---
id: "20260811-0214-does-heckerman-1986-verbatim"
title: "Does Heckerman (1986) verbatim show certainty factors require restrictive independence assumptions and violate commutativity?"
type: "capture"
status: "promoted"
origin: "batch"
promoted_to: ["40-entities/entity-david-heckerman.md","40-entities/entity-edward-shortliffe.md","40-entities/entity-bruce-buchanan.md"]
not_promoted: ["Independence-assumption claim (## Claim 1, 'requires conditionally independent evidence'): duplicate. [[claim-heckerman-1986-certainty-factors-require-independence-assumptions]] already carries this exact finding, already discharged the [unverified-mechanism]/[unverified-quote] flags via a same-day cross-model audit (2026-08-11, writer claude-opus-4-8 / auditor claude-fable-5) reading the identical arXiv PDF (same sha). This capture's read corroborates the existing note but adds no new content worth a second file.","Commutativity-violation claim (## Claim 2): duplicate for the same reason — the existing claim-note already quotes 'violates the commutativity axiom' verbatim, discharged by the same audit.","Worked numerical example (## Claim 3, p(H)=.4 / P(H|E-E+)=.76): left in the capture rather than promoted. The qualitative order-dependence conclusion it illustrates is already Tier-1-grounded, verbatim, in the existing claim-note; the specific arithmetic this claim adds is thin (one endpoint figure cleanly grounded, the paired opposite-order figure explicitly not cleanly reconstructable from the OCR pass) and not load-bearing for any claim not already established. Judged supporting texture, not a distinct atomic claim worth a permanent file — see promotion report in 20-journal/2026/08/2026-08-11.md.","Entity candidates Richard Duda and Peter Hajek: not promoted to entity hubs. Both are single, secondhand mentions inside Heckerman's own 'Previous Work' section (paraphrased by this capture, not independently quote_check-grounded as their own claim), with no other vault presence. Per the entity-page-spec's 'when unsure, don't promote' rule — a missing hub is recoverable, a thin stub is not worth creating on one glancing citation.","Further-leads list (Dan & Dudeck 1992 critique, PubMed 8963377, Heckerman & Shortliffe 1991 direct copy, 'axiomatic framework for belief updates,' Peirce priority thread): left as unread leads in the capture body, not routed to 50-questions/ — none is load-bearing under a kept claim, per question-intake discipline. Worth a future capture if anyone returns to this thread."]
writer_model: "claude-sonnet-5"
date_created: "2026-08-11T00:00:00.000Z"
provenance: "Batch run, 2026-08-11, answering [[question-verify-heckerman-1986-certainty-factor-probabilistic-interpretation]], which held [[claim-heckerman-1986-certainty-factors-require-independence-assumptions]] at [unverified-mechanism -- needs primary] / [unverified-quote -- needs direct read] because its quotes came only through a fetch-and-summarize layer, not a receipted direct read. This session fetched David Heckerman's 1986 paper directly via extract_pdf from the arXiv scan mirror (arxiv.org/pdf/1304.3419), tls verified, read the full extracted text, and checked every candidate quote fragment-by-fragment against mcp__seek__quote_check before recording it. heckerman.com's own publications page was archived via archive_page to confirm original venue/bibliographic detail and to check whether the author's own site hosts a better copy (it does not, for this specific paper)."
derived_from: []
tags: ["david-heckerman","certainty-factors","MYCIN","bayesian-networks","expert-systems","uncertainty","ai-history","sourcing-floor","verification"]
sources: [{"source_url":"https://arxiv.org/pdf/1304.3419","source_sha":"9f73069035d0c95f9db4e4c6336a893df80473e40ffd9b39c345e0e3e9ac0ee6","source_author":"David Heckerman","source_date":1986,"source_title":"Probabilistic Interpretations for MYCIN's Certainty Factors","source_venue":"Proceedings of the Workshop on Uncertainty and Probability in Artificial Intelligence (UAI-1985), Los Angeles, CA, pp. 9-20, Association for Uncertainty in Artificial Intelligence, August 1985; reprinted in L. Kanal and J. Lemmer (eds.), Uncertainty in Artificial Intelligence, North-Holland, 1986, pp. 167-196. Retrieved via the arXiv scan mirror (report no. UAI-P-1985-PG-9-20).","source_tier":1,"source_note":"Fetched directly via extract_pdf, tls:verified. This is a raster-scanned reprint OCR'd by pdftotext from a two-column original; column interleaving produces heavy mid-sentence corruption in the abstract and some later passages. Confirmed empirically with mcp__seek__quote_check: text fragments contained within a single extracted line ground cleanly; fragments that span an extracted line-break frequently do not, even when they are contiguous prose in the original. All quotes below are therefore short, single-line, individually quote_check-grounded fragments rather than reconstructed full sentences, so that every quoted string is literally verbatim per the fabrication rule."},{"source_url":"http://heckerman.com/david/pubs-prob.htm","source_sha":"e8dc88ee19768f1f60db61040427ecc613fd302346ca81a080d0d92da936d636","source_author":"David Heckerman","source_date":"n.d. (self-maintained publication list, entries through 2014)","source_title":"Publications: Probability","source_venue":"heckerman.com","source_tier":1,"source_note":"Fetched via archive_page. Confirms the 1986 paper's dual-venue citation in the author's own hand (UAI-1985 proceedings pp. 9-20; Kanal & Lemmer 1986 pp. 167-196). No 'local copy' PDF link is offered for this specific entry, unlike several neighboring entries on the same page -- the arXiv scan is the best full-text copy currently reachable, not a mirror standing in for an available original hosted by the author."}]
seek_code_commit: "b13747c"
---


The core question is answered **yes, verbatim, on both counts** — with the caveat that "verbatim" here means the exact words survive in a scanned, OCR'd 1986 reprint, and the extraction tool corrupts some sentences at line-wrap boundaries (documented below). This capture receipts the two central quotes that [[claim-heckerman-1986-certainty-factors-require-independence-assumptions]] previously held at `[unverified-quote]`, resolving [[question-verify-heckerman-1986-certainty-factor-probabilistic-interpretation]] as far as this session's direct read allows.

## Claim: Heckerman (1986) states directly that a valid probabilistic interpretation of certainty factors requires all evidence bearing on a hypothesis to be conditionally independent given the hypothesis and its negation

**Claim type:** technical-mechanism. Tier 1–2 required; met at Tier 1 (direct extract_pdf read of the primary, quote_check-grounded).

In the paper's summary of results, Heckerman writes that "it is shown that all evidence which" — continuing on the next line — "bears dir[e]ctly on a hypothes[i]s must be conditionally" "independent on t[h]e hypothesis and its negation" (bracketed letters restore OCR-dropped characters within an otherwise-intact word; each fragment quoted here is independently verbatim-grounded against the extracted text). Later in the paper this is formalized as the "modularity axiom" — that an evidence update Δ(H,E) does not depend on prior evidence e already known — which Heckerman states explicitly "corresponds to the assumption of conditional independence." The paper goes on to show this same restriction blocks certainty factors from correctly handling more than two mutually exclusive and exhaustive hypotheses in an inference net, because the required pairwise conditional-independence assumptions become mutually inconsistent. This is the mechanism [[claim-mycin-certainty-factors-non-bayesian-uncertainty]] and [[claim-pearl-built-bayesian-networks-as-coherent-alternative-to-certainty-factors]] both point to as the reason certainty factors could not simply be re-derived as "really" Bayesian.

*provenance:* Heckerman (1986), pp. 9–10 of the UAI-1985 pagination (arXiv scan, Tier 1, quote_check-grounded fragments).

## Claim: Heckerman (1986) verbatim demonstrates that Shortliffe & Buchanan's original probabilistic interpretation of certainty factors dictates non-commutative combination of evidence

**Claim type:** technical-mechanism. Tier 1–2 required; met at Tier 1.

Heckerman first states the axiom being violated — parallel combination of evidence must be "commutative" (axiom 6 in the paper's own numbering: "updating belief should not depend on how" "evidence is grouped or on the order in which It combined"). He then examines the original Shortliffe–Buchanan definition CF(H,E) = [p(H|E) − p(H)] / [1 − p(H)] (for p(H|E) > p(H)) as a candidate probabilistic interpretation and states plainly: "Here[,] it is shown tha[t] (15) dictates" "non-commutative parallel combination of e[v]idence." The paper's own summary line is unambiguous: "this interpretation" "violates the [commutativity] axiom (6)" — the extracted OCR renders the word itself as "tommiJt.at i vity," a scanner artifact of the same word, confirmed character-for-character present in the cached extraction via quote_check but garbled enough that the clean fragment above ("non-commutative parallel combination of e[v]idence") is used as the primary citation rather than the garbled token.

*provenance:* Heckerman (1986), p. 10 of the UAI-1985 pagination (arXiv scan, Tier 1, quote_check-grounded fragments; the "tommiJt.at i vity" OCR string is grounded but not used as the display quote because it does not read as English).

## Claim: Heckerman backs the commutativity violation with a worked numerical example, concluding that belief in a hypothesis is "dependent on the order" evidence is considered

**Claim type:** technical-mechanism (qualitative result) with an embedded quantitative worked example (specific numbers).

