C.17:14 - SoTA-Echoing and source use
Source-use boundary. The sources below change what a C.17 user inspects or reports. They do not install one creativity theory, automated judge, metric, or search algorithm as the FPF default. Reopen this source-use judgement when a cited source is corrected, retracted, or materially superseded; when new cross-domain evidence overturns one of the stated consequences; or when a proposal would make one automated metric, corpus, encoder, QD descriptor, or proxy score normative. Use G.11 for that refresh.
| Current practice and source | Source-use decision | Concrete C.17 consequence |
|---|---|---|
Judge novelty and usefulness as distinct, context-dependent questions. Harvey and Berry, Toward a Meta-Theory of Creativity Forms: How Novelty and Usefulness Shape Creativity, Academy of Management Review 48(3):504-529 (2023), DOI 10.5465/amr.2020.0110; Sen et al., Automated Creativity Evaluation of Language Models Across Open-Ended Tasks, ACL 2026, DOI 10.18653/v1/2026.acl-long.1061. | Adopt the separation of creative breadth from task fulfilment and the dependence of usefulness on the practical situation. Adapt it by using the named comparison basis, Use-Value, and ConstraintFit already defined here. Reject a context-free creativity score and the cited paper’s particular semantic-entropy or automated-judge machinery as universal FPF measures. | The first move names both what the bearer differs from and which objective or must-criterion matters. Approval-facing use cannot substitute Novelty for Use-Value or ConstraintFit; CC-C17-1, CC-C17-5, and CC-C17-8 test that boundary. |
Preserve a collection of different locally strong alternatives when the question needs coverage rather than one winner. Qin et al., A survey on Quality-Diversity optimization: Approaches, applications, and challenges, Swarm and Evolutionary Computation 100:102240 (2026), DOI 10.1016/j.swevo.2025.102240. | Adopt the quality-diversity and illumination insight that coverage and local quality can matter together. Adapt it as Diversity_P, Illumination, or another declared retained-set reading. Reject MAP-Elites, a QD score, feature descriptor, container, or quota as a default C.17 method or selection rule. | A set reading stays telemetry with its bearer, rule, Scale, and evidence. Use C.18 for Archive and Front maintenance, C.19 for pool treatment, and G.5 for selector-facing declarations; CC-C17-9 and CC-C17-10 keep these moves separate. |
Test whether a result survives reasonable corpus, metric, and representation choices. Lu et al., Rethinking Creativity Evaluation: A Critical Analysis of Existing Creativity Evaluations, EACL 2026, DOI 10.18653/v1/2026.eacl-long.297; Stein et al., Exposing Flaws of Generative Model Evaluation Metrics and Their Unfair Treatment of Diffusion Models, NeurIPS 2023, DOI 10.52202/075280-0165. | Adopt sensitivity checking and comparison with evidence suited to the receiving domain. Adapt it by making the corpus, inclusion rule, Method, model or encoder, distance, calibration, and uncertainty part of the claim. Reject transfer of one metric across domains, minor prompt or implementation stability as validity, and leaderboard standing as evidence of the bearer characteristic. | For a load-bearing value, inspect neighbours and run the applicable corpus/Method sensitivity or invariance probe. If the conclusion changes, report that dependence or incomparability instead of hiding it in one score; CC-C17-4 and CC-C17-11 make this visible. |
Check whether optimizing a proxy stops improving the result that matters. Gao, Schulman, and Hilton, Scaling Laws for Reward Model Overoptimization, ICML 2023, PMLR 202:10835-10866, https://proceedings.mlr.press/v202/gao23h.html. | Adopt the warning that further proxy optimization can reduce performance under a separate target judgement. Adapt it by retaining primitive coordinates, gates, evidence, and held-out or delayed observations and by giving the local result a stop or reopen condition. Reject the reward-model setting or its fitted scaling law as a universal degradation model, and reject a rising proxy score as evidence that the bearer improved. | The design and policy cases keep tooling and legal or equity gates visible even when another value improves. Scalarization never erases the primitive coordinates; later target evidence can reopen only the claims that relied on the proxy. |
These decisions reinforce the existing route rather than add another assurance layer. In the pump case, inspect the admitted design set and tooling constraint; in the hospital case, keep the held-out result separate from unsupported transfer; in the policy case, keep the missing subgroup evidence as an eligibility gap. The source-use decisions therefore change the comparison and robustness work already required by the cases and checklist, not the practitioner-first entry.