Lecture 20 · Act V · Applications

Uncommon domains.

Cyber, law, health, finance, art, and education do not share a vocabulary. They share a form: a case, a rule, a decision, a trace, and a person who signs the exception.

6 domains 3 verdicts 1 interactive lab Self-paced
00 · Learning objectivesLecture 20 · Uncommon domains

By the end of this lecture

Four things you can do.

01

Map six domains

Cyber, law, health, finance, art, and education — and the symbolic object each one turns on.

02

Read the case tuple

(N, F, R, ≺, E, d, τ, o): source, facts, rules, precedence, exceptions, decision, trace, override.

03

Audit a decision trace

Name the rule that fired, the exception that waived it, and the human who signed.

04

Name the domain limits

What the symbolic part guarantees, and which part of the world it cannot certify.

S/V · Lecture 20
00 · The hookLecture 20 · Uncommon domains

Six rooms, one form

The domain changes the vocabulary.

Cyber

A path through a network, composed from exploits.

Law

A rule and a precedent, and the reading that joins them.

Health

A guideline and a patient who may not fit it.

Finance

A posting, an invariant, and a control.

Art

A grammar of forms, and a selection.

Education

A solution graph, and a hint that must be licensed.

The audit loop does not change. Normative text, case facts, a decision, a trace, and a human who signs the exception.
S/V · Lecture 20
00 · CyberLecture 20 · Uncommon domains

Cyber · three artifacts

A single alert is not a story.

Log

An event was recorded. It does not say whether the event is malicious, or whether the record is complete.

Alert

A detector matched a pattern. It does not say that the pattern is an attack rather than a false positive.

Path

A sequence of exploits is possible under a vulnerability model. It does not say that the sequence occurred.

Denning (1987) made detection statistical. Ning, Cui, and Reeves (2002) correlated alerts into scenarios. The scenario is an abduction: which attack would make these alerts expected?

S/V · Lecture 20
00 · CyberLecture 20 · Uncommon domains

The symbolic object

The attack graph is a plan.

internet web · user db · user db · root goal exploit 1 exploit 2 exploit 3 observed alert nodes = security conditions · edges = exploits · a path = a composition of preconditions and effects
Schematic, not a real network. After the attack-graph literature (Phillips & Swiler 1998; Sheyner et al. 2002; MulVAL, 2005).
Symbolic graph + correlation rules.
Neural alert and anomaly classifiers.
Guarantee valid under the model — not proof of compromise.
S/V · Lecture 20
01 · Law & complianceLecture 20 · Uncommon domains
01 / 08

The statute is a program with an open texture.

Rules, precedents, and the score that cannot explain itself.

S/V · Lecture 20
01 · Law & complianceLecture 20 · Uncommon domains

Three symbolic objects

Rules, cases, and a score.

Statute as program

Catala writes the legal text and the rule together, with defaults for exceptions. It can show which rule produced the calculation. It cannot certify that the encoding is faithful.

Case-based

HYPO compares a case to precedents along dimensions and builds claim, response, and reply. It can show which case is closer. It cannot weigh a fact no precedent anticipated.

Actuarial score

COMPAS ranks risk from features and weights. It can rank. It cannot name the rule or the facts that produced the ranking.

The encoding is a legal act. Choosing what to encode and which exceptions to make explicit is a policy decision, not a neutral translation.
S/V · Lecture 20
01 · Law & complianceLecture 20 · Uncommon domains

The cautionary case

A score is not a trace.

ProPublica, 2016. The analysis argued that COMPAS error rates differed by race.
Dressel & Farid, 2018. The response disputed that conclusion and argued the answer depends on which error rate one chooses.
This lecture does not adjudicate the dispute. It uses the case to make a narrower point about form.
Explainability is not derivation. A post-hoc reason can be useful; it is not the rule that fired.

Fairness is a property of the policy and its effects, not of the engine. The audit trail is a prerequisite for asking, not the answer.

S/V · Lecture 20
02 · HealthLecture 20 · Uncommon domains

The oldest case study

Rules with confidence.

MYCIN (Shortliffe, 1976) asked for findings, chained backward through production rules, and recommended therapy. Its certainty factors were a hand-designed calculus, not calibrated probabilities. It remained a research system; its lesson is the trace.

