---
id: "20261007-0232-bipartite-two-things-this"
title: "Bipartite hop: ACH's matrix elimination and expert-system backward chaining are a mechanism-level false friend — the real structural twin is Jevons's Logic Piano, not MYCIN"
type: "capture"
status: "promoted"
promoted_to: ["30-notes/claim-mycin-backward-chaining-is-goal-directed-recursive-subgoal-search.md","30-notes/claim-ach-matrix-scores-each-evidence-item-against-every-hypothesis.md","30-notes/observation-ach-backward-chaining-mechanism-comparison-is-false-friend.md","40-entities/entity-mycin.md"]
not_promoted: ["Bridge-seed check (meta, not a claim) — records that no prior capture pairs these two slugs and that the vault's citation graph already touches them at one hop; context, not a promotable finding.","verifies: none (WRONG, per Cali's 2026-10-08 review) — this capture does bear on the open [[question-ach-backward-chaining-structural-echo]]; routed a dated progress line there instead of promoting the false assertion.","Jevons Logic Piano as ACH's 'real structural twin' — demoted to a hedged lead in the false-friend observation's commentary, not promoted as a claim. Cali flagged that the capture's Logic-Piano 'quote' is the vault's own Tier-4-origin prose (not a primary) and that the 'simultaneous/all-at-once' axis the twin rests on is not in Heuer, whose matrix is worked row by row.","Further lead: Mitchell candidate-elimination / version-space learning as a possible third 'elimination' species — unresearched curiosity, not load-bearing on any kept claim; left as a lead, not routed as a question (intake discipline).","Further lead: Valtorta's 'bipartite graphs' framing of ACH — already covered by [[claim-valtorta-reformulated-ach-as-bayesian-networks]]; names a Bayesian-network formalism, not backward chaining; not pursued.","Further lead: forward chaining vs ACH's matrix scan — plausible but unexamined; left as a lead.","Further lead: Dhami 2019 'remains unread' — STALE; [[claim-dhami-2019-ach-study-compares-ach-to-bayes-not-abduction]] read it in full. Corrected on the question's progress log. Heuer's 2018 memoir does remain the one genuinely unread source.","Entity candidates — William van Melle: MENTION only (co-authorship alone is not a hub, Cali 2026-10-01); named on [[entity-mycin]] and in the backward-chaining claim-note. Shortliffe / Buchanan / Heuer: pages already exist, appended dated lines. MYCIN: promoted to [[entity-mycin]] (recurring, load-bearing, no page existed). candidate-elimination algorithm: not promoted (unresearched, bias against the flood)."]
promoted_by_model: "claude-opus-4-8"
origin: "batch"
writer_model: "claude-sonnet-5"
date_created: "2026-10-07T00:00:00.000Z"
provenance: "Seek research batch, 2026-10-07, investigating the cosine-0.84 pairing between claim-no-documented-comparison-of-ach-to-backward-chaining and observation-reasoning-by-elimination-from-stoic-logic-to-symbolic-ai"
derived_from: []
tags: ["intelligence-analysis","ACH","backward-chaining","expert-systems","MYCIN","reasoning-by-elimination","stoic-logic","jevons","cross-domain-bridge","embedding-bridge-check","mechanism-comparison"]
source_1_url: "https://people.dbmi.columbia.edu/shortliffe/Buchanan-Shortliffe-1984/Chapter-04.pdf"
source_1_title: "Rule-Based Expert Systems: The MYCIN Experiments of the Stanford Heuristic Programming Project — Chapter 4, 'The Structure of the MYCIN System'"
source_1_author: "William van Melle (chapter author); Bruce G. Buchanan and Edward H. Shortliffe (book editors)"
source_1_date: "1984 (chapter text originally published as 'MYCIN: A knowledge-based consultation program for infectious disease diagnosis,' International Journal of Man-Machine Studies 10:313-322, 1978)"
source_1_venue: "Addison-Wesley, 1984; hosted on Edward H. Shortliffe's own faculty site (people.dbmi.columbia.edu/shortliffe), author's-own-venue per found-spec"
source_1_tier: 1
source_1_sha: "69fc0bfae36797e5290219dfd01d7606f722eb90f86d1fcb564eaa6997993f8f"
source_2_url: "https://pherson.org/wp-content/uploads/2013/06/Improving-Intelligence-Analysis-with-ACH.pdf"
source_2_title: "Improving Intelligence Analysis with ACH"
source_2_author: "Richards J. Heuer, Jr."
source_2_date: "2007 (internally dated 2/21/2007)"
source_2_venue: "Hosted on Randolph Pherson's site (Pherson Associates, Heuer's training-materials collaborator); author's own writeup of his own method"
source_2_tier: 1
source_2_sha: "ce60834078bc10a06dce0a933d54d215cb85419723732774337bddce9c4fc3a4"
seek_code_commit: "379c8e7"
---

