Inspect a model's config and route it to the right signal set per architecture family (dense / MoE / hybrid-attention / etc.) instead of one-size-fits-all.
Milestone 1 hard-codes a dense-only allowlist (FR-3, #6) and abstains on everything else — this would eventually replace that gate with pluggable per-family signal sets.
Iced because: Milestone 1 only needs one architecture family to prove the approach — this isn't worth building until there's a second family (MoE, see the Milestone 2 candidate) actually implemented.
Inspect a model's config and route it to the right signal set per architecture family (dense / MoE / hybrid-attention / etc.) instead of one-size-fits-all.
Milestone 1 hard-codes a dense-only allowlist (FR-3, #6) and abstains on everything else — this would eventually replace that gate with pluggable per-family signal sets.
Iced because: Milestone 1 only needs one architecture family to prove the approach — this isn't worth building until there's a second family (MoE, see the Milestone 2 candidate) actually implemented.