A lightweight evaluation framework for LLM systems combining:

• deterministic evaluation (critical gates + heuristics)
• probabilistic evaluation (LLM-as-judge)
• regression detection across runs

LLM systems are non-deterministic.

Traditional pass/fail testing is not enough.

This project explores a layered evaluation approach:

Dataset
↓
System Under Test Response
↓
Critical Gates (hard constraints)
↓
Heuristic Scoring (deterministic)
↓
(Optional) Judge Ensemble (LLM)
↓
Regression Comparison

The project now supports multiple evaluation domains via self-contained example packs.

• recommendation-style evaluation
• structured list outputs
• qualitative scoring (taste, tone, diversity)
• baseline reference task

examples/wine_recommendation/

Demonstrates:

• recommendation tasks
• support assistant evaluation
• retrieval-grounded responses (RAG-style)
• simple agent workflows (mock tools)
• structured output expectations

examples/retail_support/

.
│
├── examples/
│   ├── wine_recommendation/
│   └── retail_support/
│
├── configs/
│   ├── tasks/
│   ├── systems/
│   └── judges/
│
├── results/<task_name>/
├── baselines/<task_name>/
│
├── runner.py
├── scorer.py
├── task_loader.py
├── tool_simulator.py
├── schemas.py
├── regression_compare.py

python3 runner.py --task-config configs/tasks/wine.yaml --mode mock

python3 runner.py --task-config examples/retail_support/task_config.yaml --mode mock

⸻

python3 runner.py --task-config configs/tasks/wine.yaml --mode mock --write-baseline
python3 runner.py --task-config examples/retail_support/task_config.yaml --mode mock --write-baseline

⸻

Using explicit paths:

python3 regression_compare.py baselines/wine_recommendation/baseline_results.json results/wine_recommendation/latest_results.json

Using task shortcut:

python3 regression_compare.py --task wine_recommendation
python3 regression_compare.py --task retail_support

⸻

•	How to evaluate LLM outputs beyond simple correctness
•	How to combine heuristics and LLM judges
•	How to detect regressions in non-deterministic systems
•	How to design evaluation datasets and rubrics
•	How to structure reusable evaluation tasks

⸻

This is a V1 learning lab project.

Focus: • clarity over completeness • simplicity over abstraction • experimentation over production design

⸻

•	judge disagreement visualization
•	evaluation analytics across runs
•	cross-model judge comparison
•	richer agent workflow evaluation
•	dashboard / visualization layer