An invisible space breaks AI eval graders in Claude Code, OpenEvals, Promptfoo
maverick_man1111 · reddit · 2026-10-28
A developer found Claude Code's native eval grader marks correct Python code wrong and vice versa due to a single invisible Unicode character (e.g. NBSP). LangChain OpenEvals, Promptfoo and Nvidia SkillEvaluator share the flaw.
The fix is small: auto-generate a note listing the exact Unicode characters and positions without altering the answer, which drastically improves judgments. The author reproduced it in Claude Code's plugin grader and submitted PRs to Nvidia, OpenEvals and Promptfoo; Nvidia has approved. Full report: driftproofhq.com/invisible-space
Related event: Invisible Space Character Tricks Multiple AI Graders Into Wrong Verdicts(3 posts)→
More from coding & agent
- Avalon-M1 trains LLM agent organisations via collective-reward RL, lifting avg score 33.5% to 45.7% — RobertTLange · 2026-10-28
- Creator chains Halo and Runway via MCP connectors in GPT for a music video workflow — aziz4ai · 2026-10-28
- Dev uses Qwen in an opencode harness to root an LG TV and strip out spyware — Switchboy1 · 2026-10-28
- Clawifi MCP server cuts agent web-reading tokens by 75%, beats Firecrawl on benchmarks — atakanelik34 · 2026-10-28
- Amazon's EMMA treats multi-agent configs as episodic memory, beating baselines by up to 32.5% — omarsar0 · 2026-10-28
- A 5-minute visual explainer of what happens when you call an LLM API — HowDevelop · 2026-10-28