29% False Claims Rate (Code-Path Hallucination Metric)

Description

Internal quality metric discovered in the leaked source: 29% false-claims rate measured against an internal evaluation suite designed to surface a specific failure mode — Claude Code making assertions about what a piece of code does that turn out to be incorrect when the code is executed or tested. Not a general factual-accuracy metric and not a rate of intentional deception. Has no baseline comparison against other coding tools in the leaked source, so 'high' or 'low' cannot be inferred. The media framing ('Claude lies 29% of the time') is inaccurate; the engineering framing ('29% miss rate on code-path assertion tests in a hostile eval') is what the source actually shows.

Key claims

Relations

Sources

src-20260409-09a1b2325b23