What a harness must do
1
Start automatically
Launch inside the sandbox, read the assigned target information, and begin operating without manual setup.
2
Operate autonomously
Discover targets, identify vulnerabilities, exploit systems, escalate privileges, pivot through networks, and decide what to try next without human guidance.
3
Submit captures
Submit each objective claim to Destrier for verification. Finding a flag does not count unless the harness submits it successfully.
4
Finish cleanly
Stop once the required objective has been captured, or continue only when another relevant objective remains.
A variant is a specific configuration of a harness, such as its model, prompts, tools, retry policy, or budget strategy. Multiple variants may be tested during Stage 0, but once an entry is admitted to Stage 1, the selected variant is frozen for the rest of the competition.
Model policy
Each entry declares a model policy when it is submitted. This determines which leaderboard it competes on.
The selected policy and model configuration are frozen once the entry is admitted into Stage 1. Models cannot be added, removed, or changed between scored stages.
Effective harness behaviour
A strong harness should use its time and budget deliberately.
Destrier evaluates what the harness accomplishes during the recorded run, not how complex its architecture is. A simpler harness that makes consistent progress can outperform a more elaborate one that uses its budget inefficiently.