Engineer's Site AI Guardrail Can Only Refuse, Never Revoke: Fail-Safe by Design
TejasKumar_ · x · 2026-10-02
Tejas Kumar explains the core rule of the jev AI guardrail on his personal site: it can only add a refusal or an approval, never take one away — with no key, a timeout, or a weird response, the site behaves exactly as before. The model is pinned to jev-1.13.0 since all thresholds were measured on it. He also tried 16 prompts to detect AI-written text; the plainest one scored his 2020 human-written blog posts as more AI-like than actual AI text, so that feature never shipped.
More from coding & agent
- Review AI code by blast radius, not volume: a 6-step path to production-ready code — MaryamMiradi · 2026-10-02
- LangChain Academy updates Deep Agents course with one-command Managed deployment to Slack — LangChain · 2026-10-02
- Dev building GTA 4-style SCP game entirely with Claude Opus 5.5 shares first demo — imjustnewatai · 2026-10-02
- Claude Code Mods: spin off forked agents to build your own memory harness — trq212 · 2026-10-02
- The Hard Part of AI Memory Is the Write, Not the Read — Tiwaryswarnim · 2026-10-02
- Codeck presentation tool fully integrated inside Codex as a plugin, runs live prompts on slides — Dimillian · 2026-10-02