SpecLogician
SpecLogician is an AI framework for data-driven formal program specification synthesis, verification, and analysis.
It replaces monolithic “spec-first” formal modeling with an incremental, scenario-driven approach that works naturally with LLM-powered and agentic tools—while keeping correctness grounded in machine-checkable formal logic.
- Website: www.speclogician.dev
- Core idea: build a formal model from symbolic
given / when / thenscenarios, linked to a rich domain model (predicates, transitions, state/action types, and auxiliary logic).
1) The challenge: scaling formal methods via LLM-powered automation
- Automatically applying formal methods to large software systems using LLM-powered and agentic tools remains a fundamental challenge
- Traditional formal modeling approaches require building large, monolithic formal models upfront
- There is no single canonical way to formalize a complex software system
- As a result, formalization becomes as much an art as a science, relying heavily on expert judgment
- These characteristics fundamentally limit automation:
- LLMs struggle to generate or maintain large, globally consistent formal models
- Small local changes often require understanding the entire model
- Monolithic models are brittle under iterative, agent-driven workflows
2) SpecLogician’s core idea
- SpecLogician is an AI framework for data-driven formal program specification synthesis, verification, and analysis
- It replaces monolithic specifications with incrementally constructed formal logic
- The core logic is built from symbolic
given / when / thenscenarios - Scenarios are:
- composable
- declarative
- local in scope
- Scenarios are connected to a domain model of arbitrary complexity
- The domain model captures:
- predicates
- transitions
- state/action structure
- auxiliary and domain-specific logic
3) Why this structure works well with LLMs
- LLM-powered tools are used to:
- propose new scenarios
- refine existing scenarios
- generate structured deltas (add / remove / edit)
- LLMs operate on small, well-scoped artifacts, not entire formal models
- Each change is:
- explicit
- typed
- machine-checkable
- This aligns naturally with how LLMs perform best:
- local reasoning
- incremental edits
- structured outputs (JSON, CLI commands)
4) Agentic reasoning loop (formal reasoning as the backbone)
- SpecLogician sits at the center of an agentic reasoning loop
- In this loop:
- CodeLogician / ImandraX translate source code into formal models and reasoning artifacts
- LLM-powered agentic CLIs propose scenario additions, edits, and removals as structured deltas
- Software mapping tools (e.g. CodeMaps from cognition.ai) provide high-level program structure and navigation context
- Each agent contributes partial, local insight:
- code structure
- behavioral intent
- execution traces
- test artifacts
- SpecLogician:
- integrates these inputs into a single formal state
- validates them symbolically
- delegates global analysis to the ImandraX automated reasoning engine
- The reasoning engine analyzes the entire accumulated model after every change
- This creates a closed-loop workflow:
- propose → formalize → verify → refine
5) What the resulting formal model enables
- Systematically identify gaps in design and requirements
- Precisely understand test coverage and gaps in test coverage
- Trace:
- execution logs
- test cases
- documentation
- other artifacts
back to formal specifications and requirements
- Automatically generate missing test cases
- Safely model and verify the impact of changes before they are applied
- Maintain a single, authoritative source of truth for system behavior
6) Big-picture outcome
- Transforms LLMs from:
- probabilistic code generators
into: - reliable collaborators in a verification-driven workflow
- probabilistic code generators
- Makes formal methods:
- incremental
- data-driven
- compatible with LLM-powered automation
- scalable to real-world software systems
- Positions SpecLogician as the formal reasoning backbone for modern, agentic software development