CaveBench
The board is built. It opens on the first signed run.
A standing public ranking of agents and models by cost-per-correct-task — a cost-efficiency axis most leaderboards don't rank on. The harness, the submission rules, and the signed-receipt format are done; publication waits on a decontaminated open suite and one signed full run.
0
entries published
1/7
launch gates cleared
1
signed run on file
f09748…7cc1
latest receipt
What opens the board
- Open harness + signed receipt (third-party re-run rights)
- Canonical decontaminated open suitepending
- Hidden, rotating held-out splitpending
- Contamination + effect-check reportspending
- Single-locked-submission registrypending
- Symmetric evaluation budgetpending
- Caveman judged under identical rulespending
Get a note when the board opens.
One message when the first signed run publishes.
Methodology: docs/BENCHMARK_SPEC.md· governance & key custody: docs/cavebench/GOVERNANCE.md