# Evals and testing for LLM apps

> Measuring and regression-testing LLM apps and agents: eval datasets, benchmarks you run yourself, CI gates, tests on tool calls and traces.
> 3 recommendations, newest first. Canonical: https://wpnews.pro/playbooks/evals-testing

- When continuing harness optimization across a new benchmark or policy model, log each edit's rationale, predicted and observed outcomes, and eval context so later attempts can reuse or revise proven changes.  
  — [Self-Improvement Memory Enables Continual Recursive Self-Improvement](https://wpnews.pro/news/self-improvement-memory-enables-continual-recursive-self-improvement), source: <https://arxiv.org/abs/2611.18856>, 2026-11-17
- When prompting coding models for API calls, Dockerfiles or CI workflows, include a dated list of known upstream breaks with the fix for each; the eval cut broken-route picks from 37% to 16% for claude-sonnet-5.5, while a bare date did little.  
  — [Dated-break eval: do coding models write code that stops working on a known date?](https://wpnews.pro/news/dated-break-eval-do-coding-models-write-code-that-stops-working-on-a-known-date), source: <https://discuss.huggingface.co/t/dated-break-eval-do-coding-models-write-code-that-stops-working-on-a-known-date/191992#post_1>, 2026-11-17
- Check independent leaderboards like the Vals Index alongside vendor benchmarks before adopting a flagship model; GPT 6.2 Astra placed 7th there despite its launch.  
  — [GPT 6.2 Astra Places 7th On Vals Index, Behind Even Gemini 4 Argon & Gemini 4 Hydrogen](https://wpnews.pro/news/gpt-6-2-astra-places-7th-on-vals-index-behind-even-gemini-4-argon-gemini-4), source: <https://officechai.com/ai/gpt-6-2-astra-vals-index/>, 2026-11-17
