Hand-scoring my AI action checker: most checks were wrong more often than right
chich2017 · reddit · 2026-11-19
A developer followed up on their AI checker that flags when an assistant claims it did something but didn't — by hand-scoring its checks, with humbling results:
- One check judged the wrong party, mostly tracking the user's own promises instead of the assistant's.
- Another opined on work it couldn't see (files, commits, terminal commands outside the chat): 3 right, 7 wrong — retired.
- A nice-looking 10/0 score was withdrawn because the check counted things rather than verifying claims.
- The blind labelling gate failed twice.
The trap: every fix invalidated the accuracy history, since it belonged to the old version — constant fixing with no number to stand behind.
The fix: freeze the two verdict-deciding checks, with a test that fails if their code changes before being re-scored on 30 real panels at that exact version (currently 0/30, honestly). Shaky checks still run as info only. The author plans a pass using only the original ask and final artifact, and asks: how do you check your checker?
More from coding & agent
- The Resolution Hypothesis: agent evolution stalls because noisy evals can't spot real improvements — ShanghaiTech · 2026-11-19
- Microsoft ships Copilot SDK 2.0 with Rust core, powering all Copilot agents — PaulShellDev · 2026-11-19
- Rediscovering Notion as a Full Context Front End for AI Workflows — ivanhzhao · 2026-11-19
- WebMCP polyfill lands on npm: let web pages expose tools to browser agents — philnash · 2026-11-19
- Ex-DeepSeek core member open-sources Linger, a local AI agent claiming 5x faster inference — Aiden_Tech_Ai · 2026-11-19
- Meta's Pyrefly 1.5 hits 100% typing spec conformance, checks 40% faster — tdhopper · 2026-11-19