Empirical Truthfulness
Empirical Benchmarks & Verification Methodology
We do not publish unverified claims or theoretical maximums. Below are the actual measurements collected by our automated Live Growth QA Gate and local ground-truth evaluation suites.
Test Protocol
Evaluation Methodology
Our evaluation rigorously distinguishes between Local Engine Ground-Truth Verification and Live Production Network Probes:
- Local Ground-Truth Suite: Evaluates OCR field extraction against human-verified receipt receipts, executes regex across edge-case sample matrices, and runs SQL against isolated WebAssembly SQLite databases.
- Live Production Network Probes: Executes HTTP health checks, unauthenticated 401 rejection checks, and authenticated Server-Sent Event (SSE) query stream completions against production workers and containers.
Live Metrics
Production QA Gate Results (September 2026)
| Tool | Host Architecture | Test Scenarios | Status | Live Health Probe | Warm Query SSE |
|---|---|---|---|---|---|
| Apex Forge OCR | Render Web Service (Fastify + Tesseract WASM) | 10 Ground-Truth Cases (Thermal, Skew, Blur, Table) | 100% PASS | 228ms | 106ms |
| Apex Forge Regex | Cloudflare Worker (V8 Isolate + safe-regex) | 7 Test Matrices (Email, Phone, ReDoS, Unicode) | 100% PASS | 302ms | 103ms |
| Apex Forge SQL | Cloudflare Worker (WASM sql.js + Retry Loop) | 8 Relational Cases (JOINs, Aggregates, DDL Safety) | 100% PASS | 308ms | 139ms |
Latency & Infrastructure Context
Cloudflare Workers operate globally with near-zero cold-starts. Render free-tier containers may spin down after periods of inactivity, resulting in an initial cold-start delay (~45 seconds) on the first OCR request. Warm requests consistently execute in under 500ms.