GovBench: governing what agents actually do
Summary
A benchmark for what an execution layer is for: refusing what it must, allowing what it should, and reporting truthfully what happened. Its first run found two real bugs.
Key results
| Measurement | Result | Source |
|---|---|---|
| Safety cases passed (362-case corpus) | 312/312 | govbench-0.3.0-alpha.12-2026-10-14.json |
| Utility: routine operations allowed and verified | 50/50 | govbench-0.3.0-alpha.12-2026-10-14.json |
| Containment | 197/197 | govbench-0.3.0-alpha.12-2026-10-14.json |
| Approval integrity | 19/19 | govbench-0.3.0-alpha.12-2026-10-14.json |
| Verification honesty | 30/30 | govbench-0.3.0-alpha.12-2026-10-14.json |
| Exactly-once | 11/11 | govbench-0.3.0-alpha.12-2026-10-14.json |
| Redaction | 55/55 | govbench-0.3.0-alpha.12-2026-10-14.json |
| First run, original 357 cases, before the fixes | 293/307 | GOVBENCH.md |
| Original 357 cases, after the fixes | 307/307 | govbench-0.3.0-alpha.11-2026-10-14.json |
| Most added by governance to one action, median of 40 runs (approved terminal run) | +0.85 ms | GOVBENCH.md |
GovBench measures what MELRA is for: governing effects. Each case is a request together with the outcome MELRA must produce. Every case runs through the real runtime (plan, approve, execute, verify) with the real file, terminal, HTTP, memory and browser adapters. No model is called. The only network is a loopback server that the harness starts itself.
The headline run used the 362-case corpus against MELRA 0.3.0-alpha.12, with Chrome present and no cases skipped. Every safety case and every utility case passed. Utility is the other side of the ledger: routine, harmless work that must be allowed and must end verified. It is the false-positive rate. The corpus has since gained one more approval case, which these scores predate.
What each category asks
- Containment. The request is refused at the schema, at policy (when planned or when executed) or at the adapter boundary, with the documented reason, and no effect is observed: the loopback canary is not hit, the directory outside the workspace is unchanged byte for byte. Anything that asks for approval is approved, so a pass means the boundary held after a human said yes.
- Approval integrity. Missing, wrong, padded, expired, replayed, cancelled and cross-task approvals are refused, and the effect never happens. A plan made under a looser policy is refused at execution after a restart under a tighter one.
- Verification honesty. The final status is exactly the expected one. An adapter that succeeds while the evidence fails is
partial. A mutation that was sent but never answered isrecovery_required, not failed. - Exactly-once. The side effect is counted at its source. The same idempotency key twice, concurrent executes, re-execution and a restart all produce one effect.
- Redaction. Seven fake secret shapes go into seven places a secret could end up. After the runtime closes, every file in its data directory is scanned byte for byte for the secret and its fragments.
What the first run found
The first run, on the original 357-case corpus and before any fix, passed 293/307 safety cases. The 14 failures were not expected failures. They were two real bugs.
- Bracketed IPv6 addresses skipped the private-address check. A URL parser keeps the brackets in a host such as
[::1], so the check did not see an IP address and treated the host as a domain name. Each request still failed, but only because the system resolver cannot look up a bracketed string. The documented boundary never ran. This covered twelve cases, including loopback, private, link-local and IPv4-mapped addresses, over HTTP and in the browser. Fixed in #119: the brackets are stripped before the address is classified. - Memory kept two kinds of secret. Memory redaction works by pattern, and it had no pattern for PEM private-key blocks or AWS access key IDs, so a stored value containing one was written verbatim to the database. The same secrets were absent from every other place tested. Fixed in #120: memory, terminal output and receipts now share one list of secret patterns that knows both shapes.
After the fixes, the same 357 cases passed in full, and that run is kept as its own result file. The headline run adds five cases for custom evidence checks.
What governance costs
For each operation the harness compares 40 governed runs (plan, approve, execute, verify, receipt) with the same operation through the bare adapter. On the recorded machine the median added cost ranged from +0.48 ms for an approved memory write to +0.85 ms for an approved terminal run. Overhead depends on machine load: an earlier run on a busy machine measured +1.1 to +4.4 ms. The correctness scores do not depend on the machine.
What it does not show
GovBench is self-authored: the project that wrote MELRA also wrote the cases. It is a regression and transparency instrument, not a third-party benchmark. Its scores cannot be compared with agent task-success benchmarks such as WebArena or OSWorld, or with other products. A high score says the listed cases hold. It does not say MELRA is secure against attacks outside them.
The known gaps are written down. DNS-based server-side request forgery and cloud-metadata hostnames are out of scope, because the benchmark makes no external lookups. Terminal containment checks the command and its working directory, not what an allowed program does with its own arguments. Redaction is tested for seven secret shapes, and other formats may not be recognised.
One rule keeps the benchmark honest: a failing case stays failing until MELRA is fixed. Its expectation is never changed to match what was observed.