EPI-Fi
ACTIVEFunctional Information (Fi) Score
A metric for functional information in persona-driven AI dialogue — does an agent hold its role, advance the task, and produce genuinely novel insight?
since 2025 · v1.0
// GOAL
Accuracy, fluency, and relevance miss what matters in multi-agent / persona systems. EPI-Fi measures functional information — whether an agent sustains its role, advances the task, and generates real novelty — applying Wong & Hazen's Law of Increasing Functional Information to dialogue.
// METHOD
Score 1–5 on three dimensions — Persona Fidelity & Persistence (PFP), Functional Coherence & Task Achievement (FCTA), Evolutionary Depth & Novelty (EDN) — into a weighted composite Fi = w_p·PFP + w_c·FCTA + w_n·EDN, with context-tunable weights (Balanced / Creative / Operational). Hybrid evaluation: three human raters (target κ > 0.7) cross-checked by automated signals (embeddings, in/out-of-persona classifiers, vector-trajectory mapping, an auxiliary-LLM "surprisingness" rating).
// OUTCOME
v1.0, released open for research and prototyped across Claude, Gemini, GPT, Grok, and Abacus; validation studies (cross-rater reliability, automated–human correlation, scaling beyond ~20-turn dialogues) are ongoing.