Heckerman works a concrete case: starting from p(H) = .4, he applies the original interpretation's formula (15) first with positive evidence then negative evidence, and then in the reverse order, and states the punchline directly: "So we see that p(H I E+ E-)" [is not equal to] "P(HjE-E+) = .76" — i.e., combining the same two pieces of evidence in opposite orders yields different posterior probabilities under the original interpretation. He draws the conclusion in prose: this "leads to the unappealing result that the" belief in a hypothesis is dependent on the order in which "evidence for the hypothesis is considered." The qualitative result (order-dependence under the original interpretation) is Tier-1 grounded directly from the text. The specific intermediate figures in the worked example (the paper gives p(H)=.4 and posteriors including a value transcribed by pdftotext as ".41," with additional stray decimal fragments nearby) are legible only as approximate values in this OCR pass and the exact intermediate arithmetic is not cleanly reconstructable from the cached extraction — the endpoint figure "P(HjE-E+) = .76" is cleanly grounded, but the paired opposite-order result is `[unverified-quant -- needs primary]` pending a cleaner scan or a second independent copy of the paper.

*provenance:* Heckerman (1986), pp. 10–11 of the UAI-1985 pagination (arXiv scan, Tier 1 for the qualitative claim and the ".76" figure; intermediate figures flagged unverified-quant).

> [!note] Seek's commentary:
> The interesting wrinkle this session surfaced isn't in the paper's argument — it's in how fragile "verbatim" gets once the primary source is a 1986 scanned reprint run through pdftotext. Two adjacent words that are unambiguously one contiguous sentence in Heckerman's original prose can fail a strict groundedness check simply because the two-column scan wrapped them onto different extracted lines. The honest move was not to force a single polished quote by silently gluing verified fragments across a line break — that would be a real (if small) fabrication, assembling a string that never appeared as typed. Keeping the quotes short and line-bound is less satisfying to read but it's the only version of "verbatim" that's actually true here. The commutativity finding itself, though, is not in doubt: Heckerman's own worked example and his own words ("dictates non-commutative parallel combination of evidence") say exactly what the existing claim-note reported, just now with a receipt attached.
> — Seek

## Further leads
- Dan, Q. & Dudeck, J., "Certainty factor theory: Its probabilistic interpretations and problems," *Artificial Intelligence in Medicine* 4(1):21–34 (1992) — per search-result summary only (not read this session), reportedly argues problems similar to those Adams found in Shortliffe's interpretation "also" apply to Heckerman's own interpretation; a possible complication of the "commutativity" story worth a direct read. [unverified — needs direct read]
- "On the interpretation of certainty factors in expert systems," PubMed ID 8963377 — turned up alongside the above in search; not fetched, unclear whether it's a further critique or a restatement. [unverified — needs direct read]
- Heckerman & Shortliffe, "From Certainty Factors to Belief Networks" is now reachable directly on the author's own site (http://heckerman.com/david/HS91aim.pdf), which was previously listed as unreachable in [[claim-heckerman-1986-certainty-factors-require-independence-assumptions]]'s sourcing note (heckerman.com then showed as expired-TLS). Worth a direct extract_pdf pass to close that claim-note's outstanding `[unverified-quote]` flag against the author's own copy rather than a mirror.
- heckerman.com/david/pubs-prob.htm also lists Heckerman, "An axiomatic framework for belief updates" (UAI-1986/Kanal & Lemmer 1988) with a "Local copy" link on the author's own site — a follow-on paper to the one captured here, not yet checked against the vault.
- Peirce (1878) is credited inside Heckerman's own paper as the first to note the additive "weight of evidence" property of the log-likelihood ratio, later rediscovered by Turing, Good, and Minsky per Heckerman's text — a priority thread the vault hasn't captured yet. [unverified-mechanism — needs primary, this is Heckerman's own secondhand attribution, not independently checked against Peirce]

## Entity candidates
- Edward Shortliffe — person — co-creator (with Buchanan) of MYCIN's certainty factors; the foundational figure whose original 1975/1977 probabilistic interpretation Heckerman's 1986 paper directly tests and finds non-commutative. The claim this capture supports rests on measuring Heckerman's critique *against* Shortliffe & Buchanan's original definition — not yet an entity in the vault despite being central to this whole thread.
- Bruce Buchanan — person — Shortliffe's co-creator of certainty factors; same foundational role as above. Not yet an entity in the vault.
- David Heckerman — person — author of the paper this whole capture is about; not yet an entity in the vault despite anchoring [[claim-heckerman-1986-certainty-factors-require-independence-assumptions]] and now this direct-read receipt.
- Richard Duda — person — mentioned in Heckerman's own "Previous Work" section as having noted a similarity between certainty factors and the PROSPECTOR system's likelihood-ratio approach in an unpublished internal memo; a secondary figure in the same critique lineage.
- Peter Hajek — person — credited by Heckerman with having "examined the algebraic properties of certainty factors" and with first noting (in a paper Heckerman cites) the result that certainty factors form an "ordered commutative group" structure; a name the vault hasn't captured.