Home

Welcome to Neomycin

⚠️ NOT FOR CLINICAL USE
Neomycin is a research artifact. It is not a medical device, a decision aid, or a diagnostic tool, and it must never inform a health decision for any human or animal. Every certainty number, drug, dose, and susceptibility in it is illustrative — chosen to make the reasoning machinery legible, not measured from clinical data.

Neomycin is a hybrid symbolic/LLM engine for bacterial identification and therapy selection, using Claude as a natural-language clinical assistant and Lisa, a production-quality expert system shell written in Common Lisp. It is strictly a research vehicle.

Neomycin began as a reconstruction of MYCIN, the Stanford medical expert system of the 1970s, rebuilt on a modern Common Lisp production-rule inference engine and fitted with a conversational front end powered by a large language model. Due to numerous improvements during the development period, Neomycin is related only distantly now to its ancestors.

The system holds a body of medical knowledge as explicit rules — statements of the form if these findings hold, the organism is one of this set, and here is how strongly I believe it. A user describes a case in ordinary English. The language model turns that description into structured facts, hands them to the inference engine, and then explains what the engine concluded. The engine does all the reasoning and all the algebra. The language model does none of it.

That division of labor is the point of the project. The model is good at language and bad at being auditable. The engine is the reverse. Keeping them strictly separate produces a system that can be talked to like a person and inspected like a ledger.

• Welcome to Neomycin
• Introduction
• The Historical Problem
• Project Purpose
• What Neomycin Is
• The Architecture
• The Symbolic Inference Engine
• Reasoning Under Uncertainty
• Neomycin is Worth Building
• The Conversation
• Answering "Why?"
• Neomycin is Actually Interesting
• Real Versus Schematic
• Flies in the Ointment
• What to Read Next
• References

Introduction

Neomycin began as a reconstruction of MYCIN, the Stanford medical expert system of the 1970s, rebuilt on a modern Common Lisp rules engine and fitted with a conversational front end powered by a large language model. It is now no longer a reconstruction of either MYCIN or its descendants. Rather, Neomycin has evolved into a novel implementation in its own right.

The system holds a body of medical knowledge as explicit rules — statements of the form if these findings hold, the organism is one of this set, and here is how strongly I believe it. A user describes a case in ordinary English. The language model turns that description into structured facts, hands them to the rules engine, and then explains what the engine concluded. The engine does all the reasoning and all the arithmetic. The language model does none of it.

That division of labor is the point of the project. The model is good at language and bad at being auditable. The engine is the reverse. Keeping them strictly separate produces a system that can be talked to like a person and inspected like a ledger.

The Historical Problem

MYCIN (Shortliffe, 1976) diagnosed bacterial infections of the blood and recommended antibiotics. It was one of the first programs to perform at the level of a human specialist in a narrow domain, and it did so by encoding roughly 450 rules elicited from infectious-disease physicians.

Two things about MYCIN mattered more than its diagnoses.

It could explain itself. A physician could ask why are you asking me this? or how did you reach that conclusion? and receive the actual chain of rules that produced the answer. Explanation was not a feature bolted on afterward; it fell out of the fact that the reasoning was made of inspectable parts.

It reasoned under uncertainty. Medicine rarely offers certainty, so MYCIN attached a number to every conclusion — a certainty factor — and combined those numbers as evidence accumulated.

MYCIN was never deployed. The reasons were partly practical (a mainframe login was an absurd way to consult about a patient in 1976) and partly institutional. Its architecture, however, was extracted into EMYCIN — "Essential MYCIN" — a shell that separated the reasoning machinery from the medical content so the same machinery could host other domains.

A few years later, William Clancey went further with a system called NEOMYCIN. He observed that MYCIN's rules mixed two very different kinds of knowledge: what is true about the world (gram-negative rods in burn patients suggest pseudomonas) and how a good diagnostician proceeds (establish the broad category first, then discriminate within it). Tangled together, neither could be taught, reused, or reasoned about. Clancey pulled them apart.

This project takes its name from that separation, because that separation is its whole architecture — with one substitution. Where Clancey wrote the diagnostic strategy as an explicit second body of rules, Neomycin lets a language model play that role.

Project Purpose

Neomycin is not an attempt to build a better diagnostic tool. It is an instrument for studying three questions.

