Chapter 1 — What IRIS™ Is and Is Not
IRIS Customer Guide
Intrinsic Rank Inspection System — inspecting model training from the inside.
IRIS helps AI teams inspect what is happening inside a model during training.
Most training tools show external signals: loss, accuracy, throughput, GPU use, and logs. Those are important, but they do not show how the model's internal representations are forming, compressing, separating, rotating, or reorganizing.
IRIS adds that missing layer. It captures diagnostic snapshots during training and turns them into visual measurements of internal structure, motion, and domain interaction.
IRIS helps researchers ask better questions:
- Which layers are carrying structure?
- Are domains separating or overlapping?
- Are some domains pulling against each other?
- Which heads, experts, or layers carry concentrated motion or residual activity?
- Is the model reorganizing smoothly or sharply?
- Which probes deserve closer inspection?
IRIS is an inspection instrument. It reports measurements rather than recommendations, giving research teams evidence they can use to decide what to investigate next.
What IRIS Does
IRIS observes a model while it trains. At selected frames, it runs a fixed set of labeled inputs, called probes, through the model and records how the model represents them internally.
From those snapshots, IRIS measures:
- layer structure and effective dimensionality
- concentration or distribution of spectral energy
- internal rotation over time
- domain separation, overlap, affinity, or opposition
- residual concentration by layer, head, or expert
- parameter or covariance-space motion
- optional user-defined training signals
IRIS presents these measurements in a live dashboard and post-run reports.
How to Use IRIS Measurements
IRIS measurements support inspection, comparison, and audit.
A chart may show that a layer is concentrated, that domains pull in different directions, or that a group carried more residual activity during a run. Those readings identify places to inspect. They are not automatic conclusions about model quality, safety, understanding, or failure.
The research team interprets IRIS measurements in the context of its model, task, objective, and validation results.
Security and Customer Data
IRIS is designed to inspect model training without requiring customers to
upload full model weights. The tracker runs inside the customer's environment
and creates diagnostic artifacts such as reports, charts, .iris bundles,
probes, notes, derived analysis, and optional exported prompts.
These artifacts are not a full copy of the model, but they may include useful details about the run, such as architecture, probe labels, internal structure, training behavior, and user-defined signals — plus probe TEXT when the run used Saccadic's built-in probe set. A run on your OWN probe corpus is data-minimized: the bundle carries a verifiable reference to your prompt, never the prompt text itself, so your probes stay on your own disk. Treat IRIS outputs as internal model-inspection records, and review exported prompts before sharing them with outside systems.
The full security checklist — covering artifact handling, network posture, and review practices in detail — is in Chapter 14.
The Basic IRIS Workflow
A typical IRIS workflow has four steps.
1. Add the IRIS tracker. The tracker runs inside the training script,
observes selected training steps, and writes an encrypted .iris bundle.
2. Provide a model spec and probe set. The model spec describes the architecture. The probe set contains labeled inputs from the domains the customer cares about, such as arithmetic, code, grammar, legal text, medical text, customer support, reasoning, factual recall, safety prompts, or product-specific categories. Probes do not replace a validation set — they are used to inspect internal organization.
3. Train as usual. Training continues normally. IRIS observes and records diagnostic frames.
4. Open the run in the bench. Point the iris launcher at the run
directory to start the live dashboard:
iris /path/to/run-dir
The bench reads the diagnostic bundle captured during training — without reloading the model — and serves the dashboard, charts, and JSON data. While a run is still training, the same command attaches live and follows new frames as they land. Use the Export as PDF button in the bench to generate a PDF report from the captured bundle.
A Simple Way to Think About IRIS
A model trains in a space too large to see directly. IRIS projects parts of that high-dimensional activity into views humans can inspect.
These views do not show everything, but they preserve useful structure. When IRIS shows a cluster, rotation, compression, or concentration, it is showing a measured pattern in the model's internal geometry.
Loss tells you how the model scored. IRIS shows how the model organized itself while getting there.
