shiyao-huang/awesome-agent-evolution/projects

PinchBench Skill Model Card

One Sentence

PinchBench is one of the clearest benchmark anchors in this corpus: not a self-evolving runtime, but the evaluator substrate that makes self-improvement claims comparable.

Three Sentences

It spans productivity, research, writing, coding, analysis, email, memory, and skill-discovery tasks rather than a single narrow benchmark lane. That makes it important for this survey’s “benchmark as selection pressure” thread. The 2026-06-17 morning packet nudged the public star count again, so the benchmark anchor also needed a same-day metadata refresh.

Model Card

Teaching Use

Use PinchBench to explain why benchmark infrastructure matters as much as the agent loop itself. Without evaluator strength and saved transcripts, improvement stays anecdotal.

Evidence And Limits

The raw capture now reflects a GitHub metadata packet observed on 2026-07-05: 1,261 stars, 144 forks, 383 commits, 21 open issues, and 0 open pull requests. This packet is fresher than the previous authenticated packet at 2026-07-05 01:38 +0800 where a delta was observed. This run did not execute the repository locally, validate workflows end to end, or independently rerun benchmark claims. Product, memory, benchmark, and automation claims therefore remain repository-scoped unless separately tested.