Can a language model serve as the strategic layer of a symbolic system without contaminating it? MYCIN decided what to ask next by working backward from its hypotheses. That is a fine mechanism, but a rigid one. A language model is unusually good at deciding what to ask next and at conducting a conversation that feels natural. The open question is whether it can be given that job while being structurally prevented from touching the reasoning itself.

What does uncertainty look like when you model ignorance explicitly? Certainty factors compress everything into one number, so no evidence either way and strong evidence in both directions end up looking identical. Dempster-Shafer keeps them distinct, and Neomycin's rules are written for it: a rule states the set its evidence narrows to, and what it declines to claim stays visible as ignorance.

Neomycin once ran certainty factors and Dempster-Shafer over the same rules for comparison. It no longer can: a rule's answer is now a set, and certainty factors have no set algebra to reason over one. Both remain in the Lisa substrate for Lisa's own examples. The three-way comparison is reproducible on the v0.10.0 tag and not after it — an honest casualty of the representation getting better.

What does auditability cost, and what does it buy? Every conclusion can be unwound into the rules that produced it, what each of them answered, and the published sources behind each rule's clinical claim. The system is built so that a narrated explanation is a quotation of that record, not a reconstruction from memory. There is no arithmetic chaining one belief through another to quote, because the representation has none.

The medical domain is a vehicle. The architecture is domain-agnostic: substitute claims adjudication, underwriting, or equipment fault diagnosis and the shape of the system is unchanged.

What Neomycin Is

Neomycin began as a MYCIN/EMYCIN reconstruction and is no longer one. The divergence was not a goal; it accumulated, one representational problem at a time, and it is now large enough that comparing results against MYCIN's would be a category error. This is not a claim that Neomycin is better. It answers different questions, and it answers them about a corpus a fraction of MYCIN's size.

What changed, and why:

A rule states the SET its evidence narrows to, not "this organism is more likely". Exclusion is never authored — it falls out when answers are intersected and nothing is left. There are no ruling-out rules and no negative beliefs anywhere in the corpus.

Epidemiological rules GRADE their answers. A burn does not make Pseudomonas certain and Klebsiella impossible; it makes Pseudomonas likelier while excluding nothing. That is a mass function over several sets, and no single set can express it.

Dempster-Shafer over an open frame. The set of possible organisms is never enumerated, so a pathogen the corpus cannot name keeps its plausibility instead of being silently excluded by omission.

No organism classes and no chaining. A genus is a set, not a thing, and no belief is the product of two others.

A therapy phase MYCIN's illustration did not have: an exact minimum-set-cover solver over a schematic drug knowledge base, with explicit policy dials and an opt-in site-local antibiogram overlay.

An LLM strategy layer with enforced separation. The model conducts the interview and narrates; machinery — provenance records, /why, a queryable rule catalogue, guards over the prompt — makes it structurally hard for it to reason instead.

The rulebase is 46 rules against MYCIN's roughly 450, and every belief in it is a schematic teaching figure. :origin :paip-subset on a rule still means what it says — this association was inherited from the PAIP/EMYCIN illustration — but it does not make the historical treatment of that rule authoritative.

The Architecture

There are four components, each with a job it does not share.

Lisa. A forward-chaining expert system shell written in Common Lisp, using the Rete algorithm. It holds working memory and fires rules. Neomycin is a fork of Lisa that keeps the engine's packages un-renamed: Neomycin uses Lisa rather than absorbing it, though engine-level changes are made when the research genuinely calls for them.

The rulebase and belief system neomycin/rules/ holds 46 medical rules in a controlled vocabulary, grouped by cluster across a handful of files. Every one of them is confirming: it states the set of organisms its evidence narrows the answer to. Underneath sits a pluggable belief system — Dempster-Shafer over an open frame is Neomycin's, and certainty factors and a per-hypothesis Dempster-Shafer remain in the Lisa substrate for its own examples.

The bridge. A small HTTP service that runs inside the same Lisp image as the engine. It exposes the engine's capabilities as a handful of web endpoints: assert a fact, run inference, retrieve conclusions, retrieve the rule trace, ask which rules are close to firing, ask why a conclusion holds, request a therapy recommendation, and reset the session. The bridge is deliberately thin. It validates and translates; it does not reason. Its purpose is to make the Lisp image reachable from a process that is not Lisp — in practice, a Python client — without either side needing to know much about the other.

