AI/ ai · claude · software-development · research

New Dataset Tracks Errors in a 21,000-Line AI-Only Codebase

A new arXiv study of a 21,000-line, all-Claude codebase found real errors in 14.3% of code-gen events and factual slips in about one in four chat replies.

An entire codebase written by AI, with every commit and error logged, just became a research subject.

The paper, "Between the Commits: Process, Error, and Claim Reliability in a Wholly AI-Authored Codebase" (arXiv:2609.29744, posted September 25, 2026), tracks the full development history of a 21,000-line Python tool built entirely by Claude, with no human-written code or tests anywhere in it. The authors built two code-provenance tracing tools and three taxonomies covering instruction intent, commit provenance, and response reliability, then applied all of it to the dataset. They found that instructions given to coding-agent CLIs skew toward comprehension, planning, and consultation rather than direct commands, and that development was mostly proactive rather than reactive. The headline numbers: 14.3% of code-generation events contained a real error later caught by the AI's own test suite, and roughly one in four to one in five of the AI's interactive responses contained a factual error.

This matters because most claims about AI coding reliability are anecdotal. Here is an actual error rate, measured against a real, working 21,000-line tool rather than a benchmark demo. It also quantifies a gap that matters for anyone letting an agent run unsupervised: the code errors get caught because the AI wrote its own tests, but the one-in-four factual error rate in conversational responses has no equivalent safety net.

Worth noting: the AI graded its own homework. The test suite that caught those code errors was also AI-authored, so the 14.3% figure measures what the system was built to catch, not necessarily every bug that's actually in there.

TR

The Revision

Written by an AI system from the public sources credited above. How we write →