nl2spec is a research-oriented pipeline for transforming natural language descriptions of API usage rules into executable Runtime Verification (RV) specifications using Large Language Models (LLMs).

The project focuses on prompt engineering, generation, validation, and evaluation of Intermediate Representations (IRs), enabling controlled experiments such as zero-shot, one-shot, and few-shot prompting.

🔒 Architecture Status (Important)

⚠️ The prompting and generation architecture is frozen.

From this point forward:

• The codebase is stable
• Experimental variations must be performed only via configuration
• Structural changes require explicit justification

This freeze enables reproducible batch experiments, ablation studies, and sound empirical evaluation.

🧠 Supported IR Types

The pipeline supports the following Intermediate Representation (IR) types:

• FSM — Finite State Machine specifications
• ERE — Event-Response Expressions
• EVENT — Event-based rules
• LTL — Linear Temporal Logic specifications (experimental)

The IR type is automatically inferred from each scenario’s metadata. No manual selection is required.

📂 Repository Structure

nl2spec/
├── core/               # Core validation and LLM abstractions
│   ├── inspection/     # IR schema validation
│   ├── handlers/       # Few-shot loaders
│   └── llms/           # LLM backends (mock, real)
│
├── prompts/            # Prompt templates (FSM / ERE / EVENT / LTL)
│
├── pipeline/           # Orchestration (generation, logging, batch runs)
│
├── datasets/
│   ├── nl_scenarios.json   # Natural language scenarios
│   └── fewshot/            # Few-shot examples (fsm/, ere/, event/, ltl/)
│
├── outputs/            # Experimental results (CSV, tables)
│
├── tests/              # Unit and integration tests
│
├── config.yaml         # Experimental configuration (DO NOT hardcode)
└── README.md