The driver. A Python program that connects a Claude model to the bridge using tool-use. The model is given a description of the bridge's capabilities and the engine's fact vocabulary, and is permitted to call those capabilities as tools. Everything the model learns about the case comes back through those calls.

The important structural property is that the model has no path to the answer except through the engine. It cannot compute a belief, because there is no tool that lets it write one; it can only assert facts, ask the engine to run, and read what came back.

The Symbolic Inference Engine

A rule pairs a set of conditions with a conclusion. Here is one (slightly elided) from the Neomycin rulebase:

(defrule burn-blood-aerobic-gram-neg-rod-narrows-to-opportunist-rods
    (:belief 0.4
     :provenance (:origin :paip-subset
                  :evidence ("NCBI Bookshelf, Medical Microbiology 4th ed. ch.27 (Pseudomonas), NBK8326"
                             "NCBI Bookshelf / StatPearls, Pseudomonas aeruginosa Infections, NBK557831"
                             "Highly Drug-Resistant Pathogens Implicated in Burn-Associated Bacteremia, PMC4128596"
                             "Gram-Negative Bacilli Blood Stream Infection in Patients with Severe Burns: a 9-Year Cohort, PMC11476612")
                  :belief-basis :illustrative
                  :note "Pseudomonas aeruginosa is a classic cause of bacteraemia in seriously burned patients, but it is not the only one..."))
  (organism (id ?o) (culture ?c))
  (culture (id ?c) (patient ?p))
  (culture-site (value blood) (of ?c))
  (gram (value neg) (of ?o))
  (morphology (value rod) (of ?o))
  (aerobicity (value aerobic) (of ?o))
  (burn (value serious) (of ?p))
  =>
  (assert (candidates (value '((0.20 :pseudomonas)
                               (0.07 :klebsiella)
                               (0.05 :enterobacter)
                               (0.08 :e-coli :proteus :serratia)))
                      (of ?o))))

The conclusion is worth a second look, because it is not what most rule engines assert. The rule does not say "this is Pseudomonas". It says "one of these six organisms, and here is how my confidence is spread across them" — 0.20 on Pseudomonas, 0.07 on Klebsiella, 0.05 on Enterobacter, and 0.08 shared across E. coli, Proteus and Serratia, which the evidence does not separate. The four masses total 0.40, exactly what :belief declares. A burn makes Pseudomonas likelier. It does not make Klebsiella impossible.

The clauses in the middle are patterns matched against working memory, the set of facts asserted so far. The ?o, ?c, and ?p are variables: they bind to whatever satisfies the first pattern and must then match consistently throughout, which is how the rule insists that the blood culture, the organism growing in it, and the burned patient are all the same case rather than three unrelated facts. The final clause asserts a new fact.

The :belief 0.4 says how much this rule's firing should move confidence in its conclusion. The :provenance block records where the rule came from and which published sources support the clinical association, and marks the belief value itself as illustrative — a machine-readable field that the explanation facility later reads back.

Forward chaining means the engine works from facts toward conclusions. When a new fact arrives, the engine finds every rule whose conditions are now satisfied and fires it, which may assert further facts, which may satisfy further rules. It runs until nothing more can fire. This is the opposite of MYCIN's original approach, which started from a hypothesis and worked backward looking for supporting facts.

Lisa Rete is a highly optimized version of Charles Forgy's original algorithm that makes forward chaining efficient. The naive implementation re-tests every rule against every fact on every cycle, which is quadratic and unusable at scale. Rete instead compiles the rules into a network that stores partial matches — the intermediate state of every rule that is partway to firing. Asserting a fact updates only the affected parts of that network. The engine trades memory for time, and the payoff is that the cost of adding a fact is proportional to what that fact actually affects.

That stored partial-match state turns out to have a second use, and it is the one Neomycin cares about most. Because the network already knows which rules are one condition short of firing, the engine can be asked directly: what single additional fact would let something new conclude? That question is what /partial-matches answers, and it is how a forward-chaining engine recovers the goal-directed behavior — knowing what to ask next — that MYCIN got from backward chaining. It is not a workaround. It is the same information, read off a structure that already existed.

Rules are given descriptive names, as in the example above, so that a trace of what fired reads as clinical prose rather than as a list of identifiers. That matters when the trace is going to be narrated back to a person.

Reasoning Under Uncertainty

This section assumes no probability background. The two systems below are both ways of answering: given several pieces of imperfect evidence, how confident should I be?

