Livenerf Tracks Opus 5.5 Drift

Livenerf is a thoughtful attempt to replace recurring “Claude got dumber” anecdotes with a reproducible time series, though its value currently lies more in the methodology than in results.

• –Frozen and procedural task panels help distinguish genuine drift from prompt memorization or changing workloads.
• –Pure-function graders avoid introducing instability from an LLM judge.
• –Pinned Claude Code versions and hermetic runs reduce the risk of confusing harness changes with model degradation.
• –The benchmark measures the Claude Code subscription experience, not necessarily Anthropic’s raw API model.
• –Baseline collection and calibration are still underway, so anecdotal dips remain inconclusive. [GitHub repository](https://github.com/ninjahawk/livenerf), [community update](https://www.reddit.com/r/ClaudeAI/comments/1wtlrnu/is_opus_55_entering_a_nerfed_phase_livenerf/)