First-Order Logic in Software Engineering

Model any system as facts: depends_on(service_a, database_b). Query for cycles, transitive dependencies, or circular references. LLMs see names and guess relationships — Prolog proves them.

LLMs guess who depends on whom. Prolog proves it — including the hidden 3-step chains nobody noticed.

Express routing rules as Prolog clauses: handles(billing, Channel). Query for the correct channel given a message type, payload constraints, or user context. Deterministic, not probabilistic.

LLMs say "route to the finance team." Prolog proves which queue, with which priority, and finds routing gaps before they become dead messages.

Encode schedules, capacity constraints, and time windows as Prolog facts. Query for valid assignments that satisfy every constraint simultaneously.

Constraint satisfaction is what Prolog was built for. LLMs propose schedules; Prolog proves they're valid.

Agents accumulate facts over time: user_preference(alice, theme, dark), session_context(bob, project, prolog_mcp). Query the agent's own knowledge base to make context-aware decisions. Survives restarts.

LLMs forget between sessions. Prolog facts persist — the agent remembers what the user prefers, across every conversation.

You have 15 requirements written by 3 teams. An LLM says "they look fine." But requirement #4 says "All users must have a unique email" and #12 says "Users can share emails across departments."

LLMs read text. Prolog proves consistency — or finds the exact pair that conflicts. Before a single line of code is written.

Team A produces JSON with user_id: string. Team B expects user_id: integer. They discover the mismatch in production on Friday at 23:00.

Express every team's input/output schema as Prolog facts. Query all_interfaces_valid before the teams even meet.

Instead of "the system should handle edge cases" — write acceptance criteria as executable logic:

Every test case passes or fails deterministically. No "looks good to me." The spec is the test oracle.

"Service A needs port 8080. Service B needs 8080. They can't share. C depends on A but not B. D depends on both. What's the valid port assignment?"

LLMs guess. Prolog searches the entire solution space and returns every valid configuration.

RBAC with 8 roles, 40 permissions, 12 resource types, and escalation rules. How do you know no role can accidentally access admin functions?

Security audits by LLMs say "this looks secure." Prolog explores every role-path combination and finds the one that breaks your model.

"Can the system ever reach a state where a payment is captured but no order exists?"

LLMs can't reason about all state transitions. Prolog explores the transition graph and finds the exact sequence that breaks the invariant.

Precondition: user is authenticated AND has role (admin OR editor) AND item is not locked AND item belongs to user's workspace AND workspace is active.

LLMs generate "likely" test cases and miss the ones that break your system. Prolog generates every combination.

GDPR says PII must never reach analytics. Your system has 40 services. Does any data path carry PII from the user service to the analytics pipeline?

Query pii_leak(source, sink) and get the actual path. Not "review your data flow." The actual path.

You're refactoring a 10-year-old function. The LLM says "looks equivalent." But does it handle every input the same way?

Query equivalent(old, new, Input) for every input class. Find the one case where your refactor changes behavior.

25 microservices. What's the optimal deployment order that minimizes downtime and never starts a service before its dependencies?

Topological sort with constraints. Prolog finds the order, parallelizes where safe, and detects cycles.

47 cron jobs across 12 servers. Some share databases. Some share rate-limited APIs. Which pairs can never run simultaneously?

LLMs "estimate" conflicts. Prolog checks every time overlap against shared resource rules.

GDPR, HIPAA, SOC2 — express regulations as Prolog rules, assert your workflow facts, query for violations:

Query compliant(my_workflow)? Provable yes/no. Show this to an auditor.

Your architecture docs say "no service should bypass the API gateway." Six months later, 3 services call each other directly. Which services violate the architectural rule?

Architectural rules as Prolog clauses. Query the actual codebase facts. Find violations automatically — before the architecture document becomes fiction.