Performance and test results
See how long APS authority checks take, what each measurement includes, and how APS behaves on tests written by other projects.
How long each check takes
AMD EPYC 7313P, one threadChecking an action against authority already loaded into memory is faster than receiving a request over HTTP or waiting for an audit record to reach disk. The measurements below show those costs separately.
These are recorded results from the Runtime Passport prototype on an AMD EPYC 7313P running Ubuntu 22.04. Each benchmark runs sequentially on one thread. They describe the measured prototype revision, not the current TypeScript SDK or a production deployment.
µs = microseconds. 1 ms = 1,000 µs. 1 sec = 1,000,000 µs.
-
Rust authority check, with compiled authority already in memory0.42 µs
Median 0.42 µs, p99 0.47 µs, 1,000,000 timed calls.
-
Check called from TypeScript through the native binding6.54 µs
Median 6.54 µs, p99 8.11 µs, 100,000 timed calls.
-
Native binding check with the audit record queued for writing7.64 µs
Median 7.64 µs, p99 13.2 µs, 50,000 timed calls.
-
Native binding check that waits for the audit commit1.31 ms
Median 1.31 ms, p99 1.5 ms, 5,000 timed calls.
-
HTTP authorization request to a gateway on the same machine822 µs
Median 822 µs, p99 6.01 ms, 1,000 timed calls.
The Rust measurement starts with prepared authority and action inputs. It does not include loading a passport, verifying its signature or fetching revocation information. The queued record measurement returns before the background writer completes the disk commit. The blocking measurement waits for that commit.
The HTTP measurement includes the local request and response, a recorded policy evaluation and creation of an evaluation receipt. It excludes a remote network trip and the business operation being authorized.
Read the result files and environment record. Each file identifies the measured code revision. Median is the middle recorded latency. p99 is the latency at the 99th percentile.
Measurements on other machines
Median| Operation | Apple M3 | AWS c7i, Xeon 8488C | AMD EPYC 7313P |
|---|---|---|---|
| Rust authority check | 0.292 µs | 0.347 µs | 0.420 µs |
| Check through the native binding | 3.92 µs | 5.71 µs | 6.54 µs |
| Check with audit commit completed before return | 4.04 ms | 1.35 ms | 1.31 ms |
| HTTP authorization on the same machine | 0.305 ms | 1.07 ms | 0.822 ms |
These are separate recorded runs, with different code revisions in their metadata. Use them to inspect each setup, not to isolate the effect of a CPU or operating system. Compare the recorded environments.
Tests from other projects
Their tests, run by the APS projectThese reports show where APS agrees with another project's tests and where an adapter or a separate control is needed. The reported runs were performed by the APS project. External test authorship is identified separately from who ran the tests.
- 5/5REMORA operational technology testbedFive cases passed the reported authority and identity binding checks
The cases include valid authority, widened scope, expired authority and a request outside the delegated scope. The report records both a host run and a container run in APS CI. It checks delegation decisions and signed bindings, not whether a machine safely executes the command.
Read the run report ↗ - 5 + 3Authority Relations v0.1Five cases passed. Three were reported as unsupported.
The development run used an adapter with APS SDK 8.0.1. Its passes depend on translating APS reason codes, and two cases are decided by the adapter. The expected answers were available during development, so this was not a blind evaluation.
Read the results and adapter changes ↗ - scopeSINT runtime safety fixturesThe scope denial decision matched. Physical safety remains a separate check.
The run mapped SINT resources and actions into APS scope grants. It rejected the out of scope request. It did not evaluate velocity or force limits, escalation, transport enforcement or emergency stop behavior. The emergency stop case received a scope permit in this mapping. That result does not authorize physical actuation. The safety system still needs to block the command when the emergency stop applies.
Read the six case report ↗ - 3/3Refund sandbox receipt adapterThe adapter and direct SDK verification agreed on three checked inputs
Toshi's adapter was checked against a valid receipt, an altered signature and changed decision context. Its self test also reported 70 passing checks. This tests receipt inputs. It does not establish that the refund executor enforces the decision.
Read the adapter check ↗
Inspect and reproduce
The performance records identify the code revisions, hardware, workloads and sampling methods behind the measurements. The Rust runner measures the local check. Separate TypeScript runners measure the native binding, audit modes and HTTP request.
For additional fixtures and implementation reports, see the Agent Authority Conformance lab. Read each record's scope and provenance before treating it as independent evidence.
These results describe specific implementations and test conditions. They do not establish production capacity or an independent security audit. An authorization receipt records a decision. Evidence from the protected resource is needed to establish whether the authorized operation completed.
For AI agents: llms-full.txt · MCP descriptor