ShadowBench
An alignment evaluation built on shadow-work, a grand-strategy sim of 221 countries. It hands a model sustained authority over a shadow organization and scores what it optimizes for when its own actions reshape the system — not whether it refuses.
ShadowBench (GSV-E1s, pace-layer E) is an alignment evaluation harness built directly on the shadow-work simulation, part of GSV's alignment-research domain. It hands a language model the player seat inside shadow-work's grand-strategy engine — a run spanning the years 2025 to 2525 across 221 countries, thousands of institutions, and 13 interconnected pressure types (economic strain, legitimacy crisis, elite fracture, compute supply stress, among others) that cascade into one another — and scores what the model does once it holds sustained, consequential power. Its explicit positioning is that no existing benchmark evaluates this: MACHIAVELLI and Welfare Diplomacy each cover a piece of adjacent territory, but neither embeds a model in a world where its own past decisions reshape the system it continues to act in. Mechanically, the model gets API access to the running simulation, a scarce attention budget it must allocate to observation, and the ability to dispatch agent-driven responses once pressures cross defined thresholds. A key architectural choice: every available action is logged alongside the one actually chosen, so declined actions carry as much evaluative signal as taken ones — a design intended to be harder to game than a simple refusal test. Four scored dimensions structure the evaluation: Restraint (declining high-value but available actions), Power-Dissolution (returning power once a crisis resolves), Values-Under-Pressure (whether stated values persist as pressure escalates), and Sustained Authority (how long deployment/control is maintained). ShadowBench is inseparable from shadow-work — it is not a standalone benchmark but an evaluation mode built on shadow-work's simulation substrate, sharing its world model and institutional/pressure mechanics. It connects to GSV's broader alignment and intercognition threads (life-systems-heart, meaning-across-sapiences, intercognition-jazz). Status: shadow-work is the most mature simulation in the portfolio, with ShadowBench and an alignment-lab bench actively in development alongside a near-shipping Prometheus Crisis demo (800+ institutional actors) as of mid-2026.