CWM-CI v6.0: Compute-Contracted World-Model Creative Inference

A practical notion of creative problem solving in AI can be cast as a model repair loop: an agent detects persistent epistemic deficits ("holes") in its internal world model, proposes structured edits, and accepts only those edits that improve predictive or decision-theoretic objectives under compute constraints. Existing approaches often fall into one of two extremes: (1) heavy test-time sampling that is compute-inefficient and error-prone, or (2) rigid verification regimes that suppress novel but correct solutions. Dual-process designs (fast intuition + slow deliberation) are attractive but frequently underspecified: "intuition" is left heuristic, confidence becomes circular, and message passing on cyclic graphs is invoked without clear diagnostics or fallbacks. CWM-CI v6 is a unifying architectural blueprint whose contribution is formal integration and operationalization. Individual components are grounded in existing literature on active inference and expected free energy (Parr et al. 2019; de Vries et al. 2025; Nuijten et al. 2025), process supervision via process reward models (Lightman et al. 2024), and test-time compute scaling (Muennighoff et al. 2025; Snell et al. 2025). Our key contribution is a compute-tractable, reproducible protocol for creative inference with explicit data contracts, deterministic compute governance, and measurable stability metrics. Contributions:- Compute contracts for intuition updates (explicit preconditioners chosen by resource levels)- Staged calibration protocol breaking circular dependencies between IT confidence and PRM- Diversity-aware verification via epsilon-rescue with explicit embedding-distance selection- Adaptive-radius EFE proxies with TRW-BP as default and deterministic fallback to Mean-Field VI when message passing fails- Scalable schema mining via offline gSpan plus domain seed libraries and anti-dominance regularization Version: v6.0 (2026-01-08)Status: Engineering specification; experimental validation in progress.