Stress-tested, not just demoed
Our public demo (CardDemo) is 44 programs — a clean, well-behaved AWS reference app. Real mid-tier bank and insurer estates aren't clean. They're decades old, thousands of programs, full of copy-pasted code nobody's touched since the author retired. So we went and found one of those to test against — and used it honestly, warts and all.
The corpus
A real Brazilian business/financial system (checks, payments, accounting) built 1998–2004, later archived as open source under GPL-3.0 by its author. No relation to MainframeLore or any customer — we picked it specifically because it's public, properly licensed, and nothing like our own tidy demo corpus.
The run
| Stage | Time | Result |
|---|---|---|
| Ingest & classify | 4m 21s | 12,499 files classified — including ~6,500 compiled binaries/listings mixed into the export, correctly filtered as noise, zero crashes |
| Doc + flowchart generation | ~20s | 2,131 program pages, structural facts only (no AI calls) |
| Q&A indexing | 5.8s | 65,569 searchable chunks with citations |
Ran on a laptop. No mainframe access, no upload to us required to reproduce — same pipeline that runs on every estate we're handed.
The dependency map
kello has no JCL — it's a Windows/Micro Focus shop, not classic mainframe batch — so there's no
job-level estate map like CardDemo's. What it has instead is copybook fan-out most estates only
discover the hard way: change one shared field definition and hundreds of programs are affected.
This is the real, unfiltered dependency graph — top copybooks ranked by how many programs COPY
them.
591 distinct copybooks referenced · 32,632 COPY relationships total · 402 copybooks shared by 3+ programs
What we found
Real legacy shops copy-paste a template file and never update the internal PROGRAM-ID.
In this estate, 39 completely different source files —
different logic, different purpose — all declared PROGRAM-ID. GALHOCOD050. Our doc
generator keyed each output page by that name. Every duplicate silently overwrote the last.
No error, no warning — just missing documentation for 70% of a real estate.
Pages are now keyed by the actual file, not the name it happens to claim. When two files
collide on the same PROGRAM-ID, both get documented (NAME,
NAME~2, ...) with a visible banner explaining the collision instead of hiding it.
Re-ran the same estate: 2,131 of 2,131 programs
documented.
A tool that claims to document "undocumented mainframes" and then silently drops 70% of a messy real estate isn't safe to sell. We'd rather find that on a public GPL repo before a client finds it on their own source. Parsed, not guessed — including about our own bugs.
Verify it yourself
Three unrelated files, same declared program name — open them on GitHub, no sign-in needed. Line 2
of each says PROGRAM-ID. GALHOCOD050.
Copy either file's source from GitHub, paste it into our live parser — same engine that produced every number on this page, runs entirely in your browser, nothing uploaded — and watch it extract the structure itself.
The full 2,131-program estate is public and unmodified — github.com/paulohl/kello, GPL-3.0, archived by its own author. Count the files, read the code, form your own opinion.