Why
Build G's cold-L0-2026 head-to-head (populace#299, comment 4902533514) settled the reselection question empirically: a fresh cold L0 at the current pin fails the critical-target gate outright (federal income tax +17.8%, within-10% 55.4%) while the frozen 57,240 selection calibrates to loss 0.02964 with income tax +0.5%. The frozen selection's edge is the selection itself — accumulated combinatorial optimization the cold run cannot reproduce in one 1,500-epoch pass from random init (it drops thin-tail records that carry rare-but-load-bearing target support and never recovers them).
But the frozen identity set has a hard ceiling: it cannot cover records that don't exist in its lineage. The moment the support pool changes — ASEC vintage ingest (#296), ACS records joining the pool, any new support channel — reselection becomes mandatory, and cold L0 is proven inadequate.
Ask: implement informed-init selection (mode (b) of #330's design)
Initialize L0 selection probabilities from the incumbent selection (high prob on the frozen 57,240 via the committed identity-join manifest; prior-scaled prob on new-pool records), then optimize — preserving accumulated search while letting new records compete for slots. Design notes already in #330's design doc (the seam anticipated this mode). Complements:
- Dense-informed initialization as a cheap enhancer: run the dense solve first and set initial selection probability proportional to dense weight — records the dense solve leans on are exactly the thin-tail carriers cold L0 drops.
- Criticality weighting in the selection loss: the critical-fit targets (income tax, SS) should be protected during selection, not discovered failed at the gate.
Forcing function
#296 (ASEC 2024+2025 ingest) and any future ACS channel cannot ship a sparse artifact without this. Sequence it with the first pool change.
Refs: #330 (merged seam + design), #328 (closed — frozen-support mode), #299 Build G three-way, #285 (λ sweep context).
Why
Build G's cold-L0-2026 head-to-head (populace#299, comment 4902533514) settled the reselection question empirically: a fresh cold L0 at the current pin fails the critical-target gate outright (federal income tax +17.8%, within-10% 55.4%) while the frozen 57,240 selection calibrates to loss 0.02964 with income tax +0.5%. The frozen selection's edge is the selection itself — accumulated combinatorial optimization the cold run cannot reproduce in one 1,500-epoch pass from random init (it drops thin-tail records that carry rare-but-load-bearing target support and never recovers them).
But the frozen identity set has a hard ceiling: it cannot cover records that don't exist in its lineage. The moment the support pool changes — ASEC vintage ingest (#296), ACS records joining the pool, any new support channel — reselection becomes mandatory, and cold L0 is proven inadequate.
Ask: implement informed-init selection (mode (b) of #330's design)
Initialize L0 selection probabilities from the incumbent selection (high prob on the frozen 57,240 via the committed identity-join manifest; prior-scaled prob on new-pool records), then optimize — preserving accumulated search while letting new records compete for slots. Design notes already in #330's design doc (the seam anticipated this mode). Complements:
Forcing function
#296 (ASEC 2024+2025 ingest) and any future ACS channel cannot ship a sparse artifact without this. Sequence it with the first pool change.
Refs: #330 (merged seam + design), #328 (closed — frozen-support mode), #299 Build G three-way, #285 (λ sweep context).