Follow the signal
The Flow Map shows measured connections between layers and experts for different input domains.
IRIS™ · Intrinsic Rank Inspection System
A closer look
at AI training.
Observe how a neural network’s internal structure changes while the run is still underway.
Explore IRIS ↗See a measured run ↓Same model. Different learning times.
In a modular-arithmetic experiment, the operations reached high held-out accuracy at different stages. The overall accuracy curve hides some of that progression.
Look inside the run
Measurements that help turn an observation into a testable question.
Follow the signal through the model
The accuracy curves show what the model achieves. The Flow Map adds a view of the measured signal carried between experts and layers during training.
Compare where domain paths overlap and how traffic is distributed across layers. Revisit the recorded run in IRIS to inspect how those patterns change over time.
Static capture of the IRIS Bench. The controls shown inside the image are part of the captured interface.
Inspect representation structure
Track how activation variation concentrates and how its dominant directions move during training.
Compare expert routing
Inspect which subjects reach each expert, separately for each layer, and how routing changes over time.
Observe collaboration and conflict
Compare sharing and training-gradient agreement across domains, alongside loss and task results.
Read the measurement in context. IRIS records observations. Use them with validation, gradients, and task results to decide what to test next.
A training movie, built from measurements
A consistent view during training, and a record to revisit afterward.
Capture
The tracker measures the model using consistent probe inputs. Capture intervals can be fixed or adapt to organizational activity.
Record
Measurements accumulate in a self-contained .iris file. Training checkpoints keep their usual role.
Inspect
The local Bench reads the growing record during training and opens it afterward, without loading the model.
Work in your own environment
IRIS fits existing PyTorch training loops. Capture and inspection can run on CPU, Apple MPS, or NVIDIA CUDA.
Support includes GPT / LLaMA / Mistral, BERT-family encoders, and supported mixture-of-experts models. User-defined architectures use a measurement adapter to declare capture points.
Coverage depends on model layout. Shared-weight configurations are currently unsupported.
Model and data handling
Model weights, activations, and probe text are not sent to Saccadic. Licensing and metering contact Saccadic.
Keep your existing experiment tracker
Optional configured relays send selected scalar measurements to Weights & Biases, TensorBoard, MLflow, and Aim.
Read, explain, and export
Each graph includes an Explain guide. Define labeled scalar signals, derive formulas from captured channels, and export custom charts as PNG images, Excel workbooks, or JSON data.
Research and joint evaluations
One model. One run.
One concrete investigation.
For teams training or fine-tuning models that need a clearer internal view. A joint evaluation can test whether the readings focus a real investigation and measure overhead in your workload.
Academic researchers are encouraged to request credit grants.
Commercial IRIS pricing depends on model size, with costs of just pennies per run for models below 2 million parameters. See the product site for pricing and licensing terms.
Discuss an evaluation ↗Product details ↗