Benchmarks

Benchmarks

How fast can the gateway evaluate whether an agent should be allowed to act? The numbers below come from the APS benchmark suite, measured at four layers from the raw verifier core to a full gateway round-trip, on three machines, with both median and tail latency.

<2msp50 Policy Eval
403Ops/sec
<1msDeny Latency
15Constraint Dims

Runtime performance

APS is measured at four layers, from the raw verifier core to a full gateway authorization round-trip, on three environments. Bare-metal Linux (AMD EPYC 7313P) is the canonical reference; Apple M3 and AWS c7i (Intel Xeon 8488C) are shown for range. Core layers run single-threaded over 1,000,000 iterations, the gateway round-trip over 1,000. Both p50 (median) and p99 (tail) are reported.

LayerWhat it measuresMac M3AWS c7iEPYC (canonical)p99 (EPYC)
L0 · verifier coresignature and delegation-chain check, hot cache292ns347ns420ns470ns
L2 · SDK callthe same check as your code calls it, through N-API3.9µs5.7µs6.5µs8.1µs
L3 · durablethe call plus a persisted audit record (queued commit)4.0µs6.7µs7.6µs13.2µs
L4 · gateway round-tripfull authorization across the enforcement boundary305µs1.07ms822µs6.0ms

The fast-reject deny path matches L0 within noise. A synchronous fsync per action, instead of queued commit, is disk-bound: about 1.3ms on server SSD. Durability and round-trip tails are environment-dependent.

Reproduce it: per-environment results and env_capture.json live in benchmarks/prototype-1/results/ in the SDK repo. The TS SDK via napi-rs adds approximately 13-16x latency over pure Rust on all three platforms tested. Receipt minting happens inside the measured L4 round-trip.

Full enforcement stack

The L4 figure above is the lean round-trip. The product gateway does more on every call: cascade revocation, escalation checks, charter governance, settlement, drift detection. Measured against that full stack:

OperationLatencyNotes
Policy evaluation p50<2msfull constraint check across all applicable facets
Policy evaluation p99<10mscold cache, worst case
Denial<1msfail-fast on the first failing facet

Full-stack figures from the gateway benchmark suite, 500 iterations per scenario.

Constraint dimensions

Every action is evaluated against all applicable dimensions. The gateway short-circuits on the first failure, which is why denial is sub-millisecond.

Time

Approval TTL and delegation expiry

Spend

Per-action and cumulative limits, delegation budget

Scope

Tool name, parameter patterns, target restrictions

Reputation

Tier threshold, minimum evidence count

Values

Values floor F-001 through F-008, graduated enforcement

Revocation

Delegation revoked, cascade revocation

Reversibility

Tentative, compensable, or irreversible classification

Governance

Governance artifact version, stale attestation

Identity

Passport verification, signature validity

Replay

Request ID reuse detection

Cross-chain

Taint tracking, confused-deputy prevention

Audience

Proof bound to a named recipient

Escalation

Bounded escalation grant validity

Fidelity

Substrate behavioral fidelity

Data

Data source access terms, contribution tracking

Architecture

The gateway is both judge and executor. The agent requests an action. The gateway evaluates all constraints, executes the action if permitted, and generates the receipt. Agents never touch tools directly. Enforcement is applied at the gateway rather than left to the agent.

ComponentCountNotes
Protocol modules10784 core + 23 v2 constitutional
MCP tools150Full protocol surface via MCP
Tests4,326Including adversarial + invariant suites
Research papers8Published on Zenodo

How it compares

APS is the governance layer under orchestration frameworks, not a replacement for them. The dedicated comparison lives on the compare page.

Test Coverage

4,326 tests across 107 modules. The test suite includes unit tests, integration tests, adversarial scenarios (replay attacks, confused deputies, scope escalation), invariant checks (monotonic narrowing, cascade completeness), and stress tests (concurrent gateway load).

Among the suite are the cross-implementation conformance fixtures: canonical-JSON vectors that any verifier, in any language, can run to confirm byte-parity against the spec. Those are the tests that make receipts portable.

Independently cited by PDR in Production (University of British Columbia), which references the APS earned-reputation model. One external citation. Zenodo

For AI agents: Visit llms.txt for machine-readable documentation, llms-full.txt for the full technical reference, or .well-known/mcp.json for MCP server discovery.