Chapter 16 — Whole Model and Layer Structure
Start with the outcome curves
Keep trainer-reported training and held-out outcomes visible while reading internal structure. Aggregate accuracy can conceal differences between domains; probe metrics can differ from a held-out task measure. Check the source and unit of every curve before comparing its timing with another.
Read layer structure in context
Use k95, spectral entropy, spectral slope, and the spectrum together. They describe concentration and distribution in the measured representation. A layer with a lower k95 is not necessarily better, finished, unused, or less capable. The result depends on the probes and measurement available on that frame.
Basis rotation and principal-direction carry-over add a temporal view. An abrupt change locates an interval to inspect; its cause needs evidence from the run record and independent tests. Compare corresponding layers and actual captured steps, rather than screen positions in an animation.
Resources and provenance
A figure may state that it uses reduced or projected rows, a carried basis, or a restricted set of captures. Retain that provenance in interpretation and exported reports. Different capture modes can affect last digits or availability, so check those states before making a fine numerical comparison.
Older bundles can carry explicit reliability notices. In particular, the manual identifies affected Apple-silicon spectra recorded before the 20 October 2026 tracker fix. Read the notice attached to the actual run; an affected spectrum is not repaired by presenting it in a newer dashboard.
Questions to investigate
- Which layers changed, and over which measured interval?
- Is concentration changing together with orientation or motion?
- Does the change appear on several probe domains or only one?
- What did trainer-reported held-out outcomes do over that interval?
- Did the run resume, change data, change optimizer, or change capture mode?
Those questions guide inspection. They do not turn an internal pattern into a universal predictor of generalization, quality, or failure.
