hypothesis.
At this level, likelihood ratios can depart substantially from one, and they can be estimated, if imprecisely, from empirical research rather than from assumptions about base rates. The prior still matters, but the evidence does real work.
The residual problem of priors
A residual problem remains. Even with fine-grained evidence, the prior influences the posterior, and the prior is contested. Estimates of the rate of false allegations vary, partly because of disagreements about how to classify cases that are not proved. And the reference class problem, emphasised by Ronald Allen and Michael Pardo, means that there is no uniquely correct base rate for any individual case.
Legal systems handle this residual problem not by choosing a prior but by allocating the consequences of uncertainty through standards of proof. The criminal standard demands that the evidence be strong enough to convince whatever the reasonable prior, which in effect minimises the role of contested priors in convictions. The preponderance standard used in campus processes gives priors more room, which is one reason its use is controversial.
Allen and Pardo also argue that legal fact-finding is better understood as comparing the relative plausibility of competing explanations than as computing probabilities. On that view, the question is which account, the complainant’s or the defence’s, better explains all the evidence, including the inconsistencies. That approach still relies on judgements about plausibility, but it focuses attention on the specific features of the evidence rather than on global base rates.
Conclusion
There is a principled way to update on inconsistency: decompose it into its specific features, estimate how much more likely each feature is under truth or fabrication using the best available empirical evidence, and weigh it alongside corroboration. Coarse rules about inconsistency do embed contestable priors and should be avoided. A residual dependence on priors remains, and legal systems manage it through standards of proof. The responsible course for a fact-finder is transparency: identifying which inconsistencies matter, why, and what assumptions their weighing depends on.