A.19.CPM:11 - SoTA-Echoing
The following research lines inform comparator choices. Their methods require their own conditions; no source family alone establishes the CPM outcome or admission rule.
Concrete comparator-family SoTA packages are cited through their current Part G pack or claim sheet when one governs the use. CPM’s kernel semantics remain unchanged.
| Research contribution | How it connects to CPM | Adoption status in FPF |
|---|---|---|
| Fair ranking and constrained ranking (e.g., Zehlike et al., 2017; Biega et al., 2018) | Reinforces the “no hidden tie‑breaks and thresholds” stance: fairness constraints belong in explicit comparator and acceptance policies, not as silent kernel constants. | Integrate via ComparatorSpec editions in CG‑Spec.ComparatorSet + policy pins; CPM remains unchanged. |
| Uncertainty-aware and set-valued inference (e.g., Romano et al., 2019; Barber et al., 2021) | Prediction intervals or other uncertainty results retain their own semantics. A declared comparator determines which relation tokens follow; an interval is neither a CPM token set nor by itself a reason for abstention. | Model as comparator families (or supporting method families) packaged in G.2; wired into declared ComparatorSpec. |
| Differentiable sorting and learned comparators (e.g., Grover et al., 2019; Blondel et al., 2020) | Replay and drift assessment use the learned ComparatorSpec edition and the comparison’s policy, argument and result bindings. Add an A.10 evidence-provenance account when the bounded source-to-use grounds must remain replayable. | Use a sorting or learned-comparison component only under a ComparatorSpec that establishes the returned relation semantics and required conditions; an algorithm label is insufficient. |