---
title: "Benchmark Evidence and Reproducible Results | Korve"
url: "https://korve.dev/evidence.md"
canonical_url: "https://korve.dev/evidence"
description: "Read Korve's benchmark method, exact local regression results, limitations, and reproduction command without confusing lab measurements for production promises."
type: "resource"
korve_skills: "curl -fsSL https://korve.dev/skills.sh | sh"
---

Public evidence

# The benchmark, the method, and the limits.

This page publishes Korve's verified app-auth and database-explorer regression baseline exactly as captured. It is a local, warm-path engineering measurement—not a production latency promise and not a performance comparison.

[See exact results](#results) [Markdown evidence](https://korve.dev/evidence.md)

Captured July 21, 2026 · Public page updated July 29, 2026.

Configuration

## A bounded local regression baseline

The harness creates throwaway databases, seeds deterministic rows, warms each path, records serialized samples, and removes the databases when the run finishes.

| Input | Captured value |
| --- | --- |
| Captured | July 21, 2026 |
| Host | Apple M5 Pro, arm64, macOS 26.4.1 |
| Runtime | Bun 1.3.13 |
| Database | PostgreSQL 17.10 in a local container |
| Dataset | 10,000 event rows, one compound index, and one view |
| Network | Loopback with warm connections and caches |

Reproduce the harness

```
TEST_DATABASE_URL=<local-test-postgres-url> bun run benchmark:auth-data
```

The harness rejects non-loopback database hosts before creating or dropping anything.

Exact captured results

## Percentiles, sample counts, and maximums

Values below are copied from the checked benchmark artifact. They have not been rounded again, extrapolated, or converted into an uptime or latency guarantee.

| Path | Runs | Mean | p50 | p95 | Max |
| --- | --- | --- | --- | --- | --- |
| App-session API verification | 200 | 3.29 ms | 2.68 ms | 5.11 ms | 19.91 ms |
| Opaque-token SHA-256 digest | 2,000 | <0.01 ms | <0.01 ms | <0.01 ms | 0.13 ms |
| OAuth state + nonce + PKCE material | 500 | 0.02 ms | 0.02 ms | 0.03 ms | 1.08 ms |
| Schema, columns, keys, and indexes | 25 | 6.74 ms | 5.56 ms | 8.94 ms | 23.15 ms |
| Filtered and sorted 50-row browse | 50 | 7.58 ms | 6.90 ms | 10.82 ms | 26.63 ms |
| Read-only grouped SQL query | 100 | 3.06 ms | 2.84 ms | 4.49 ms | 11.11 ms |

Across 175 bounded database operations, the observed process-heap change was 5,338,895 bytes (5.09 MiB). The sample recorded no timeouts, cancellations, rejected statements, failed sessions, or leaked result contents.

## What this supports

- The benchmark harness exercises real app-session verification and the bounded database explorer.
- Every sampled explorer result stayed under configured row and byte ceilings.
- The measured paths are suitable as a repeatable local regression baseline.

## What this does not support

- It does not measure regional production network latency or a customer workload.
- It does not establish a production service-level objective or compare Korve with another vendor.
- It does not measure database provisioning, branching, queue dispatch, cron skew, backup, or restore time.

## Read who stands behind the claims.

Korve's author page documents the editorial, disclosure, and correction standard used for technical articles and comparisons.

[About the author](https://korve.dev/about)