SPEC

The shift from statistical models to large foundation models, reinforcement-learning policies, and agentic systems in investment management has outpaced the diagnostic vocabulary that allocators bring to manager selection. Existing artifacts approach the problem from four sides — manager-side model validation, allocator-side AI-claim detection, firm-side governance, and trading-system-side backtest hygiene — but none integrates them into a single, falsifiable, per-strategy diagnostic that an allocator can apply at the manager-selection layer.

SPEC — Specification, Performance diagnosis, Explanation capability, Configuration sensitivity — is a four-axis, architecture-agnostic, falsifiable due-diligence framework that resolves to a Y/N verdict per axis. For each axis the protocol defines a falsifiable test, cites 2–3 exemplar papers documenting the failure mode the test catches, and maps the axis to existing regulatory vocabulary (IOSCO PD684/788, SR 26-2, BoE-PRA SS1/23, ESMA Algo-Trading Briefing 2026, MAS FEAT + Veritas + MindForge, CFA Ethics & AI). Four anonymized audit case studies — one per axis — show what allocators would have approved without SPEC and what they declined with it. The appendix is a 68-question diagnostic library mapped to the four axes, intended for direct paste into IC memos and DDQs.