benchmarks 5
- In my rerun, Headroom saved 0 tokens on its GSM8K eval prompts
- GameDevBench's 68.8% top score is the only strictly confined row
- SoL-Pi's Capability Floor Is Per Mechanism, Not Per Harness
- Harness Engineering: What a Benchmark Score Actually Measures
- GLM-4.7’s scorecard needs its test protocol beside it