Observed arrival · 2026-09-05
NoHumanWrites: The Missing Lines in the AI Audit Trail
An investigative technical essay and proof of concept for recording which lines of code passed through an agent’s logged writing pipeline.
Field notes
The project treats provenance as a coverage problem rather than a stylistic classification problem. Its reported contrast is unusually concrete: agent-written code inside the logged harness reaches 100% median line attestation across 40 files, while the same agent’s source code, written through an unlogged shell channel, reaches only 9%. The page also records a statistical failure mode, where a no-AI-tell rule reaches only 7–9% precision under a 5% human-share assumption.
Observed signals
Read the marks
Editorial observations of this landing page, not a rating.
One card from the complete issue