Judea Pearl built Bayesian networks as a probabilistically coherent alternative to expert systems' certainty factors

AI-historybayesian-networksjudea-pearlcertainty-factorsexpert-systemsprobabilistic-graphical-modelscross-domain-bridgeinference

In his own retrospective technical report, Judea Pearl describes the combination rules behind MYCIN's certainty factors as "not based on probability theory or any other principled methodology," and argues that their unprincipled character made them fail on basic evidential phenomena — notably explaining away, where confirming one cause of an observed effect should lower belief in an alternative cause of the same effect. Pearl states that in 1982 he proposed an efficient belief-updating algorithm running Bayesian probabilities over structured networks, an approach he named Bayesian networks by 1985, motivated by explicit design goals including "coherence with orthodox probability theory and Bayesian reasoning."

This is the constructive half of the bridge whose critical half is claim-heckerman-1986-certainty-factors-require-independence-assumptions: where Heckerman proved the old scheme incoherent, Pearl built the replacement. The program kept the expert-system architecture — a structured knowledge base plus an inference procedure (claim-inference-classic-ai-engines) — but swapped the ad hoc uncertainty arithmetic for genuine statistical inference over a probability distribution. It is the documented mechanism that makes the vault's "two senses of inference" pairing a real connection rather than an embedding false friend (contrast observation-ml-ad-en-gedi-cosine-pairing-is-embedding-false-friend).

The single most concrete piece of evidence needs no internal quote: Pearl's 1982 paper is consistently cited as titled "Reverend Bayes on Inference Engines: A Distributed Hierarchical Approach" — a title that textually fuses "Bayes" (statistical inference) with "inference engines" (the classic expert-system term) in one document. That the same pattern recurs in an unrelated domain — intelligence analysts recast ACH as Bayesian networks, not rule-chaining — suggests a general shape: AI heuristic methods superseded by Bayesian-network statistical inference.

Sourcing: the source is Pearl's own self-hosted report (Tier 1 by venue; a cut section of The Book of Why, cf. claim-atheist-bible-truth-chapter-cites-pearl-batens-chaitin), and two independent research passes converged on identical quote wording. But convergence across fetch-and-summarize layers is not a direct read, and no receipt exists this session, so the mechanism is held at [unverified-mechanism -- needs primary]; the specific quotes are [unverified-quote -- needs direct read]. See question-verify-pearl-r476-bayesian-networks-vs-certainty-factors.

Audit 2026-08-11 (cross-model, claude-fable-5): R-476 fetched and read directly via extract_pdf (sha256 c9b54ac4…, 9 pp.); the paragraph above is the capture-time record, retained unedited. All three receipts appear verbatim on p. 2 of the report: "These rules of combination, however, were not based on probability theory or any other principled methodology, and therefore tended to produce unintended results. In particular, they could not exhibit the 'explain away' effect"; the design desiderata including "to insist on coherence with orthodox probability theory and Bayesian reasoning"; and "In 1982, I entered the arena with a paper showing an efficient belief-updating algorithm using Bayesian probabilities, an approach that I would eventually (1985) call 'Bayesian networks.'" The title "Reverend Bayes on Inference Engines: A Distributed Hierarchical Approach" is given in Pearl's own text, no longer resting on "consistently cited as." The [unverified-*] flags are discharged.