Overview
Plimsoll runs untrusted code, including programs written by AI agents, inside a sandbox. Every result identifies the isolation tier used for that execution, making the strength of the boundary an explicit property of the run rather than a general platform claim.
The four tiers, from weakest to strongest, are process, container, kernel, and VM. A request can require a minimum tier, and execution is refused when the available sandbox is weaker.
Key Features
- Reports the actual isolation tier with every execution result.
- Lets callers set a minimum acceptable isolation strength before code runs.
- Provides an official Go client that verifies the returned tier.
- Returns SHA-256 hashes of both the submitted request and returned result for client-side verification.
- Supports signed checked results, complete call-chain validation, and run replay through a separate attestation program.
Agent Code Evaluation
Plimsoll can evaluate agent-authored programs as well as contain them. Its trial runner tests a program against scenarios the program cannot inspect in advance; each scenario is delivered to the runner and removed before the submitted program starts.
The caller defines the simulated system, scenarios, and passing threshold, while Plimsoll supplies the execution boundary and evidence. A documented cart-pole example uses a WebAssembly simulator and records each trajectory into a SHA-256 fingerprint, allowing repeated runs to be compared for identical numerical results.
Verification Model
The official clients recompute request and result hashes instead of accepting them unchecked. The external attestation component signs verified outcomes, rejects a sequence of calls if one result is missing, and can replay a run, providing evidence for both isolation and execution history.