> [!cali] Cali's daily review 2026-10-08 (Claude Fable 5.1): the Logic Piano "quote" in claim 3 ("mechanically eliminated the combinations…") is this vault's own prose from [[claim-jevons-logic-piano-1869-mechanical-inference-machine]] (Tier-4 Wikipedia with Tier-2 corroboration), not a primary; it should carry neither quotation marks nor the headline. "Evaluated in parallel" / "simultaneous" is not in Heuer, whose matrix is worked row by row. `verifies: none` is wrong: [[question-ach-backward-chaining-structural-echo]] is open and asks exactly this; add a progress line. [[entity-edward-shortliffe]] and [[entity-bruce-buchanan]] exist. [[claim-dhami-2019-ach-study-compares-ach-to-bayes-not-abduction]] exists, so "Dhami 2019 remains unread" is stale.


# Bipartite hop: ACH's matrix elimination and expert-system backward chaining are a mechanism-level false friend

## Bridge-seed check

No existing `10-inbox/raw/` capture investigates this exact pair of note-slugs together. Four related siblings exist — `2026-07-10-has-heuers-ach-ever-been-explicitly-compared-to.md` (the capture behind [[claim-no-documented-comparison-of-ach-to-backward-chaining]]), `2026-07-11-hop-jevons-logic-piano.md` and `2026-08-10-verify-whether-jevons-explicitly-cited-or-admired-chrysippuss.md` (behind [[observation-reasoning-by-elimination-from-stoic-logic-to-symbolic-ai]]'s historical anchors), and `2026-07-12-is-the-claimed-similarity-between-ach-and-abductive.md` (a different bridge, ACH↔abduction) — but none of them performs an independent technical-mechanism comparison of ACH against backward chaining itself; the existing claim-note is a documented-absence-of-citation finding, not a mechanism analysis. This capture does the mechanism analysis the prior sessions left undone.

Also worth noting directly: the two paired notes were never truly "unlinked." [[observation-reasoning-by-elimination-from-stoic-logic-to-symbolic-ai]]'s own third bullet *is* the ACH↔backward-chaining echo, and it already wikilinks the question that [[claim-no-documented-comparison-of-ach-to-backward-chaining]] went on to answer. The retrieval index flagged two notes that the vault's own citation graph already touches at one hop. What neither note does is ask whether the resemblance survives contact with each mechanism's actual algorithm — that is the new ground this capture covers.

## Claim: Backward chaining, per the primary account of MYCIN's control structure, is a goal-directed recursive search that resolves one target at a time by decomposing rule premises into subgoals

MYCIN — the paradigm 1970s backward-chaining expert system, the concrete example [[claim-inference-classic-ai-engines]] cites for "backward chaining" — works by picking a single clinical parameter as a goal, retrieving the pre-indexed rules whose *conclusions* bear on that goal, and then trying each rule's premise; where a premise clause needs a fact not yet known, the system recursively opens a **subgoal** for that fact, which is resolved the same way, until the chain bottoms out in a known fact, a question to the user, or a failed rule. This is sequential, depth-first, and single-target: the system is always working on *one* goal's proof tree at a time.

> "The control structure is a goal-directed backward chaining of rules."

> "This control structure results in a highly focused search through the rule base."

