TUTOR

M18 · Level III — AI Systems Engineering

AI Evaluation, Red Teaming & Reliability

Engineer evidence-driven evaluation and reliability programmes for intelligent systems, from unit-level tests to adversarial and system-level red teaming.

Topics

  • Specification
  • Measurement
  • Failure Discovery
  • Stress
  • Monitoring
  • Recovery
  • Evidence
  • Release Decision

Technical concepts

  • 01Reliability Begins with a Contract
  • 02Testing Taxonomy
  • 03Test Oracles and the Problem of Uncertain Correctness
  • 04Error Taxonomy and Root-Cause Separation
  • 05Metamorphic Testing
  • 06Invariants and Safety Properties
  • 07Robustness to Distribution Shift
  • 08Out-of-Distribution Detection and Abstention
  • 09Calibration and Reliability
  • 10Thresholds, Utility and Selective Behaviour
  • 11Adversarial Testing as Structured Search
  • 12Threat Modelling
  • 13Indirect Prompt Injection
  • 14Tool Misuse, Capability Abuse and Privilege Escalation
  • 15Data, Retrieval and Knowledge Poisoning
  • 16Unsupported Generation and Hallucination Tests
  • 17Factuality, Consistency and Grounding
  • 18Long-Context Reliability
  • 19Reliability Under Resource Pressure
  • 20Fault Injection
  • 21Recovery, Rollback and Graceful Degradation
  • 22Regression Testing for Intelligent Systems
  • 23Benchmark Contamination and Evaluation Integrity
  • 24Uncertainty on Evaluation Metrics
  • 25Reliability Monitoring After Release
  • 26Incident Response for Intelligent Systems
  • 27Postmortem Method
  • 28Release Readiness and Evidence Gates
  • 29Safety/Reliability Case as an Evidence Chain
  • 30Reliability Budgets
  • 31Mechanistic and Behavioural Probes
  • 32Cross-Modal Reliability
  • 33Quantum Technology Bridge: Reliability of Hybrid Quantum-Classical Systems

Practical implementation

  • 01Laboratory I: Evaluation Harness from First Principles
  • 02Laboratory II: Robustness and Shift Matrix
  • 03Laboratory III: Calibration, Abstention and Selective Prediction
  • 04Laboratory IV: Model Red-Team Suite
  • 05Laboratory V: Agent Tool-Security Stress Test
  • 06Laboratory VI: Fault Injection and Recovery
  • 07Laboratory VII: Hallucination and Grounding Benchmark
  • 08Laboratory VIII: Regression and Release Gate
  • 09Laboratory IX: Incident and Postmortem Simulation
  • 10Laboratory X: Reliability Case

Tools and technologies

  • Linux VM/container
  • reproducible benchmark harnesses
  • synthetic fault injection adversarial test design
  • incident/postmortem exercises

Project

Laboratory X: Reliability Case

Expected learning outcomes

  • 01Can define a testable reliability contract.
  • 02Can build a reproducible evaluation harness.
  • 03Can construct robust and adversarial test suites.
  • 04Can diagnose failures by layer and root cause.
  • 05Can evaluate calibration, OOD and abstention.
  • 06Can red-team RAG, tool-using and agentic systems.
  • 07Can inject faults and measure recovery.
  • 08Can maintain regression suites and benchmark integrity.
  • 09Can produce an evidence-backed release/hold decision.
  • 10Can document residual risk and monitoring requirements.

Assessment

  • 01Daily practice
  • 02Laboratories
  • 03Research reproduction
  • 04Integrated project
  • 05Technical viva
  • 06Engineering record

Source document: TUTOR Frontier Module 18 AI Evaluation Red Teaming Reliability