Certainty Factors: A Single Number

MYCIN's original scheme attaches a single number between −1 and +1 to each hypothesis. Positive means the evidence supports it; negative means the evidence argues against it; zero means nothing is known.

The interesting part is how two independent confirmations combine. Suppose one rule contributes 0.4 and another contributes 0.6. The result is not 1.0, and it is not their average. It is 0.76, and the reasoning is intuitive: after the first rule you are 0.4 convinced, so 0.6 of your doubt remains. The second rule claims 0.6 of what is left, which is 0.36. Add that to the 0.4 you already had and you get 0.76. Each new piece of evidence takes a bite out of the remaining doubt, so confidence approaches 1.0 without ever arriving.

The scheme's weakness is that one number cannot say two different things. A hypothesis about which nothing is known and a hypothesis with strong evidence on both sides both land near zero, and the system cannot distinguish them. In medicine that distinction is exactly the one you want.

Dempster-Shafer: Two Numbers and the Gap Between Them

Dempster-Shafer theory replaces the single number with a pair, written [bel, pl].

bel — short for belief — is the evidence that positively commits to the hypothesis. Read it as a floor: at least this much supports it.
pl — short for plausibility — is one minus the evidence committed against the hypothesis. Read it as a ceiling: at most this much room is left for it to be true.

The gap between them, pl − bel, is ignorance: evidence that has not been committed either way, or simply evidence you do not have.
A hypothesis nobody has said anything about sits at [0, 1] — a floor of zero, a ceiling of one, and complete ignorance in between. That is a genuinely different statement from "the evidence balances out," and it is the distinction certainty factors cannot make.

The mechanics are easier than the notation suggests, and in Neomycin they rest on a single idea: a rule states the set of organisms its evidence narrows the answer to. The Gram stain says "one of these eight". The aerobic result says "one of these seven". The burn says "one of these six, leaning Pseudomonas"; the compromised host says "one of these six, leaning E. coli". Each of those is an answer, with a weight, and answers combine by intersection — the organisms that survive are the ones in every answer.

That single mechanism does all the work, including the work that looks like it needs something else:

Exclusion is never written down. {pyogenes, agalactiae} intersected with {pneumoniae} is empty, and that emptiness is the ruling-out. No rule ever argues against an organism, and no rule can.

A genus is just a set. "Is this a staphylococcus?" is a question about {aureus, epidermidis, saprophyticus}, answered directly. Nothing chains, and no belief is the product of two others.

Belief committed to one organism is belief unavailable to its rivals, so evidence for one lowers the ceiling on the others automatically. An organism nobody has mentioned can still be effectively excluded.

Belief that lands on an impossible combination is the conflict, K — and it is reported with a companion, the margin, because K alone does not mean what it looks like it means. Two answers naming different organisms are incompatible totally, so K counts how much rival evidence was overruled, and it therefore climbs as the winning side strengthens. A high K can describe a case that is perfectly clear-cut. The margin — how far the leading answer sits above the nearest answer that contradicts it — is what says whether the disagreement resolved.

The frame is open. The system never enumerates every organism that exists — only the sets its rules actually name. So the question "could this be something you don't model?" has a real answer rather than an embarrassed silence: it is the belief that has not been committed anywhere, and the system will quote it.

Neomycin is Worth Building

Here is a case with three lab findings: the organism ferments lactose, tests indole-positive, and produces a red pigment on the plate. Lactose and indole together point to E. coli. Red pigment points to Serratia. One organism cannot be both.

Nothing in the rulebase says so. There is no rule anywhere that argues against E. coli or against Serratia — the corpus contains no such rule and cannot express one. The findings simply give answers that do not overlap, and the arithmetic does the rest:

Organism	Belief (floor)	Plausibility (ceiling)
E. coli	0.670	0.758
Seratia	0.242	0.303

with conflict K = 0.736 and a margin of 0.427.

Read the ceilings first. Both have fallen well below 1.0, which is the signature of genuine contradiction; in an ordinary case where the evidence agrees, both would sit at 1.0. E. coli leads because three answers admit it and only one admits Serratia. But nearly three quarters of the belief in this case landed on a combination that cannot be true, and the two specific answers driving it — {e-coli} at 0.80 and {serratia} at 0.80 — are equally strong and cannot both hold. The right thing to tell a clinician is not "E. coli, 67%"; it is that the bench results contradict each other and should be repeated.

