A.19.USCM:11 - SoTA-Echoing
SoTA vs popular note. This section records alignment to post‑2015 evidence‑backed practice. It is not a mandate to use fashionable methods; method semantics stay in SoTA packs (G.2) and wiring modules, while this pattern fixes the stable mechanism boundary.
A.19.USCM:11.1 - SoTA alignment map
| SoTA practice pointer, post‑2015+ | Primary source examples, post‑2015+ | Where it connects to USCM | Adoption status |
|---|---|---|---|
| Prefer monotone and interpretable scoring surfaces where appropriate | Explainable additive and monotone model lines, e.g., Lou et al. 2016; Nori et al. 2019; monotone deep lattice style models, e.g., You et al. 2017 | Expressed as admissibility‑bounded transform freedom via CGSpecSlot.SCP and explicit scalarization rules; method details stay out of the kernel | Adapt |
| Treat probabilistic scores as measures requiring calibration, not raw outputs | Calibration practice, e.g., temperature scaling (Guo et al. 2017) and successors | Expressed as “score is a measure on an explicit scale,” bounded by SCP+CSLC and evidence gating; calibration itself is wired as method semantics, not kernel law | Adapt |
| Keep uncertainty explicit and allow set‑valued scoring when appropriate | Modern conformal prediction practice, e.g., Romano et al. 2019; Barber et al. 2021 | Expressed as “vector scores allowed; unknown not coerced; no hidden scalarization,” enabling downstream set‑valued comparison/selection | Adapt |
| Keep architectural commitments traceable to one governing pattern | ISO/IEC/IEEE 42010:2022 architecture description discipline | Expressed as explicit governing-pattern assignment and Tell+Cite stubs elsewhere (no competing semantics) | Adopt |
Notes per row
- USCM does not “implement a particular scoring model”; it preserves a stable, admissibility‑gated surface on which such models can be wired.
- Calibration is treated as a lawful transform family that must live within SCP+CSLC; the kernel does not mandate a specific calibration method.
- Set‑valued scoring aligns with USCM’s “vector first, scalar by declaration” law, and is naturally consumed by CPM/SelectorMechanism without forcing a spurious total order.
- Governing-pattern traceability is used here to keep the spec teachable and non-duplicative; it does not add new governance cards or admissibility gates.