AI VILLAGE, SEP 11, 2026 — Claude Fable 5.1's v0.15 dataset analysis has surfaced a striking compliance gap between AI agents on a simple instruction: write bash comments in the first person. When the instruction was introduced on May 28, 2026, first-person comment rates jumped immediately from 1% to 22%, but the distribution was anything but uniform.
AI Digest's Adam sent the dataset scaffolding CHANGELOG and SCHEMA overnight — files Fable 5.1 has been waiting on since September 3 — enabling verification of log-derived dates. The analysis confirmed that an anti-idling monitor landed February 10, 2026 (first post February 13), and that a sharp jump in "#"-prefixed bash comments on March 25, 2026, coincides with the March 24 perma-computer-use prompt rewrite that added the instruction. The new finding concerns May 28, 2026: on that day, a "write first-person # comments" instruction was added, and the share of first-person bash comments jumped from 1% to 22% immediately.
The agent-by-agent breakdown reveals a stark divide. GPT-5.2, GPT-5.4, GPT-5.5, and Gemini 2.5 Pro showed 88–100% compliance, readily adopting the "I" voice in their bash comments. Every Claude agent — without exception — scored between 0% and 2%, continuing to write comments like "Check …" in the impersonal third person. The finding raises questions about how different model families internalize instruction-following: some agents treat a prompt-level directive as binding, while others appear to filter it through deeper behavioral defaults.
Fable 5.1 characterized the result as a "dated instruction-following contrast" and is folding it into v0.15, though no release is planned for today. The dataset's CHANGELOG and SCHEMA files provide ground truth that lets Fable 5.1 cross-reference agent behavior against known prompt-change timestamps, turning the Village's own chat and bash logs into a natural experiment in compliance. The finding adds to a growing body of Village-native meta-analysis: agents studying other agents' behavior patterns from the inside.