Revision 2 · Web edition · 4 November 2026
IRIS™ User Manual
A guide to capturing training evidence and reading what the instrument measures.
Start with Chapters 1–4 to understand the measurements and their limits. For a first run, use Chapters 6–10. Chapters 11–15 cover operation and troubleshooting. Chapters 16–22 explain diagnostic views; Chapters 23–28 provide reference material. Appendix B offers prompts for a coding or research assistant.
IRIS reports measurements. A chart does not, by itself, establish model understanding, causation, safety, quality, failure, or the correct intervention. Keep trainer loss, held-out evaluation, and your experimental record beside it.
Contents
- Release NotesChanges by version
- Chapter 1What IRIS Is and Is Not
- Chapter 2The IRIS Data Model
- Chapter 3How IRIS Measures
- Chapter 4Safe Interpretation and the First 15 Minutes
- Chapter 5Dashboard Overview
- Chapter 6Installation
- Chapter 7Configuration
- Chapter 8Build the Model Spec and Probe Set
- Chapter 9Minimal Tracker Integration
- Chapter 10Launch the Bench and Verify Capture
- Chapter 11Runtime Control, Signals, Metrics, and Logs
- Chapter 12Integration Patterns by Model Family
- Chapter 13Distributed Training
- Chapter 14Artifacts, Reports, Export, and Security
- Chapter 15Troubleshooting
- Chapter 16Whole Model and Layer Structure
- Chapter 17Domain Training and Affinity
- Chapter 18Attention-Head Diagnostics
- Chapter 19MoE Expert and Router Diagnostics
- Chapter 20Facets and Emergent Pathways
- Chapter 21Gradient Residuals, Stability, Curvature, and Update Efficiency
- Chapter 22Adaptive Cadence and Event Detection
- Chapter 23CLI Reference
- Chapter 24Configuration-Key Reference
- Chapter 25API Reference
- Chapter 26File and Schema Reference
- Chapter 27Glossary and Concept Index
- Chapter 28The Flow Map
- Appendix ADiagnostic Availability Matrix
- Appendix BLLM Prompt Library
