Performance Objectives
This page lists a small set of latency objectives for the hosted service, how each one is measured, and how far each has been checked.
What these objectives are, and are not
- They are design objectives. They describe what the service is built and operated to achieve. They are not a service-level agreement.
- They are not an availability commitment. Nothing here states an uptime, availability, high-availability, recovery, error-rate, throughput, or capacity commitment, and no credits or remedies attach to them.
- They do not replace enforced limits. The values in Limits and Configuration, including the decision timeouts that fail a call closed, are enforced behavior and are unchanged by this page.
- They are provisional until validated. The Status column says how far each objective has been checked. Targets may change, and a change is recorded on this page.
How latency is measured
- Server side. Time starts when the service receives a request and ends when it finishes sending the response. Your network, your client, and the upstream call you are governing are not included, except where an objective says otherwise.
- Percentiles. p50 is the median, p95 and p99 are the values that 95% and 99% of measured requests finish within. Hosted objectives are defined over a rolling 30-day window.
- Counted requests. Requests for a registered asset that succeed. Requests answered with a client error, including
429when a rate limit is exceeded, are not counted. Time spent waiting for an approval to be resolved is not counted. - Unregistered tools. A call for a tool that is not registered is allowed without a decision record and is not counted.
Decision objectives
| Objective | p50 | p95 | p99 | Status |
|---|---|---|---|---|
| Policy evaluation, computation only | 10 ms | 30 ms | 50 ms | Validated against measurements of the software |
Decision handling, server side (POST /v1/evaluate) | 30 ms | 60 ms | 120 ms | Pending hosted validation |
| End-to-end decision latency at the API edge | 100 ms | Pending hosted validation |
- Policy evaluation, computation only. The time to evaluate the published, active policies of one asset against a request, without storage or network. The objective applies to assets with up to 100 policies.
- Decision handling, server side. The time to handle one
POST /v1/evaluatecall for a registered asset: authentication, evaluation, recording the decision and its audit event, and the response. - End-to-end decision latency. The time seen by a client in the same region as the service, from sending the request to receiving the response. It includes network and TLS time. This objective applies to the hosted service only.
Read objectives
| Objective | p95 | Status |
|---|---|---|
Reading one page of up to 100 rows from GET /v1/audit or GET /v1/decision-explorer | 500 ms | Pending hosted validation |
Reading the recorded detail of one operation with up to 1,000 events, GET /v1/audit/operations/{operation_id} | 1 s | Pending hosted validation |
These objectives cover the API response only. They do not cover rendering in the Console.
Objectives to be set after measurement
No target is published yet for these. They will be added here when they have been measured on the hosted service.
- Start of asynchronous delivery. The time from an approval email or webhook event being created to the first delivery attempt. Later attempts follow the fixed retry schedule in Limits and Configuration. The objective will not cover whether the recipient accepts the delivery.
- Audit export delay. The time from a decision being recorded to its copy appearing in write-once storage. See Export to write-once storage for which events are exported.
Self-hosted deployments
The hosted objectives do not apply to a deployment you operate: latency there depends on your infrastructure, database placement, and load. To judge your own deployment, measure the same boundaries described above on your own traffic.
The table below is a reference measurement of the software, not a target and not a commitment for any environment. It was taken on 15 November 2026 on a single development host with the database on the same host, no TLS, no network, and one request at a time. Decision rows use 400 samples after warm-up, and computation rows use 20,000.
| Measurement | p50 | p95 | p99 |
|---|---|---|---|
| Policy evaluation, computation only, 10 policies | 0.06 ms | 0.15 ms | 0.34 ms |
| Policy evaluation, computation only, 100 policies | 0.6 ms | 1.3 ms | 1.7 ms |
POST /v1/evaluate, registered asset, no policies | 14.6 ms | 18.0 ms | 20.1 ms |
POST /v1/evaluate, registered asset, 10 policies | 19.7 ms | 23.9 ms | 27.2 ms |
Decision handling is dominated by recording the decision and its audit event, not by policy evaluation. A database that is not on the same host adds its own round-trip time to every decision.
Related pages
- Limits and Configuration lists enforced limits and timeouts.
- Policy Engine describes how decisions are evaluated.
- Audit describes what is recorded and how it is read and exported.