Chapter 20 — Facets and Emergent Pathways
What a facet means
A facet groups probes with similar recorded paths through the model's representations. Domains are human-assigned labels; facets are observed groupings. A facet can cross domain boundaries or share a domain with other facets. It is not an established semantic category or an identified causal circuit.
Read evidence before counting
The Bench can display a grouping without stating that it supports a meaningful facet count. Read the structure/evidence sentence. A count or identity can be absent because evidence is insufficient or still pending. Do not turn a null count into zero groups or describe every visible cluster as a distinct learned algorithm.
Probe diversity and coverage matter. Very small or degenerate sets can produce trivial or unstable groupings. Where a count is bounded or a view is restricted, retain that qualifier. A repeated result on one fixed set does not validate a semantic interpretation against labeled ground truth.
Identity across frames
An identified facet can be followed across recorded frames. An identity is different from a frame-local label. Read continuation, appearance, disappearance, ambiguity, split, and merge annotations as the Bench states them. A gap is not a measured transition, and provisional identities should not be presented as final when the run is still being read.
Similar subject mixes do not establish identity. Where the Bench reports ambiguity, preserve it instead of selecting one story. A split or merge label describes the observed grouping relationship, not a causal claim that a network circuit was created or destroyed.
Relationships and separation
Facet-pair gradient readings concern the sampled probe pairs behind them. Read their coverage and uncertainty; insufficient evidence is not zero agreement. A reported layer of separation is a measurement under a stated criterion, not proof that the model performs an operation at exactly that layer or nowhere else.
Availability
Facets require the appropriate per-probe representation data. Declared architectures have a separate availability limit in Revision 2: absent facet emission is not proof that such a model cannot form structure. Check the actual run's availability sentence.
Use facet views to choose probes and intervals for further investigation. Compare the proposed interpretation with independent behavior and controlled tests before naming a mechanism.
Facet identity over training — Figures 8.10–8.13
Follow observed facets across recorded frames rather than treating every snapshot as an unrelated clustering. An uncertain identity match is shown as a fork; the domain-versus-facet table compares supplied curriculum labels with observed groupings. Identity and count remain conditional on the stated evidence and captured data. They are not validated semantic classes simply because a drawing assigns them separate marks.
