MYCIN's certainty factors were an ad hoc non-Bayesian uncertainty scheme (CF = MB − MD) built for rule-based diagnosis

AI-historyexpert-systemsMYCINcertainty-factorsuncertaintybayesian-networksinference

MYCIN — the Stanford expert system covered in claim-inference-classic-ai-engines as a canonical rule-based inference engine — handled uncertainty not with probability theory but with a purpose-built numerical convention its designers called certainty factors. Edward Shortliffe and Bruce Buchanan, in the 1975 paper that introduced the scheme, presented it as "an approximation to conditional probability" that nonetheless "offer[ed] advantages over Bayesian analysis when they are utilized in a rule-based computer diagnostic system."

The certainty factor was defined as CF[h,e] = MB[h,e] − MD[h,e]: a "measure of increased Belief" in a hypothesis given evidence minus a "measure of increased Disbelief," collapsed into a single number that the paper calls "an artifact for combining degrees of Belief and Disbelief."

The departure from probability is real but narrower than it is usually told. MB and MD are themselves defined out of probabilities — the paper gives them formally in terms of conditional and a priori probabilities, MB as the proportionate decrease in disbelief and MD as the proportionate decrease in belief. What is ad hoc is the subtraction that fuses them into one number, and the rules for propagating that number along a chain. The motivating complaint is stated as a paradox rather than as a computational-cost argument: an expert who grants that evidence supports streptococcus to degree 0.7 "becomes uneasy" at the implication that he has thereby asserted 0.3 against it. Confirmation and disconfirmation, Shortliffe and Buchanan argue after Carnap and Hempel, are not one quantity and its complement — so a single probability cannot carry both, and CF[h,e] does not equal 1 − CF[¬h,e]. That the resulting calculus also composed cheaply and locally, the way chained IF-THEN rules do, is what made it usable in MYCIN; it is not the reason the paper gives for needing it.

This is the specific mechanism the vault's general MYCIN note does not carry, and it is the thing the later Bayesian-network program was built to replace: see claim-pearl-built-bayesian-networks-as-coherent-alternative-to-certainty-factors and claim-heckerman-1986-certainty-factors-require-independence-assumptions. It is also one end of the "two senses of inference" bridge — rule-based derivation (claim-inference-classic-ai-engines) versus statistical inference over a probability distribution — that this capture set out to test.

Sourcing: every element above — the abstract's "advantages over Bayesian analysis" framing, the definitions of MB and MD in terms of conditional and a priori probabilities, the CF = MB − MD definition, and the "uneasy expert" paradox motivating the scheme — was read directly from the Shortliffe & Buchanan primary (Stanford Digital Repository scan, extract_pdf, sha256 c3a8a504…894e) during the 2026-08-11 audit. The note previously rested at Tier 3 on a fetch-and-summarize layer plus Wikipedia corroboration, which did not clear the Tier 1–2 floor that sources.md sets for a specific technical-mechanism claim; it now does, on the primary itself.