The margin is what licenses reading it that way, and it is worth seeing the contrast. Take two answers constructed to isolate the effect — {pseudomonas} at 0.928 against {klebsiella} at 0.60. That pair runs K = 0.557, comparably high, with a margin of 0.740: the rival is decisively overruled and the answer is simply Pseudomonas. Same shape of number, opposite meaning. Here the margin is 0.427 and the two specific answers are tied, so the high K is what it appears to be. Neither figure is interpretable alone; the pair is.

Those two are answer masses, not a consultation. No scenario in this project produces them, and they are quoted here only because a constructed pair isolates what K and the margin each measure. The contradiction figures above, by contrast, are what the engine actually returns for that case.

That is the property worth building for. A system that reports its own inputs disagree is more useful than one that smooths the disagreement into a confident-looking average — and here the disagreement is not a special feature bolted on, but a consequence of representing an answer as a set.

Contrast the case where the evidence agrees: a chain-forming gram-positive coccus, alpha-hemolytic, optochin-sensitive. Four answers, each nested inside the last, and every one of them admits S. pneumoniae. It reaches bel 0.850 with pl 1.000 and K = 0 — nothing lost to conflict, and no ceiling below one. Because the answers nest, belief settles at the sharpest one's own weight rather than climbing past it; evidence that overlaps without nesting does climb past it, and the E. coli case later in this document reaches 0.884 that way. Same machinery, opposite reading.

The Conversation

Here is a hypothetical consultation, from the user's first sentence to the system's answer. The user types into a prompt; nothing else is involved.

A person describes a case in ordinary English. For example: a patient with serious burns and a compromised immune system has an aerobic gram-negative rod growing in a three-day-old blood culture, and what organism is this?

The model translates that into structured facts. It has been told the exact fact vocabulary the engine understands — which attributes exist, which values are legal, which entities facts attach to — and it maps the free text onto that vocabulary. Severe burns become a fact about the patient. Gram-negative and rod-shaped become facts about the organism. Each is submitted individually through the bridge, and the bridge validates it before it enters working memory. Malformed facts are rejected there, not silently absorbed.

The model asks the engine to run. The engine fires every rule whose conditions are met, each one asserting the set of organisms its evidence narrows the answer to, and records which rules said what.

The model reads back what concluded. For the case above, a three-way differential: E. coli at belief 0.232 with plausibility 0.564, pseudomonas at 0.176 / 0.468, and klebsiella at 0.165 / 0.458. Nothing reaches plausibility 1.0, because belief supporting one organism is belief the others cannot also have — and nothing has been excluded either, because the only evidence so far is a stain and two facts about the patient. A further 0.234 sits on the set of seven aerobic gram-negative rods without naming a member: the honest statement that the family is better established than any species in it, and often the right headline. K = 0.180 records the belief that went to combinations which cannot all hold.

The reading a clinician should take from this is that the culture has not been discriminated yet. A burn raises Pseudomonas and an immunocompromised host raises E. coli, and the engine says both — but epidemiology ranks organisms, it does not identify one. A lactose or indole result would. An earlier version of this rulebase returned pseudomonas at 0.613 here and looked far more decisive; it reached that number by asserting that a burn made Klebsiella, E. coli, Enterobacter, Serratia and Proteus impossible, which is not true and is not what burn-unit surveillance data say. The smaller numbers are the more honest ones.

The model narrates the result. It reports the differential, the beliefs, and the reasoning — all of it read from the engine rather than composed.

Three details in that flow deserve emphasis.

The model conducts an interview, not a form. If the case description leaves out something the engine needs, the model asks for it in natural language. It decides what to ask by consulting the engine's partial-match state — the rules that are one fact short of firing — so its questions are the ones that would actually change a conclusion. This is the clinician's-assistant role: the model manages the dialogue, notices what is missing, asks in ordinary language, and keeps the conversation coherent across turns. What it never does is decide anything about the case.

Uncertainty in the input is carried, not flattened. A person who says "I think the stain was gram-positive" is not making the same claim as one who says "the stain was gram-positive." The model attaches a reduced confidence to the asserted fact, and the engine's belief arithmetic carries that reduction through to every conclusion downstream.

Evidence of different sharpness combines without any of it being scaffolding. An aerobic gram-negative rod answers "one of these seven". A lactose result narrows to four of them. Indole narrows to two. The combination that names E. coli answers with one. Those are four answers at four resolutions, and because each set contains E. coli they agree — they reinforce to 0.884, higher than any single rule's own weight, with conflict zero.

