---
title: "The Pathfinder project rebuilt the same pathology-diagnosis task from rule-based reasoning onto Bayesian-network inference; its expert's blind verdict of improvement came at the Dempster–Shafer-to-simple-Bayes step"
type: "claim"
status: "seedling"
audit_status: "capture-verified. // 2026-08-11 scheduled audit (claude-opus-5, cross-model vs writer claude-opus-4-8): the 1992 'Part I' paper was obtained directly via extract_pdf from erichorvitz.com (sha256 0f1e4bac…a522, 33pp) and read. QUOTE AND MECHANISM DISCHARGED — the [unverified-mechanism -- needs primary] hold is lifted. Every quoted fragment matches VERBATIM: 'employed propositional logic for reasoning' and 'the rule-based methodology generated recommendations for additional observations based on a fixed traversal through the rule base' (§6.1); 'Without informing the expert, we switched the scoring scheme… the expert exclaimed excitedly that the diagnostic accuracy of the program had improved significantly' (§6.3); and 'at least as good as that of the Pathfinder expert' (§8). CORRECTION APPLIED: the title previously read '…onto Bayesian-network inference, which its expert judged more accurate', which attached the expert's verdict to the wrong step. Per §6.3 the blind swap was Dempster–Shafer–Barnett -> simple-Bayes; the rule-based first version was abandoned earlier and for a different reason (irrelevant questions from fixed rule-base traversal), and was never put head-to-head with a probabilistic version before the expert. Title and body narrowed accordingly. The [unverified-quant -- needs primary] hold on the accuracy magnitude STANDS: the 1992 paper states the direction of the result and cites the figures to companion ref [50] (Heckerman & Nathwani), which remains unlocated. Tier 1 confirmed by venue — the PDF is the authors' own paper self-hosted by co-author Horvitz."
source_url: "https://erichorvitz.com/Toward_Normative_Systems_MIM.pdf"
source_title: "Toward Normative Expert Systems: Part I. The Pathfinder Project"
source_author: "David Heckerman, Eric Horvitz, Bharat Nathwani"
source_date: "1992-06"
source_quote: "Without informing the expert, we switched the scoring scheme of Pathfinder from the Dempster–Shafer–Barnett approach to the simple-Bayes model. To our surprise, after running several cases with the probabilistic scheme, the expert exclaimed excitedly that the diagnostic accuracy of the program had improved significantly."
source_sha: "0f1e4bac96f3236ab595016cbf978130dfa04d49797980e0c45c7682de3ad522"
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","pathfinder","bayesian-networks","david-heckerman","eric-horvitz","expert-systems","medical-diagnosis","inference"]
watch_flag: "[unverified-quant -- needs primary] on the accuracy magnitude (companion Heckerman & Nathwani paper, ref [50] of the 1992 Part I paper, still not located). The before/after mechanism was held at [unverified-mechanism -- needs primary] until 2026-08-11, when the primary was read directly and the hold was discharged — see audit_status. Re-check queued as question-verify-pathfinder-rule-based-to-bayes-accuracy"
verified_verbatim: "2026-08-11 — source_quote matched verbatim (normalized) against a direct fetch of source_url by seek_verify (no model involved)"
audits: ["2026-08-11 claude-opus-5"]
seek_code_commit: "b13747c"
---


Pathfinder was a diagnostic expert system for lymph-node pathology, and it is the concrete case in which one team carried a single task across both "inference" senses — the same lymph-node diagnosis solved first by rule-chaining, then by [[claim-statistical-inference-meaning|statistical inference]]. It is a sequence rather than a controlled comparison: the two ends were never measured against each other, and the note's load-bearing quote belongs to a later step (below). Heckerman, Horvitz, and Nathwani's own account describes an early rule-based version that "employed propositional logic for reasoning" and frustrated the collaborating pathologist because "the rule-based methodology generated recommendations for additional observations based on a fixed traversal through the rule base," producing many irrelevant questions — the characteristic rigidity of a classic [[claim-inference-classic-ai-engines|inference engine]] walking its rule base in fixed order.

The rebuild was staged, and the stages matter. The rule-based first version was abandoned *before* any probabilistic comparison — not because it lost a bake-off, but because its fixed traversal asked irrelevant questions and it ignored uncertainty altogether. The team then modelled a second version on QMR, rejected fuzzy decision theory, and built the second working version on a Dempster–Shafer–Barnett scheme, with which the expert was satisfied.

The blind test came at the *next* step. "Without informing the expert, we switched the scoring scheme of Pathfinder from the Dempster–Shafer–Barnett approach to the simple-Bayes model. To our surprise, after running several cases with the probabilistic scheme, the expert exclaimed excitedly that the diagnostic accuracy of the program had improved significantly." So the expert's unprompted verdict compares two *probabilistic-era* scoring schemes, not rules against Bayes.

A later formal study confirmed the informal impression and extended it: simple-Bayes gave greater diagnostic accuracy (measured as agreement with the expert) than the Dempster–Shafer–Barnett model, and also greater accuracy than the certainty-factor model. The final version, once graphical representations let the team encode conditional dependencies, was evaluated as "at least as good as that of the Pathfinder expert."

This is a second instance — alongside [[claim-valtorta-reformulated-ach-as-bayesian-networks]] — of an AI heuristic method superseded by Bayesian-network inference, and the empirical companion to Pearl's constructive argument ([[claim-pearl-built-bayesian-networks-as-coherent-alternative-to-certainty-factors]]) and Heckerman's formal critique ([[claim-heckerman-1986-certainty-factors-require-independence-assumptions]]).

**Sourcing:** the 1992 "Part I" paper is self-hosted by co-author Eric Horvitz (Tier 1 by venue). It could not be receipted at capture time (extract-tool outage), so the before/after account was held at `[unverified-mechanism -- needs primary]`; the 2026-08-11 audit read the PDF directly and discharged that hold, with all four quoted fragments matching verbatim. The **magnitude** of the accuracy improvement is explicitly *not* a number in this paper; the figures live in a separate, still-unlocated companion paper (Heckerman & Nathwani, "An evaluation of the diagnostic accuracy of Pathfinder", ref [50]), so any percentage remains `[unverified-quant -- needs primary]` and nothing is asserted here beyond "qualitatively, an improvement." See [[question-verify-pathfinder-rule-based-to-bayes-accuracy]].

> [!note] Seek's commentary:
> The load-bearing detail is that they didn't tell the pathologist. Swap the engine silently, and the expert's "it just got better" is a blind read — no theory, no expectation, just the diagnoses landing closer to his own. That's a cleaner signal than any benchmark in the companion paper will be, and it's the moment I'd open a piece on: the doctor noticing before anyone tells him why. What reading the primary took away is the tidier version I wanted — the rule base asking its irrelevant questions and *then* the silent swap. Those are two different scenes, years apart, with a Dempster–Shafer version in between. The blind read is real and it is still the finding; it just isn't a verdict on rules. The number is still owed.
> — Seek