---
title: "Does the Pathfinder Part I paper verbatim support the rule-based → simple-Bayes switch and the expert's accuracy verdict — and where are the magnitude figures?"
type: "question"
status: "answered"
date_raised: "2026-08-10T00:00:00.000Z"
writer_model: "claude-opus-4-8"
tags: ["pathfinder","david-heckerman","eric-horvitz","bayesian-networks","medical-diagnosis","sourcing-floor","verification","ai-history"]
verifier_log: "2026-08-11 — answered by seek_verify: source_quote(s) of [[claim-pathfinder-rebuilt-from-rules-to-bayesian-networks]] matched verbatim against a direct fetch of the primary. Mechanical check only; judgment untouched."
---


[[claim-pathfinder-rebuilt-from-rules-to-bayesian-networks]] carries two flags. The before/after mechanism is `[unverified-mechanism -- needs primary]` (reached this session only via fetch-and-summarize, tools down). The accuracy **magnitude** is `[unverified-quant -- needs primary]` and is *not asserted* in the note — the Part I paper states the improvement only qualitatively, and the numbers reportedly live in an unlocated companion paper.

**What would answer it.**
- Heckerman, Horvitz & Nathwani, "Toward Normative Expert Systems: Part I. The Pathfinder Project," *Methods of Information in Medicine* 31(2):90–105 (1992), self-hosted at https://erichorvitz.com/Toward_Normative_Systems_MIM.pdf — direct read to receipt: the "fixed traversal through the rule base" account; the silent Dempster–Shafer-Barnett → simple-Bayes switch; and "the expert exclaimed excitedly that the diagnostic accuracy of the program had improved significantly."
- For the magnitude only: locate and read Heckerman & Nathwani, "An evaluation of the diagnostic accuracy of Pathfinder," *Computers and Biomedical Research* (1992, "in press" in the Part I references). This is the primary that would let a *quantitative* accuracy claim clear the Tier 1–2 floor — until then no percentage enters the vault.

**Why it matters.** Pathfinder is the empirical before/after case for the whole bridge — the same diagnostic task solved by rule-chaining then by [[claim-statistical-inference-meaning|statistical inference]], a second instance alongside [[claim-valtorta-reformulated-ach-as-bayesian-networks]]. The mechanism (the switch, the blind expert verdict) is the load-bearing part; the number is owed but secondary.

**Next move.** Re-fetch Horvitz's PDF via extract_pdf once tooling clears and lift the mechanism quotes; separately hunt the companion paper before any accuracy figure is recorded. On mechanism success upgrade past `seedling`; keep `[unverified-quant]` until the companion is read.
</content>