SANDBOX OPERATIONS CENTER EXPERIMENT R4
Let the agent work.
Keep the blast radius zero.
The model writes real code, the OS contains it, a human releases it — and every run lands in a hash-chained ledger.
02 / CODING AGENT
READYModel writes the fix
Choose a real bug. The generated patch runs only in the boundary; release always waits for a human.
01MODELWrites a patch
02GATEFixed tests, no host
03SANDBOXbwrap + seccomp
04HUMANReviews the diff
A real task loaded from the fixed catalogue.
AGENT LOOP · ATTEMPTSNOT RUN
Waiting for the coding agent.
03 / PATCH AWAITING HUMAN REVIEW
04 / COMPUTED CHECKS
No patch executed yet.
The planner may only choose one allowlisted fixture. It never writes executable code.
06 / CURRENT ATTACK
READYSelect a controlled incident
No host processes will be modified.
ISOLATED PROCESS OUTPUTNOT RUN
Waiting for a sandbox run.
One short-lived process per run; no persistent sandbox state.
07 / SAFE REPAIR PREVIEW
08 / CHECKS COMPUTED FROM OBSERVED BYTES
No attack executed yet.
09 / INDEPENDENT OBSERVABLES
NETWORKUNSHARED
SYSCALLS…
RESOURCESCPU · MEM · PIDS
VULTR VMNOT CONNECTED