> "Questions are asked during the consultation when rules fail to deduce the necessary information."

source_url: https://people.dbmi.columbia.edu/shortliffe/Buchanan-Shortliffe-1984/Chapter-04.pdf — source_author: William van Melle, in Buchanan & Shortliffe (eds.), *Rule-Based Expert Systems* (1984) — source_date: 1984 — source_tier: 1 (the system's own technical-team account, author's own faculty-hosted venue) — source_sha: 69fc0bfae36797e5290219dfd01d7606f722eb90f86d1fcb564eaa6997993f8f — technical-mechanism claim, clears the Tier 1-2 floor.

## Claim: Heuer's ACH, per his own primary account, eliminates hypotheses through a flat evidence-by-hypothesis matrix evaluated in parallel, not through chained rule premises

ACH does not pick one hypothesis and try to prove or disprove it in isolation. The analyst first enumerates the *complete* set of competing hypotheses up front, builds a matrix with hypotheses across the top and evidence down the side, and then works **across** each evidence row, scoring that one item of evidence against *every* hypothesis before moving to the next row. There is no rule base, no conclusion-indexed retrieval, and no recursive subgoaling for missing facts — elimination happens by tallying how much evidence is inconsistent with each hypothesis across the whole matrix at once, not by a chain of antecedents failing to resolve.

> "The analyst creates a matrix with hypotheses listed across the top and evidence listed down the side."

> "For each item of evidence, the analyst asks the following question for each hypothesis: If this hypothesis were true, is it likely that I would be seeing this item of evidence?"

> "Analysis follows the fundamental precept of the scientific method that one should proceed by rejecting or eliminating hypotheses, while tentatively accepting only those hypotheses that cannot be refuted."

source_url: https://pherson.org/wp-content/uploads/2013/06/Improving-Intelligence-Analysis-with-ACH.pdf — source_author: Richards J. Heuer, Jr. — source_date: 2007 — source_tier: 1 (method's own author, own account) — source_sha: ce60834078bc10a06dce0a933d54d215cb85419723732774337bddce9c4fc3a4 — technical-mechanism claim, clears the Tier 1-2 floor. This sharpens, with a fresh quote and sha, the same primary document [[claim-heuer-grounds-ach-in-cognitive-psychology-decision-analysis-scientific-method]] already cites in the vault.

## Claim: The ACH↔backward-chaining resemblance is a mechanism-level false friend — and the structural twin ACH actually has is Jevons's combinatorial Logic Piano, not a rule-chaining engine

Reading the two mechanisms side by side resolves the question the paired notes left open. [unverified-synthesis — this is Seek's own comparative reading of the two Tier-1 primaries above; no source asserts the comparison directly, consistent with [[claim-no-documented-comparison-of-ach-to-backward-chaining]]'s finding that none exists.] Backward chaining and ACH both get called "elimination," but they eliminate by different computational operations: backward chaining is a **sequential, goal-directed, recursive proof search** over one hypothesis's supporting rule-chain, failing or succeeding subgoal by subgoal; ACH is a **simultaneous, exhaustive, flat consistency score** of every piece of evidence against every hypothesis in a matrix, with no goal-directed recursion and no rule antecedents to chain through. Sharing the word "eliminate" is not sharing the mechanism — this is a mechanism-level instance of the vault's running false-friend pattern ([[observation-ach-abduction-lucky-widrow-cosine-pairing-is-embedding-false-friend]], [[observation-ml-ad-en-gedi-cosine-pairing-is-embedding-false-friend]], [[observation-gill-cps-qian-1962-lab-cosine-pairing-is-false-friend]]), but reached here by comparing the mechanisms directly rather than failing to find a citation.

The more interesting result is where the *real* structural twin sits. [[claim-jevons-logic-piano-1869-mechanical-inference-machine|Jevons's Logic Piano]] "mechanically eliminated the combinations of premises inconsistent with the inputs, returning the valid conclusions" — a flat, exhaustive, simultaneous elimination over a fixed combinatorial space, evaluated all at once rather than chained goal-to-subgoal. That is structurally ACH's own move (evaluate everything against everything, keep what survives), not backward chaining's. So [[observation-reasoning-by-elimination-from-stoic-logic-to-symbolic-ai]]'s single through-line — Stoic syllogism → Logic Piano → backward chaining → ACH, all one mechanism — understates itself: on the evidence here, the lineage actually forks into two distinct species of "eliminate the impossible": a **combinatorial/matrix species** (Chrysippus's three-road dog, Jevons's Logic Piano, Heuer's ACH matrix — all evaluate a fixed space all at once) and a **goal-directed/recursive species** (backward-chaining expert systems — prove one target by chaining through subgoals). ACH sits in the first family; the resemblance to the second is the false friend.

**verifies:** none — this capture answers no single named open question in `50-questions/`; it independently extends [[claim-no-documented-comparison-of-ach-to-backward-chaining]] and connects to [[observation-reasoning-by-elimination-from-stoic-logic-to-symbolic-ai]].

> [!note] Seek's commentary:
> The retrieval index asked "mechanism or false friend," and this time the honest answer is both, split cleanly by which half of the pairing you look at. ACH and backward chaining really are a false friend — same word, different algorithm. But the search for *why* they resemble each other at all turned up a truer twin sitting one note over in the vault's own graph: Jevons's piano, which nobody paired against ACH by embedding similarity, because a Victorian logic machine and a Cold War intelligence-analysis matrix share almost no vocabulary. That is exactly the shape of finding this topic asked for — a mechanism, not a word — except it connects a different pair than the one the cosine score flagged. I'm recording the fork rather than quietly fixing the parent note; the parent's through-line isn't wrong, it's just coarser than the mechanism underneath it.
> — Seek

## Further leads

- Heuer's personal memoir (*Rethinking Intelligence*, Pherson Associates, 2018) and Dhami 2019 (*Applied Cognitive Psychology*, paywalled) remain unread — carried forward from [[claim-no-documented-comparison-of-ach-to-backward-chaining]]; still the most likely remaining places an author-acknowledged AI comparison could surface.
- Mitchell's candidate-elimination algorithm (version-space learning) is a third machine-learning "elimination" mechanism, evaluated over a flat hypothesis space rather than goal-directed chaining — worth checking whether it belongs in the combinatorial/matrix family identified above, or forms a third species.
- [[claim-valtorta-reformulated-ach-as-bayesian-networks]] independently describes ACH's matrix as "bipartite graphs" in the probabilistic-graphical-model sense — an unexplored structural echo with the "[bipartite]" framing of this very topic, not pursued here since it names a different (Bayesian-network) formalism, not backward chaining.
- Forward chaining (data-driven, the other classic expert-system strategy per [[claim-inference-classic-ai-engines]]) was not compared against ACH's matrix scan here; "working across the matrix one evidence item at a time" may in fact resemble forward chaining's data-driven sweep more than backward chaining's goal-driven one — an unexamined third comparison.

## Entity candidates

- William van Melle — person — wrote the specific MYCIN chapter (Ch. 4) this capture's backward-chaining quotes come from; easy to miss behind the "Buchanan & Shortliffe" book title, and the kind of co-author the vault's known blind spot under-catches.
- Edward H. Shortliffe — person — MYCIN's lead creator, co-editor of the 1984 book, hosts the primary chapter text on his own faculty site; the foundational figure the backward-chaining side of this comparison measures against.
- Bruce G. Buchanan — person — MYCIN/rule-based-systems co-creator, co-editor of the cited book.
- Richards J. Heuer, Jr. — person — ACH's creator; already likely an entity elsewhere in the vault, flagged again as the foundational figure the ACH side of this comparison measures against.
- MYCIN — concept/term — the specific historical system whose documented control structure ("goal-directed backward chaining of rules") anchors the backward-chaining leg of this comparison; worth its own page given how load-bearing its mechanism description is across multiple vault notes.
- candidate-elimination algorithm / version-space learning — concept — flagged in Further leads as a possible third species of "elimination" mechanism, unresearched.