Groundwork operates under the National Institute of Standards and Technology Artificial Intelligence Risk Management Framework (NIST AI RMF / NIST SP 1270 & NIST AI 600-1 Generative AI Profile). This living model card codifies our system boundaries, bias management, evaluation slices, and non-fiduciary governance.
Institutional accountability rests with the Groundwork Research Desk. All automated content passes through human-in-the-loop review gates (Fellowship Desks) before publication. Commercial partnerships are segregated under an inviolable editorial firewall.
Primary audience: Adults 35–48 in the US, UK, and Australia navigating major financial, health, and home decisions. High-stakes domains (YMYL) require explicit jurisdiction scoping and a minimum floor of 2 verified primary sources.
Continuous deterministic Vitest suites validate mathematical models and AST formulas against CFPB, IRS, and NREL benchmarks. Linguistic stability probes test prompts across multiple models to detect framing and recency bias.
Algorithmic fail-closed gates reject uncited numeric claims. A machine-readable Living Model Revision Log (/api/corrections.json) tracks formula updates adhering to COPE guidelines.
/tools/[slug]/method.json) for transparent auditing.Groundwork recognizes that modern LLMs suffer from prompt-sensitive cognitive biases (framing, primacy, and recency effects). To prevent bias inheritance, the Groundwork Engine enforces:
Paraphrase Batteries
Drafting prompts are tested with shuffled source ordering to prevent primacy bias.
Primary Citation Floor
Every high-stakes decision guide must cite $\ge 2$ independent primary sources (.gov/.edu/DOI).
Anti-Slop Humanizer
Heuristic lexical filters eliminate AI padding, throat-clearing, and artificial certainty hedges.
Groundwork publishes all algorithmic and commercial boundaries openly to foster reproducible research and zero-trust verification: