Copyright © 2026 Saccadic LLC · Patent pending.

Browse chapters
  1. Release NotesChanges by version
  2. Chapter 1What IRIS Is and Is Not
  3. Chapter 2The IRIS Data Model
  4. Chapter 3How IRIS Measures
  5. Chapter 4Safe Interpretation and the First 15 Minutes
  6. Chapter 5Dashboard Overview
  7. Chapter 6Installation
  8. Chapter 7Configuration
  9. Chapter 8Build the Model Spec and Probe Set
  10. Chapter 9Minimal Tracker Integration
  11. Chapter 10Launch the Bench and Verify Capture
  12. Chapter 11Runtime Control, Signals, Metrics, and Logs
  13. Chapter 12Integration Patterns by Model Family
  14. Chapter 13Distributed Training
  15. Chapter 14Artifacts, Reports, Export, and Security
  16. Chapter 15Troubleshooting
  17. Chapter 16Whole Model and Layer Structure
  18. Chapter 17Domain Training and Affinity
  19. Chapter 18Attention-Head Diagnostics
  20. Chapter 19MoE Expert and Router Diagnostics
  21. Chapter 20Facets and Emergent Pathways
  22. Chapter 21Gradient Residuals, Stability, Curvature, and Update Efficiency
  23. Chapter 22Adaptive Cadence and Event Detection
  24. Chapter 23CLI Reference
  25. Chapter 24Configuration-Key Reference
  26. Chapter 25API Reference
  27. Chapter 26File and Schema Reference
  28. Chapter 27Glossary and Concept Index
  29. Chapter 28The Flow Map
  30. Appendix ADiagnostic Availability Matrix
  31. Appendix BLLM Prompt Library
On this page

Chapter 14 — Artifacts, Reports, Export, and Security

What a run contains

The .iris bundle is the primary diagnostic record. The run directory holds the tracker log and runtime scratch files; the Bench keeps recomputable analysis separately on the machine. Notes, recordings, and exported reports are also model-inspection records to review before sharing. Chapter 26 lists their roles.

IRIS™ does not require uploading full model weights to inspect a local bundle. That does not make every artifact non-sensitive: labels, architecture, training behavior, probe text or references, notes, and trainer-supplied signals can reveal information about the model and data. Review files before sending them to another person or an external AI service.

Bundles and exports

A compatible Bench opens diagnostic bundles automatically; no user-managed bundle key is required. Control access to the files rather than assuming their filename or format is an access boundary.

Use Export as PDF for a report and the Bench's AI export for structured discussion. Confirm which run and frames are included. Some exports are full-run summaries while an individual figure may name a particular frame; those are different scopes. When sharing a chart, keep its caption, units, coverage, and missing-data explanation with it.

Let your own assistant read a running job

IRIS supports read-only measurement access through MCP. A person enables the Bench's MCP switch, then the assistant runs iris mcp on that machine as the same login account, without sudo or a service account. It is off at Bench startup and restart. With no assistant connected, MCP turns off after 15 minutes. After training ends, it also turns off when every connected assistant has made no calls for 15 minutes. A silent connected assistant may remain connected during an active run. --keep-alive does not keep MCP enabled. The assistant receives the measurement numbers available through Export for your AI, never personal notes. IRIS does not perform interventions on the assistant's behalf. Review what you share with your chosen assistant and retain human control over training changes.

Butler export

The optional iris-butler relay forwards captured scalars to supported experiment trackers such as W&B, TensorBoard, MLflow, or Aim. Configure its destination and credentials according to the installed guide. A run's iris.cfg is a Butler configuration and can override the global Butler defaults. Treat destination credentials as secrets and review the fields you are sending.

Network posture

The Bench has no password: anyone who can reach its port can view it. Writes and control actions require localhost on the Bench machine or a supported SSH port-forward through that machine's own SSH server as your login account. Tailscale SSH forwarding is viewing-only. Restrict viewing access through network policy and use a trusted, encrypted path. For the supported Tailscale workflow, use the machine's Tailscale name and follow your installation's network guidance.

Licensing uses network access separately from the local dashboard. Follow the installation and account instructions supplied with your distribution and check current readiness with iris --license. Do not interpret local bundle reading as a claim that every IRIS operation is offline.

Review before sharing or deleting

Check the run identity, notes, probe material, export scope, and recipient. Keep original captures when you need an audit trail. Read deletion confirmations carefully: stopping a Bench is different from deleting a recorded run, and an archived-run deletion is not reversible.