The coarse answers are conclusions in their own right, not way-stations. "One of these seven, the evidence does not say which" is frequently the most honest thing the system can report, and it is reported rather than discarded on the way to a species. Nothing chains, nothing is scaffolding, and no belief is the product of two others.

Answering "Why?"

Explanation was MYCIN's signature capability, and reproducing it faithfully is a substantial part of this project.

The naive way to explain a language model's answer is to ask the model. That produces fluent, plausible text with no reliable connection to what actually happened. Neomycin does something structurally different.

Two records are kept. Each rule carries provenance: where the rule came from (inherited from the historical MYCIN corpus, or added by this project) and which published sources support its clinical claim, each citation checked adversarially rather than accepted on the model's say-so. Separately, the engine records which rules produced each answer as they fire.

The bridge composes both into a single answer: the argument. Asked why an organism stands where it does, it returns every answer given about that culture — the set each one narrowed to, its weight, the rules that said it, and their sources — together with what those answers intersect to.

It also returns the answers that do not admit the organism, and that is the load-bearing part. Since nothing in this corpus argues against anything, an organism can only lose ground because other evidence pointed somewhere else. An explanation showing only the supporting answers would leave a reader to assume an objection that does not exist.

So when a user asks why Klebsiella, and how sure are you?, the model does not answer from memory. It requests the argument and reads it aloud. For the burn case above, four answers were given and every one of them admits Klebsiella — the Gram stain narrowing to eight organisms at 0.70, the aerobic rod finding to seven at 0.80, the burn answer at 0.40 which puts 0.07 of its mass on Klebsiella, and the compromised-host answer at 0.60 which puts 0.16 there. It nonetheless sits at bel 0.165, pl 0.458, because those two graded answers commit most of their mass elsewhere — 0.20 to Pseudomonas, 0.28 to E. coli — and belief spent on a rival is belief Klebsiella cannot have. Nothing argued against it, and nothing could.

Ask the same question about Salmonella and the answers that do not admit it are the whole story: the stain and the aerobic result still admit it, neither graded answer names it at all, and it lands at bel 0.000, pl 0.293 — untouched rather than rejected. Reading only the supporting half of either argument would misdescribe both.

There is no arithmetic composing one belief through another to quote, because nothing chains. And it reports the boundary the record itself encodes — that the sources verify the clinical association, not the number, and that the rule weights are illustrative teaching values rather than measured frequencies.

That last behavior is the one worth dwelling on. The model declines to present a schematic number as a measured one, because the engine's own record says which it is. The honesty is a property of the architecture, not of the model's disposition.

From Identification to Treatment

Identifying the organism is half a consultation. The other half is deciding what to treat with, and Neomycin handles it with the same separation of responsibilities.

Treatment selection is posed as a covering problem. Every organism whose belief clears a threshold must be covered by at least one drug the patient can actually take, and the regimen should use as few drugs as possible.

This is a classic set-cover problem. Finding the true optimum is computationally hard in general, but not at this scale — eleven drugs and a handful of organisms — so Neomycin searches exhaustively: it finds every smallest regimen that covers the case, then picks among them by a stated rule. Drugs the patient cannot take are excluded up front, with the reason recorded. Any organism no available drug can cover is reported as uncovered rather than quietly dropped. A greedy solver is kept alongside, and the test suite checks the two agree.

Fewest drugs is not the same as narrowest drugs, and the difference turned out to matter more than we expected. Using fewer drugs is good; reaching for a broad-spectrum agent when a narrow one would do is the thing that drives resistance. Those are separate goals, and an early version of this system conflated them: its code claimed to practise antimicrobial stewardship while actually optimising drug count alone. Because the broadest agents also carry the best coverage numbers, that objective reached for a last-line carbapenem almost every time — and, in a recorded session, told a clinician who asked for something narrower that no narrower option existed. Five did.

So the choice is now a dial you can turn, alongside the ones for the belief system and for how much certainty to demand, and it has three settings across two axes. One reproduces the original behaviour, honestly labelled: fewest drugs, then best coverage. The second prefers narrow-spectrum agents. The third prefers cheap ones in the stewardship sense — the WHO's Access, Watch and Reserve tiers. The same case under different settings gives different defensible answers, and the system reports what each one gave up: the narrower agent usually has a lower, and less certain, chance of working.

