We ran 6.2 billion COBOL validation passes. Zero errors. Here's what we learned.

COBOL is not dead. It's everywhere.
95% of ATM transactions worldwide run on COBOL. 80% of in-person point-of-sale transactions. An estimated 3 billion lines of COBOL are actively running in banking systems, insurance companies, and government infrastructure.
And yet — no modernization vendor has ever published a large-scale validation benchmark. Promises accumulate. Evidence remains absent.
We decided to change that.
The test
Environment: Hostinger KVM 8 VPS — 8 cores, 32 GB RAM, Ubuntu 24.04
Corpus: 9,595 real COBOL files — 4,490,720 lines
Method: 1,380 complete validation passes, 8 parallel workers
Total duration: 12.7 hours continuous
No synthetic data. No fabricated corpus. Real COBOL files — the raw material of industry.
The results

Total validations: 6,197,193,600
Errors: 0
Success rate: 100.000%
Stable speed (0–5h): 293,000 lines/second
Peak speed: 329,411 lines/second
Average speed: 283,881 lines/second
Memory leak: None
Crash: None

Milestones: 1B at 0.9h — 2B at 1.9h — 3B at 2.8h — 4B at 3.8h — 5B at 4.9h — 6B at 7.9h
The speed curve — and what it reveals
The parser held stable at ~293K lines/second for the first 5 hours. Then throughput declined progressively.
This is not a parser failure. It is the VPS being throttled by Hostinger after 5 hours of sustained 100% CPU load.
The parser did not fail. The infrastructure was externally limited.
This is the floor of an entry-level cloud VPS — not the ceiling of KIVUMIA.CODE.
What this means
Six billion validations. Zero errors. On a standard VPS.
Next step: run on local Ryzen hardware, no throttle, targeting 125 billion validations.
About KIVUMIA
Multi-agent AI platform dedicated to COBOL modernization — semantic migration to Python, large-scale validation, European digital sovereignty.
We don't conquer. We pollinate. 🐝
🌐 kivumia.com | kivumia.ai