Part V · 5 — Benchmarks and leaderboards

draft

"Without good evals, you navigate in the dark." How to measure capability for real — and why private benchmarks are your only trustworthy compass.

🎨 Figure F-V.5The evaluation dashboard. Brief: a stylized "dashboard" with gauges of several benchmarks and a red "contamination" alert over the public ones; alongside, a vault labeled "private benchmark". Compendium palette.


The evaluation dashboard

5.1 Reference benchmarks

Benchmark Focus Note
HumanEval / HumanEval+ Python functions classic, saturated
MBPP / MBPP+ basic Python
SWE-bench / Verified / Multimodal real GitHub issues the most important today
LiveCodeBench recent contests avoids contamination
BigCodeBench realistic library calls
RepoBench / CrossCodeEval multi-file
GPQA Diamond PhD-level science
Humanity's Last Exam PhD-level, multidisciplinary the frontier already matches the average human

Current leaderboards (snapshot) live in Part III, doc 08-state-of-the-art-2026.kmd (SWE-bench Verified, Arena Elo).


5.2 The contamination trap

Critical tip: the public benchmarks are partially or fully contaminated in the training of frontier models. A high score may be memorization, not capability.

Build your own private benchmarks with data from your target domain:

  • they represent the real problem you want to solve;
  • no public model was trained on them → they measure genuine capability;
  • they are your evaluation advantage (and, ultimately, your product advantage).

5.3 How to evaluate well

  • Public + private combined; never public-only.
  • Red teaming and uplift studies (does the model really help?).
  • For reasoning/code: verifiable reward (did it pass the tests?) is the

    gold standard — and becomes an RLVR signal (see 01-frontier-levers.kmd).

Evaluation is not a final stage — it is a continuous compass. It connects to stage 5 (Evaluation) and 5.5 (Homologation) of the lifecycle (Part I).