There are three settings rather than two because breadth and stewardship cost turned out to be genuinely different questions, and neither subsumes the other. Vancomycin is narrow-spectrum and exactly the sort of agent a steward holds back, so on a group A strep the narrow-spectrum setting returns vancomycin where the stewardship setting returns ampicillin.

The narrow-spectrum dial is also allowed to be visibly wrong on its own terms: it can see how broad a drug is but not how precious, so it will reach for a reserve antibiotic — the kind held back precisely so it keeps working — on the grounds that it is narrow. That could have been quietly patched. It wasn't. A second dial was added beside it and the first was left saying exactly what it means, because a research instrument that shows you where its reasoning runs out is worth more than one tidied until it looks authoritative.

What none of the three can do is trade one Reserve drug for two Access drugs. Fewest drugs is primary under all of them, and real stewardship sometimes wants exactly that trade. The limit is stated rather than approximated around.

Every regimen also lists what it did not choose: the other drugs that would have covered the case, and the other equally small regimens that lost on the tiebreak. A recommendation that shows only its winner invites the reader to assume there was no alternative, which is how the false answer above came to be given in the first place.

Susceptibilities — how well a given drug works against a given organism — are themselves belief-valued, not booleans, and they can be refined by a site-local antibiogram: the record of how many isolates on this particular ward actually proved susceptible to this drug.

That refinement is where a small amount of statistics earns its place, and it can be stated in words. Suppose the local lab tested 10 isolates and 7 were susceptible. The optimistic reading is 7 in 10. The pessimistic reading is also 7 in 10. But you have only tested 10 — so imagine a couple of additional tests whose outcomes you never saw. A pessimist assumes those unseen tests all failed; an optimist assumes they all succeeded. Those two assumptions give you a floor and a ceiling, and the gap between them is exactly the ignorance that comes from a small sample. Test 500 isolates instead of 10 and the two imagined extra tests barely move either bound, so the interval tightens on its own. No separate confidence machinery is needed: sample size and ignorance are the same quantity.

The local interval is then pooled with the curated figure from the knowledge base, so local resistance can pull coverage down and solid local data can promote a drug the curated knowledge base was unsure about. Every susceptibility carries its own provenance — where the number came from, and how many isolates stand behind it — so the model can narrate that too.

The rule that governs the whole phase is the same one that governs identification: the solver chooses the drugs, and the model explains the choice. The model never picks a drug.

Neomycin is Actually Interesting

Set the medicine aside. What remains is a pattern with three properties that are hard to get at the same time.

A natural-language interface over a system of record that stays authoritative. Most attempts to put a language model in front of a decision system end up letting the model make decisions, because the boundary is a matter of prompting and prompting is not a boundary. Here the boundary is structural: the model's only available actions are to submit facts, invoke the engine, and read results. There is no tool through which it could score a hypothesis, so it cannot, and no amount of clever prompting changes that.

Explanation that is a quotation rather than a reconstruction. The system does not ask the model to explain the answer. It asks the engine for the derivation and has the model read it aloud. The difference matters most in exactly the settings where explanation is legally or ethically required, because a reconstructed explanation can be fluent and wrong in ways nobody can detect from the text.

Uncertainty that stays legible instead of collapsing. Conflicting evidence produces a visibly wider, lower interval rather than a confident-looking average. A system that can tell you its own inputs disagree is more useful than one that smooths the disagreement away, and it is the kind of signal that any serious adjudication process wants surfaced.

There is also a historical argument. Expert systems were declared a dead end, and the reasons given were real: knowledge acquisition was expensive, and the interfaces were unusable. Both of those failures were about the human edges of the system — getting knowledge in, and getting explanations out — not about the inference core, which worked. Language models are unusually good at exactly those edges. Neomycin is a test of the resulting hypothesis: that the symbolic core was never the problem, and that pairing it with a model that handles the edges produces something neither approach reaches alone. The engine keeps the model honest, and the model makes the engine usable.

Real Versus Schematic

Being clear about this is part of the project's purpose.

Real: the inference engine, the belief algebra and its arithmetic, the set intersection and the conflict behavior, the explanation and provenance records, the therapy solver and its guarantees, the antibiogram mathematics, the release gate that checks the model's narration against the payloads it was actually given (bin/release-check.py — every quoted number must appear in something the engine returned), and the test suite — roughly 1859 assertions across 239 tests, including every rule fired in isolation, hand-verified golden values for each scenario, and corpus-wide invariants checked by introspecting the compiled rulebase so that a new rule is covered the moment it is written.

