Chapter 3 — How IRIS™ Measures
Read the measurement before the interpretation
IRIS observes model activity on a fixed set of labeled probes and records diagnostic frames. Some channels describe representations, some describe parameters and gradients, and others are values reported by your training code. They answer different questions and should not be substituted for one another. A probe measurement is not an independently sampled held-out test.
Spectral structure
- k95 counts the leading spectral directions needed to account for 95% of measured variance. A lower value means greater concentration in that measurement, not automatically better learning.
- Spectral entropy describes how evenly variance is spread across spectral directions. Read changes alongside k95 and the spectrum.
- Spectral exponent describes the slope of the spectrum on log-log axes. Compare like-for-like layers, probes, and training conditions.
- Basis rotation describes changes in principal directions between recorded frames. It measures motion, not its cause.
Spectral-signature labels summarize the measured shape. BROAD_SPECTRUM
describes broad structure, ANISOTROPIC emphasizes uneven concentration,
RANK_ONE describes strongly concentrated structure, and BASIS_ROTATING
draws attention to orientation changes. UNCLASSIFIED is an honest absence
of a supported label. These are descriptive labels, not grades of a model.
Frontier, core, and tail
Frontier energy is a reading near the boundary between dominant spectral structure and weaker directions. Core, frontier, and tail views help compare where measured variance sits. Use the chart's labels and trajectory; a band becoming larger or smaller is not a diagnosis.
Motion and relationships
Representation-motion readings describe changes in the activations produced by the probes. Weight-motion readings describe changes in the model's parameters. They are different observables. A motion spike may be a useful place to inspect held-out evaluation, optimizer events, and your run notes, but it does not identify a learning mechanism.
Training affinity compares measured probe-gradient directions. Positive agreement and negative opposition concern those measurements and sampled inputs. They do not establish that two tasks always help or harm one another. The Flow Map describes measured links between neighboring layers; see Chapter 28 for its marks and limits.
Coverage matters
Keep the probe set, model family, layer identity, sampling cadence, and available channels in view. An absent, shed, inapplicable, or pending reading is not a zero. Compare runs only after checking that their measurements and definitions are comparable. Start with safe interpretation before drawing a conclusion from a changing spectrum.