Rule

If the organism is gram-negative and the site is sterile, then the therapy is …

Certainty

Evidence combines through a simple, inspectable calculus. The number is not a posterior.

Trace

The recommendation arrives with the rules that supported it. That is what a clinician can inspect.

INTERNIST-1 / QMR took the other route: a large internal-medicine knowledge base and a scoring differential. Inspectability costs maintenance.
S/V · Lecture 20
02 · HealthLecture 20 · Uncommon domains

The interface

Perception at the boundary.

NeuralImage, signal, or note model produces a finding with a probability.
FactsThe finding becomes a documented fact, with its provenance.
GuidelineThe pathway says whether the fact changes the next step.
HumanA clinician owns the exception and the decision to deviate.
The guarantee. The recommendation follows the encoded guideline for the documented facts. It is not a diagnosis and not an outcome.
The limits. Guidelines lag evidence; patients are exceptional; a calibrated model can still be wrong; alerts can fatigue the people who read them.

Guideline languages: GLIF, Asbru, PROforma, EON. Compared by Peleg et al. (2003). Imaging benchmark: Esteva et al. (2017) — a benchmark result, not a deployment licence.

S/V · Lecture 20
03 · Finance & auditLecture 20 · Uncommon domains
03 / 08

Every posting is a proof obligation.

Invariants, controls, and the trace that lets a reviewer reconstruct the decision.

S/V · Lecture 20
03 · Finance & auditLecture 20 · Uncommon domains

What a machine can check

The ledger is an invariant.

Double entry

Every journal entry has equal debits and credits. The trial balance follows. A balanced fraudulent posting still balances.

Three-way match

Purchase order, receipt, and invoice agree. A collusive vendor can supply all three.

Segregation

Requester, approver, and payer are different people. Collusion across roles defeats the check.

The invariant is necessary, not sufficient. A machine checks the identity; a human checks the story. DMN decision tables make the hit policy explicit — first, unique, collect.
S/V · Lecture 20
03 · Finance & auditLecture 20 · Uncommon domains

Interactive · one claim, two switches

Rule, exception, verdict.

ClaimEXP-1042
Categorymeal
Amount$248.00
Attendees4
Per attendee$62.00
Cap$60.00
ExceptionMEAL-EXC
Receiptattached
ACCEPT · E1
R2 failed; E1 waived the cap.
R2 · fail · $62.00 per attendee > $60.00
E1 · applied · MEAL-EXC waives R2
→ ACCEPT with exception
S/V · Lecture 20
03 · Finance & auditLecture 20 · Uncommon domains

The audit record

The trace is the product.

OrderRuleResultWhy
1R3 · prohibited categorypassmeal is reimbursable
2R5 · duplicate claimpassno earlier match
3R6 · currencypassUSD; no conversion
4R1 · receiptpassreceipt attached
5R4 · travel approvaln/anot a travel claim
6R2 · meal capfail$62.00 > $60.00
—E1 · exceptionappliedMEAL-EXC waives R2
—R0 · defaultskippeddecision: R2 + E1
Sound. The decision follows from the rules and facts. Re-derive the decisive rule and check every premise.
Not sufficient. Sound relative to a rule set that may be wrong. The trace is a proof about the policy, not about the world.
S/V · Lecture 20
04 · ArtLecture 20 · Uncommon domains

Procedural generation

The grammar generates; the artist selects.

AARON (Harold Cohen) encoded a drawing practice as hand-written rules: composition, figures, colour. It did not learn a style from a corpus; it explored a space the artist defined.

L-systems (Lindenmayer, 1968) are rewriting rules for branching form. A derivation either exists or it does not; that is what the grammar can certify.

Generate and criticize. The grammar guarantees a well-formed candidate; a neural critic ranks; a human selects.
S/V · Lecture 20
04 · ArtLecture 20 · Uncommon domains

The guarantee is narrow

What the grammar can promise.

Guaranteed

Membership in the grammar. Satisfaction of a composition constraint. Reproducibility from a recorded rule set and checkpoint.

Not guaranteed

Quality. Originality. Authorship. Intent. In the copyright case, those are legal questions, not model properties.

Selection is where intent enters.

If the interface hides the selection, it hides the authorship. A neural critic is an optimizer, not a curator.

