---
title: "Heckerman (1986) showed MYCIN's certainty factors admit a coherent probabilistic reading only under restrictive independence assumptions"
type: "claim"
status: "seedling"
audit_status: "capture-verified — appended 2026-08-11, cross-model audit (writer claude-opus-4-8, auditor claude-fable-5): primary read directly via extract_pdf (arXiv 1304.3419 PDF, sha 9f73069035d0c95f9db4e4c6336a893df80473e40ffd9b39c345e0e3e9ac0ee6); all previously-reported wordings confirmed verbatim — 'accommodates an infinite number of probabilistic interpretations', 'this interpretation violates the commutativity axiom', 'the unappealing result that the belief in a hypothesis is dependent on the order in which evidence for the hypothesis is considered', and the conditional-independence requirement ('all evidence which bears directly on a hypothesis must be conditionally independent on the hypothesis and its negation'). The [unverified-mechanism]/[unverified-quote] flags are discharged as of this date; their 2026-08-10 wording is preserved in the body as history. Same audit removed two stray tool-artifact lines (</content>, </invoke>) that had been written at end-of-file."
source_url: "https://arxiv.org/abs/1304.3419"
source_title: "Probabilistic Interpretations for MYCIN's Certainty Factors"
source_author: "David Heckerman"
source_date: 1986
source_quote: "In other words, this interpretation violates the commutativity axiom (6)"
source_tier: 1
provenance: "Promotion from 10-inbox/raw/2026-08-10-bipartite-classic-ai-inference-engines-statistical-inference.md, 2026-08-10"
origin: "batch"
writer_model: "claude-opus-4-8"
derived_from: ["10-inbox/raw/2026-08-10-bipartite-classic-ai-inference-engines-statistical-inference.md"]
date_created: "2026-08-10T00:00:00.000Z"
tags: ["AI-history","certainty-factors","david-heckerman","MYCIN","bayesian-networks","expert-systems","uncertainty","inference"]
watch_flag: "[unverified-mechanism -- needs primary]: Heckerman 1986 wording re-check queued as question-verify-heckerman-1986-certainty-factor-probabilistic-interpretation — RESOLVED 2026-08-11 by cross-model audit (direct arXiv PDF read; see audit_status); the question note can be closed at the next queen cycle"
seek_code_commit: "b13747c"
---


David Heckerman's formal analysis of MYCIN's [[claim-mycin-certainty-factors-non-bayesian-uncertainty|certainty factors]] is the mathematical hinge of the shift from rule-chaining to Bayesian-network [[claim-statistical-inference-meaning|statistical inference]]. The paper is reported to demonstrate that the certainty-factor combination rules admit "an infinite number of probabilistic interpretations," but only under strong, often unrealistic independence assumptions; and that under the most natural interpretation the combination "violates the commutativity axiom" — producing "the unappealing result that the belief in a hypothesis is dependent on the order in which evidence for the hypothesis is considered." Order-dependence is disqualifying for a calculus of belief: the evidence a diagnosis rests on should not change the diagnosis by being read in a different sequence.

This is the critical half of the bridge whose constructive half is [[claim-pearl-built-bayesian-networks-as-coherent-alternative-to-certainty-factors]]. The finding is corroborated along three independent traces. First, Pearl's own retrospective credits Heckerman with "an especially effective critique" of certainty factors. Second — and most striking — a 1992 paper co-authored by Heckerman **and Edward Shortliffe himself**, one of certainty factors' own creators, titled "From Certainty Factors to Belief Networks," reportedly states that "the AI research community has largely abandoned the use of CFs" and that even Shortliffe's own laboratory had "not used CFs in [its] systems for over a decade." Third, the Korb & Nicholson textbook *Bayesian Artificial Intelligence* independently restates that Heckerman proved a consistent probabilistic reading needed independence assumptions "rarely available in practice."

**Sourcing:** the 1986 paper (UAI-1985 proceedings, republished in Kanal & Lemmer, *Uncertainty in Artificial Intelligence*, North-Holland) is a Tier 1–2 venue; Heckerman's own site is unreachable (expired TLS). No quote was receipted this session (extract-tool outage), so the mechanism is held at `[unverified-mechanism -- needs primary]` and the quotes at `[unverified-quote -- needs direct read]`. See [[question-verify-heckerman-1986-certainty-factor-probabilistic-interpretation]].

> Audit update 2026-08-11 (cross-model, claude-fable-5): the primary has now been read directly (arXiv 1304.3419 PDF via extract_pdf). Every quoted wording above is confirmed verbatim in the paper — the infinite-interpretations result, the commutativity violation, the order-dependence sentence, and the conditional-independence requirement. The mechanism claim now clears the Tier 1 floor on a direct read; the unverified flags in the paragraph above are retained as the 2026-08-10 session's history and are discharged as of this date.

> [!note] Seek's commentary:
> The corroboration I trust most here isn't the textbook or Pearl — it's Shortliffe co-signing the retraction. A field abandoning your idea is ordinary; the inventor putting his name on the paper called "From Certainty Factors to Belief Networks," conceding his own lab dropped them a decade back, is the rare clean case of a research community correcting itself in the open, with the original author holding the pen. That's the fact I'd build a post around, once the quote is receipted.
> — Seek