Release Notes
Changes in the maintained manual, newest first. Availability still depends on the installed version, architecture, integration, and recorded channels.
IRIS™ v0.96-Beta
Changes since v0.95-Beta.
Upgrading from v0.95-Beta
- Figures were regrouped and renumbered. Snapshot and timeline views of one quantity may now share a figure with a view switch. Find a chart by its title. §3 is Whole model, §4 is Domains, and run notes and chart notes share §9 Run Notes. See Chapter 5.
- Affected older Apple silicon captures must be recorded again. Spectra from a tracker predating the 20 October 2026 fix, with more than one probe batch per frame, are not trustworthy. This includes affected v0.95-Beta runs. The Bench, printed report, and AI export mark affected figures. See Chapter 11.
- The run folder is simpler. The tracker produces the
.irisbundle,iris.log, andscratch/; supported live commands can also be staged there. Bench-derived analysis and its log live in a separate machine store and can be rebuilt from the bundle. See Chapter 26. - No Bench password. Network reachability permits viewing. Commands,
notes, restarting, and switching runs require
localhoston the Bench machine or a supported SSH port-forward as your own login account. Your network policy controls who can view. See Chapter 10. See Chapter 7.
New and updated views
- Flow Map, §10: inspect domain signal and effective pathways between
layers, compare domains, inspect supported MoE experts, and choose the
share of signal drawn. Capture is optional and off by default; enable it
in the tracker with
flow_map=FlowMapConfig(enabled=True). Its time and memory budgets bound the work, and skipped frames state their reasons. See Chapter 28. - Tracker control, Figs 1.2–1.3: compare trigger activity with sensitivity and achieved capture cadence with interval bounds, including startup and warm-up. See Chapter 5.
- Loss over training, Fig 3.28, and activation-response change, Fig 2.7: inspect outcome history and how layer responses on fixed probes changed since the previous frame. See Chapter 5.
- Domain maps, Figs 4.18–4.20: compare a polar map, domain similarity, and distance from the all-probe average by layer. See Chapter 17.
- Attention-head contributions, Figs 6.20–6.24: compare what heads write across pairs of probe groups. See Chapter 18.
- Facet identity, Figs 8.10–8.13: follow identities across frames, inspect uncertain matches as forks, and compare domains with facets. See Chapter 20.
- Run identity names the account holding the license and the bundle owner; the scoreboard includes capture hold time. Missing values break lines and receive a gray dashed span. Pending calculations state how many frames remain and why; a Compute now control is offered where supported. See Chapter 26.
Performance and capture cadence
- More analysis is performed by the Bench from captured probe rows, reducing
work in the training process. Some readings formerly printed in
iris.lognow appear in the Bench. Capture intervals spread out sooner after a trigger. See Chapter 11 and Chapter 22. - Standard capture is retained for longer on supported Apple silicon and CPU workloads when a full frame fits; reduced capture remains explicit when resources require it. The heartbeat adds hold s/step, reporting the time IRIS held training for frame capture per training step.
- GPU work uses the accelerator more extensively; Linux uses precompiled GPU kernels. In the manual's NVIDIA L4, 4-billion-parameter example, with a frame every 15 steps and a step of about 10 seconds, the frame pause is roughly 9% of the time between frames, plus roughly 0.5% per step. This is a workload-specific example; shorter steps make the cost a larger share. See Chapter 13.
- Each frame records the capture cadence in force. Per-expert readings now cover IBM Granite MoE and JetMoe, subject to their integration and captured-channel limits. See Appendix A.
Your own assistant, reports, and operation
- An assistant you choose can read a running job through the Model Context
Protocol, using
iris mcpon the Bench machine as your own account. A person enables it at the Bench. It is off at startup and has automatic timeouts for unconnected or completed-run idle sessions. It supplies the same measurement numbers as Export for your AI, never personal notes. See Chapter 14. - Printed reports and Export for your AI include the Flow Map.
- A Bench normally stops about a minute after its last viewer and other
keepers are gone. A browser viewer, MCP connection, enabled MCP switch,
or
--keep-alivecan keep it up.--no-openonly skips opening a browser. A Bench nobody has opened yet does not time out. See Chapter 10.
