sepia argues surface-level AI humanizers fail, citing a 61,608-story study where structure alone detects AI at 93.2% F1
Created 2026-08-28 and at 554 stars within a day, sepia is a portable Agent Skill for Claude Code, Codex, Grok Build and Antigravity built on StoryScope (arXiv:2604.03136, Russell et al. 2026), which found a classifier using narrative-structure features alone hits 93.2% macro-F1 on AI fiction, and that editing surface style away moves it only from 95.5% to 93.9%. It therefore repairs narrative architecture first (themes explained by the narrator, causally tidy single-track plots, emotion rendered only as bodily sensation, no real-world references, linear time), then discourse flow, then the usual vocabulary layer, with a 30-feature rubric and per-model fingerprint corrections for Claude, GPT, Gemini, DeepSeek and Kimi. Professional prose gets separate venue-matched rule files for release notes, PR replies, postmortems and tickets.
Source
↳ Follow the thread