Schematic: the certainty numbers. Every rule weight is a teaching value, chosen to make the machinery observable, and each is explicitly marked as such in its own provenance record. The published citations attached to a rule verify that the clinical association is real; they do not verify the number.

Grounding those numbers in real frequency data is deliberately not being done. Illustrative is the honest state, and stating it loudly is better than a set of numbers that look measured and are not. What has been done is narrower and worth distinguishing: making the numbers consistent with each other, so that (for instance) a rule cannot commit less than a more general rule it displaces. That is a coherence property, not a calibration claim.

Also schematic: the drug knowledge base, its doses, its susceptibilities, and its contraindications. The rulebase is 46 rules against MYCIN's original 450 — enough to exercise every mechanism in the architecture, and nowhere near enough to be clinically meaningful. It can name 17 organisms; a real differential is not 17 organisms wide.

Redundant evidence, and why one rule can speak for several. Four of the gram-negative epidemiological rules rest on the same underlying statistics — they are one fact reported four times, under four labels. Combining them would count that fact repeatedly: a patient who was both immunocompromised and neutropenic saw the leading organism's belief inflated and the reported conflict inflated, between two rules that agreed. They are now declared as a single evidence group, and only the most committed member contributes; the rest are dropped before combination and are absent from the explanation as well as the arithmetic. Contexts that genuinely differ — a burn, a tropical journey — carry their own distributions, are in no group, and still combine normally. Measured and analysed in docs/base-rate-investigation.md.

This is worth reading as an instance of a general problem rather than a local fix: Dempster's rule assumes the evidence you combine is independent, and a knowledge base assembled from overlapping literature will quietly violate that. The corpus now declares the dependence rather than hoping it does not matter, and two invariants check the declaration from both directions.

Flies in the Ointment

As a potential clinical tool, in addition to the medical community's rejection of expert systems decades ago, Neomycin is currently considered untrustworthy. LLMs, and in the case of this project Anthropic's Claude, seem to be notorious for circumventing guards in system prompts, behaving unpredicatably and, on occasion, dangerously.

During late-stage development of Neomycin, it was found that, despite instructions (guards) to the contrary, Claude would sometimes inject its own vision of pathogen possibilities and their corresponding treatments. In such cases Claude was always wrong. Discovery of this behavior was possible because the clinical assistant driver, by default, records each session in a format that may be analyzed by Claude itself.

Divergence from any narrative that was used to exercise the system seemed to be the trigger causing Claude to "lie" and work outside of its explicit parameters and restrictions.

It is uncertain at this point whether Claude's "hallucinations" when acting as a clinical assistant can be overcome to the extent that Neomycin could be fully trusted.

What to Read Next

getting-started.md (sourceforge.net): Building, running, and testing; how to read the output
runbook.md (sourceforge.net): A guided tour of a full identification consultation
therapy-demo.md (sourceforge.net): The treatment phase, end to end
clinician-scenarios.md (sourceforge.net): Worked cases exercising the rulebase and the
why-how-provenance-design.md (sourceforge.net): The design of the explanation facility
antibiogram-overlay-design.md (sourceforge.net): The site-local susceptibility mathematics
rule-catalogue-design.md (sourceforge.net): Why the rulebase is queried rather than described to the model
demo-runsheet.md (sourceforge.net): A 15-minute live demonstration script
CLAUDE.md (sourceforge.net): Build notes and the layout of the codebase

References

• Shortliffe, E. H. (1976), Computer-Based Medical Consultations: MYCIN
• Buchanan, B. G. and Shortliffe, E. H. (1984), Rule-Based Expert Systems
• Clancey, W. J. (1987), Knowledge-Based Tutoring: The GUIDON Program
• Shafer, G. (1976), A Mathematical Theory of Evidence
• Norvig, P. (1992), Paradigms of Artificial Intelligence Programming, chapter 16, whose forward-chaining translation of the MYCIN rules this rulebase follows
• Forgy, Dr. Charles L. (1982), Rete: A Fast Algorithm for the Many Pattern/Many Object Pattern Match Problem, Artificial Intelligence. 19: 17–37

Project Members:

• David Young (admin)