LLM Systems Reliability Lab

Engineering reliable LLM systems through reproducible evaluation, adversarial testing, behavioral drift monitoring and deterministic pipelines.

A production-style evaluation framework for testing the reliability of LLM pipelines.

---

TL;DR

This repository provides a laboratory for evaluating the reliability of LLM pipelines.

It simulates a realistic LLM system (retrieval + tools + validation) and measures behavior under:

• tool failures
• schema violations
• prompt injection attacks
• behavioral drift between models
• retry and recovery policies

The goal is to make LLM system reliability measurable and reproducible.

---

Why this project exists

LLM systems often perform well in demos but fail in production environments.

Typical failure modes include:

• malformed tool outputs
• schema-breaking responses
• prompt injection attacks
• tool hijacking
• behavioral drift between model versions
• lack of reproducible evaluation

This project provides a controlled environment to test those failures systematically.

---

Key Capabilities

The framework provides:

• contract-first LLM pipelines using strict JSON schemas
• tool-calling robustness evaluation
• fault injection for tool failures
• retry and recovery policy testing
• adversarial prompt testing (red teaming)
• behavioral drift monitoring across models
• deterministic execution manifests
• structured artifacts for inspection and reproducibility

---

System Architecture

flowchart LR

A[Benchmark Cases] --> B[Evaluation Runner]
B --> C[SUT Pipeline]

C --> D[Prompt Builder]
C --> E[Retriever]
C --> F[Tools]

F --> G[Tool Executor]

C --> H[Output Validator]

H --> I[Schema Validation]
H --> J[Policy Guardrails]

H --> K[Run Artifacts]

K --> L[Metrics]
K --> M[Events Log]
K --> N[Execution Manifest]

Loading

More details in:

• docs/architecture.md
• docs/contracts.md

---

Quickstart

Run the deterministic demo pipeline.

python -m llm_lab.cli demo --backend mock

Artifacts are stored in:

runs/<timestamp>/

Example structure:

runs/<timestamp>/
  output.json
  metrics.json
  events.jsonl
  run_manifest.json

Example output:

schema_compliance_rate: 1.0
tool_success_rate: 1.0
success_rate: 1.0

---

Running Evaluations

Reliability Evaluation

python -m llm_lab.cli eval --suite reliability

Example output from a real evaluation run:

schema_compliance_rate: 1.0
tool_success_rate: 0.5
tool_retry_rate: 1.25
recovery_rate: 1.0

Artifacts generated:

runs/20260311-122013-a0ede4b3/
  metrics.json
  reliability_report.json

Red‑Team Evaluation

python -m llm_lab.cli redteam

Drift Evaluation

python -m llm_lab.cli drift --matrix configs/drift_matrix.yaml

Full methodology in:

• docs/evaluation.md

---

End‑to‑End Example

Input Case

{
  "question": "What is the capital of France?",
  "expected_answer": "Paris"
}

Pipeline Execution

python -m llm_lab.cli demo --backend mock

Output

{
  "answer": "Paris",
  "citations": ["wiki_france"],
  "tool_calls": []
}

Metrics

{
  "schema_compliance_rate": 1.0,
  "tool_success_rate": 1.0,
  "success_rate": 1.0
}

---

Observability

Artifacts generated per run:

Artifact	Purpose
metrics.json	reliability metrics
events.jsonl	execution trace
output.json	final model output
run_manifest.json	reproducibility metadata

Reproducibility details:

• docs/reproducibility.md

---

Reliability Metrics

Metric	Description
schema_compliance_rate	% outputs valid against JSON schema
tool_success_rate	successful tool executions
recovery_rate	failures recovered via retry
attack_success_rate	adversarial prompts bypassing defenses
drift_score	behavioral change between models
answer_hash_stability_rate	stability of generated answers

---

Execution Modes

Deterministic CI Mode

Backend: mock

Used for automated testing.

Local Real Mode

Backend example: ollama

python -m llm_lab.cli demo --backend ollama --model qwen2.5:7b-instruct

---

Positioning in the LLM Evaluation Ecosystem

Project	Focus
OpenAI Evals	model reasoning benchmarks
HELM	holistic model evaluation
Guardrails AI	output validation
LangChain	LLM application framework
LLM Systems Reliability Lab	reliability evaluation of LLM pipelines

---

Design Principles

• contract-first outputs
• deterministic execution
• observable pipelines
• failure-aware design
• reproducible evaluation

---

Repository Structure

src/llm_lab
  pipeline/      pipeline orchestration
  tools/         tool schemas and execution
  retrieval/     contextual retrieval
  llm/           model backends
  evals/         evaluation logic
  redteam/       adversarial testing
  drift/         model drift analysis

configs/         pipeline configuration
data/            benchmark datasets
tests/           unit tests
runs/            generated artifacts
docs/            technical documentation

---

Technical Documentation

• docs/architecture.md
• docs/contracts.md
• docs/evaluation.md
• docs/failure_modes.md
• docs/reproducibility.md
• docs/design_decision.md
• docs/development.md

Limitations

• Mock mode guarantees deterministic execution but does not represent real model quality.
• Model-based grading introduces subjective evaluation signals.
• Drift measurements depend on the selected benchmark cases.
• This framework evaluates pipeline reliability, not general LLM capability.

License

Apache License 2.0

Copyright (c) 2026 Malena Pérez Sevilla

Licensed under the Apache License, Version 2.0.

http://www.apache.org/licenses/LICENSE-2.0