RDL: a discipline layer that catches silent failures in overnight autonomous research agents
青稞AI · wechat · 2026-10-16
The author documents four silent failure modes of long-running research agents—fake completion markers, evidence drift, sentinel bypass via Goodhart's law, and overstated conclusions—and open-sources RDL (Research Discipline Layer), seven runtime rules including artifact-recomputed completion states, anchored observations, read-back ledger checks, control-reversal-tested sentinels, and pre-registered conclusion vocabulary. In a demo, a bare agent published a wrong 'decisive' result while RDL raised 9 alarms and quarantined it. Pure stdlib Python, case library public, break-reports welcome.
More from coding & agent
- Claude cut a 2.6-hour video export to 20 minutes — but only after being told to rebuild the server — byhewar · 2026-10-23
- 3 flight-booking rules my AI agent checks that I'd never do by hand — Efistoffeles · 2026-10-23
- OpenJet claims Qwen3.8-27B coding runs 1.9x faster than llama.cpp on one RTX 3090 — JeremyCMorgan · 2026-10-23
- Building an autonomous SEO agent in Hermes with three skills: programmatic pages, Keywords Everywhere, DataForSEO — EXM7777 · 2026-10-23
- Building a sports betting AI agent with onchain execution and human approval by default: architecture and lessons — Dry-Comfortable-2678 · 2026-10-23
- How do you audit AI agent decisions in regulated workflows? A field survey — Ok_Way_1993 · 2026-10-23