TLDR
Bayesian Belief Networks translated three qualitative ACH matrices into formal probability estimates: Spoliation P=0.90, Willful Blindness P=0.77, and Asset Concealment P=0.69, all starting from uniform 0.25 priors (PAPER TRAIL Project, 2026a). Alternative hypotheses -- understaffing, industry norm, legitimate estate planning -- received posterior probabilities below 0.01, effectively ruling them out as explanations for the observed evidence.
From Scores to Probabilities
Analysis of Competing Hypotheses (ACH) produces ordinal scores -- numbers that rank hypotheses relative to each other but do not express the probability that any hypothesis is true (Heuer, 1999). A score of +6.40 for willful blindness means it is more consistent with the evidence than its alternatives. It does not mean there is a 64% chance the hypothesis is correct.
Bayesian Belief Networks (BBNs) bridge this gap. A BBN is a directed graph where nodes represent variables and edges represent probabilistic dependencies. Given observed evidence and prior beliefs about hypothesis probabilities, the network computes posterior probabilities -- updated beliefs after considering the evidence -- using an inference algorithm called Variable Elimination (PAPER TRAIL Project, 2026a). The implementation uses pgmpy, a Python library for probabilistic graphical models.
Construction
One naive Bayes network was constructed per ACH matrix (PAPER TRAIL Project, 2026a). In a naive Bayes structure, each piece of evidence is conditionally independent given the hypothesis -- meaning the probability of observing each piece of evidence depends only on which hypothesis is true, not on what other evidence has been observed. This is a simplifying assumption, but it is conservative: in practice, evidence items may be correlated, which would concentrate probability mass even more strongly on the favored hypothesis.
Each hypothesis received a uniform prior of 0.25 -- expressing no prior preference for any explanation before examining the evidence (PAPER TRAIL Project, 2026a). The conditional probability tables were derived from the ACH consistency ratings: evidence items rated as consistent with a hypothesis increase its posterior, while items rated as inconsistent decrease it. The ratings were converted to probability values using a standard mapping from the five-point ACH consistency scale.
Three Posteriors
Spoliation (P=0.90). The BBN assigns a 90% posterior probability to the hypothesis that FedEx records were deliberately destroyed or withheld (PAPER TRAIL Project, 2026a). The competing hypotheses -- channel migration to personal or courier channels (P=0.087), OCR artifact or processing gap (P=0.009), and carrier switch (P=0.006) -- share the remaining 10%. The high posterior reflects the convergence of multiple independent evidence items: the abrupt October 2005 cutoff, the absence of a documented account closure, the correlation with the Palm Beach investigation timeline, and the active shipping volume in the months preceding cessation.
Willful Blindness (P=0.77). The BBN assigns a 77% posterior probability to the hypothesis that Deutsche Bank deliberately avoided investigating compliance red flags (PAPER TRAIL Project, 2026a). Revenue capture -- the hypothesis that compliance was subordinated to fee income -- receives P=0.230, still plausible but less consistent with the evidence. Understaffing receives P=0.00008 and industry norm receives P=0.00004 (PAPER TRAIL Project, 2026a). These last two are effectively zero. The evidence -- a single unverified approval email cited for five years, 97 cash withdrawals at exactly the third-party limit, and the eventual $150 million consent order -- is inconsistent with explanations rooted in resource constraints or standard practice (New York Department of Financial Services [NYDFS], 2020).
Asset Concealment (P=0.69). The BBN assigns a 69% posterior probability to the hypothesis that the Butterfly Trust functioned as a mechanism for asset concealment (PAPER TRAIL Project, 2026a). Structured payments for suspect disbursements receives P=0.29 -- a related but distinct hypothesis. Tax optimization (P=0.01) and legitimate estate planning (P=0.003) receive negligible posterior mass. The 69% posterior is the lowest of the three, reflecting that the asset concealment and structured payments hypotheses share overlapping evidence, diluting the concentration on either one.
Interpretation Boundaries
These posteriors are model outputs, not ground truth. They depend on three assumptions: the uniform prior, the naive Bayes independence assumption, and the mapping from ACH consistency ratings to conditional probabilities (PAPER TRAIL Project, 2026a). Changing any of these changes the posteriors.
The uniform prior is the most defensible choice -- it encodes no analyst bias. A skeptical prior (lower initial probability for the favored hypothesis) would produce lower posteriors but would not change the ranking. The Monte Carlo simulation confirmed that no parameter perturbation changes which hypothesis ranks first (PAPER TRAIL Project, 2026b).
The naive Bayes assumption is a known simplification. If evidence items are positively correlated -- which they likely are, since many derive from the same underlying documents -- the true posterior may be lower than the naive Bayes estimate. The BBN posteriors should therefore be read as upper bounds on hypothesis probability given the evidence, not as calibrated probabilities.
What the BBN adds to the ACH is specificity. Instead of "willful blindness is more consistent with the evidence than its alternatives," the BBN says "given uniform priors and the observed evidence, there is a 77% probability that willful blindness explains the pattern." That number, with its stated assumptions, is more useful for downstream decision-making than an ordinal rank.
References
Heuer, R. J. (1999). Psychology of intelligence analysis. Center for the Study of Intelligence, Central Intelligence Agency.
New York Department of Financial Services. (2020). Consent order: Deutsche Bank AG [Regulatory filing].
PAPER TRAIL Project. (2026a). BBN posterior estimates [Data set]. _exports/synthesis/bbn_posteriors.csv
PAPER TRAIL Project. (2026b). Monte Carlo ACH robustness results [Data set]. _exports/synthesis/mc_ach_robustness.csv
PAPER TRAIL Project. (2026c). BBN sensitivity analysis [Data set]. _exports/synthesis/bbn_sensitivity.csv
PAPER TRAIL Project. (2026d). Cross-domain synthesis engine: BBN module [Software]. Script 25b, app/scripts/25_cross_domain_synthesis.py
This investigation is part of the SubThesis accountability journalism network.