S/V · Lecture 20
05 · EducationLecture 20 · Uncommon domains
05 / 08

The hint must be licensed by the next step.

Model tracing, knowledge tracing, and the LLM that proposes while the tutor disposes.

S/V · Lecture 20
05 · EducationLecture 20 · Uncommon domains

The symbolic core

A graph of legal steps.

s0 start s1 student is here s2 correct s3 common error s4 a hint is licensed only if the next step is an edge from here valid path error branch
Domain model. The graph encodes correct and incorrect steps, not only one solution.
Limit. A wrong graph teaches the wrong procedure; a valid alternative can be marked wrong.
S/V · Lecture 20
05 · EducationLecture 20 · Uncommon domains

The modern interface

The model proposes; the tutor disposes.

LLMGenerates a hint, a grade, or an explanation in fluent language.
VerifierChecks the proposal against the next licensed step and the prerequisites.
DecisionLicensed: show it. Unlicensed: repair, simplify, or ask a question instead.
StudentActs on the hint; the tutor updates the student model.
Guarantee. The feedback is consistent with the domain model for the student's state. It is not a guarantee of learning.
Without a graph. The check becomes a rubric or a constraint, and the guarantee weakens. Effectiveness is an empirical question, not a property of the interface.

Cognitive Tutors: Anderson, Corbett, Koedinger & Pelletier (1995). Knowledge tracing: Corbett & Anderson (1994). Effectiveness: VanLehn (2011); the literature is more textured than the “two sigma” slogan.

S/V · Lecture 20
06 · The common patternLecture 20 · Uncommon domains

Six domains, one loop

The audit loop is the common form.

normative text statute · guideline · policy rule base rules · precedence · exceptions engine apply · waive · flag decision + trace rule · facts · exception human review · sign case facts extracted · verified override is logged; policy change is explicit
The common loop. The vocabulary changes; the slots do not.
Symbolic rules, precedence, exceptions, trace.
Neural perception, extraction, ranking, generation — at the boundary.
S/V · Lecture 20
06 · The common patternLecture 20 · Uncommon domains

The decision space

Three verdicts, one human act.

Accept

No rule blocks, or a named exception waives the only blocker. The default rule is R0.

Reject

A blocking rule fires. The first failure in precedence order is decisive; later failures are still reported.

Review

A required fact is missing, or two rules conflict. The case is flagged, not guessed.

Override

Not a fourth logical verdict. A human accepts a rejected or reviewed case, with a reason, and the act is logged.

Collapsing review into accept or reject hides exactly the cases a human should see. The review bucket is a feature, not an embarrassment.
S/V · Lecture 20
07 · Tuple & casesLecture 20 · Uncommon domains

The schema

Read a case as a tuple.

c = (N, F, R, ≺, E, d, τ, o)
SlotMeaningIn the lab
NNormative source and versionthe expense policy
FCase facts with provenancethe eight claims
RRule setR1–R6
≺Precedence orderR3, R5, R6, R1, R4, R2, R0
EException policyE1 meal, E2 receipt
dDecisionaccept · reject · review
τTraceevery rule, every result
oOverridea logged human act
S/V · Lecture 20
07 · Tuple & casesLecture 20 · Uncommon domains

The sharp version

The rule fired is not the rule that is right.

A rule engine can be sound relative to its rules and still be wrong about the world. The guarantee is conditional on the specification and the facts.

The trace is a proof about the policy, not a certificate about reality.

S/V · Lecture 20
08 · RecapLecture 20 · Uncommon domains

What you can do now

Four moves, one schema.

Map

Six domains, six symbolic objects: graph, rule, guideline, invariant, grammar, solution graph.

Read

The case tuple: source, facts, rules, precedence, exceptions, decision, trace, override.

Audit

Name the rule that fired, the exception that waived it, and the human who signed.

Limit

Sound relative to rules and facts, not correct about the world; fairness is a separate argument.

S/V · Lecture 20
08 · NextLecture 20 · Uncommon domains

Next · Lecture 21

Benchmarks, evaluation, and how to read a NeSy paper.

Six domains, one question: on what benchmark, under what assumptions, with what guarantee? The next lecture turns the tuple into an evaluation protocol.

S/V · Lecture 20
Speaker notes