AI is rewriting
your COBOL.
Refinery proves it’s safe.
The COBOL talent cliff is forcing banks onto AI tools, with no independent check in place. Refinery runs the modified code against the original, catches semantic drift before production, proves equivalence with Z3 formal verification, and explains it to your risk team. Every step aligned to SR 11-7, SS1/23, and DORA.
No setup fee · Independent of your vendor stack · No lock-in
Add ROUNDED to COMPUTE WS-PAYMENT to satisfy Basel III precision requirements for loan amounts exceeding £1M, prevents silent truncation at 8+ decimal places.
Run. Catch. Prove. Explain.
Four verbs, one audit trail. Not a platform you have to picture. A sequence you can watch run.
Run
Compile the AI-modified COBOL and the original, then run both against boundary-value inputs on a real z/OS emulator.
Catch
Compare outputs byte by byte. Catch the COMPUTE drift, the truncated field, the removed error handler, before production does.
Prove
Generate a signed, SHA-256 fingerprinted Change Contract mapped to SR 11-7, SS1/23, and DORA. Audit-ready evidence on paper.
Explain
A compliance-tuned AI agent knows every check, every formal proof, and every blast-radius impact. Risk asks it — not your engineers.
Live Demo
Watch Refinery catch an AI-generated COBOL defect
Real IBM Bob output. Real audit engine. Real SHA-256 fingerprinted PDF.
Five layers.
One audit trail.
Semantic Analysis
Beyond diff.
Compiler-inspired analysis.
Builds a COBOL symbol table for the supported subset, resolves field references, traces data flow signals, and fires 11 named rules including Z3 formal verification. Catches changes a diff engine would miss.
Analysis output
Estate Blast Radius
One file.
Fourteen systems at risk.
Traces every COBOL CALL, COPY, and JCL EXEC PGM to map the full estate impact before a change ships.
Change Contract
A structured PDF.
SHA-256 fingerprint.
Every audit produces a structured PDF with a cryptographic fingerprint, a plain-English risk summary, and a sign-off block for the Chief Risk Officer. Permanent and regulator-ready.
Sample report
Formal Verification
Z3 SMT solver.
Proved for all inputs.
For every COMPUTE, ADD, and MULTIPLY, Refinery encodes both program versions as Z3 integer formulas and asks the solver: does any input produce different outputs? UNSAT means mathematically proved equivalent — not just for test cases, for every possible input.
Z3 proof output
AI Explanation Agent
Risk asks Refinery.
Not your engineers.
A compliance-tuned AI agent knows every check, every formal proof, and every blast-radius impact. It briefs your CRO in plain English — on demand.
Runs on your hardware.
Gets smarter every audit.
A local-first AI agent that lives inside your infrastructure. No audit data sent to any cloud API. Every run makes it sharper.
Local inference via Ollama
The compliance agent runs on an Ollama-served model inside your own infrastructure. No query, no audit record, and no COBOL source ever touches an external API.
Self-improving audit memory
Every audit Refinery runs is indexed into an on-device retrieval store. The agent searches past audits to find similar programs, prior verdicts, and known risk patterns — automatically getting smarter the more you use it.
Plain English for your risk team
Risk officers ask questions in plain English. The agent answers with the full audit context — formal proof status, blast radius, regulatory mapping — without involving an engineer.
Agent runtime
Not a plugin.
An independent layer.
| Capability | Refinery |
|---|---|
| Semantic equivalence check | ✓ 11 named rules |
| Z3 formal proof of equivalence | ✓ PROVED_EQUIVALENT for all inputs |
| Estate blast radius mapping | ✓ JCL + VSAM + CALL |
| PDF audit certificate | ✓ SHA-256 fingerprint |
| DORA / SR 11-7 / SS1/23 | ✓ all three |
| Vendor-independent | ✓ any AI tool |
| CRO governance portal | ✓ built in |
Built for the humans
and agents that come next.
The first layer to sit between AI and production COBOL
Banks are facing a COBOL talent cliff. As the engineers who wrote these systems retire, AI tools are being deployed to fill the gap. Refinery is the independent check that makes that transition safe — and provable to regulators.
Mathematical proof, not just test coverage
Refinery's Z3 SMT engine formally proves that modified COBOL produces identical outputs for every possible input — not just the inputs in your test suite. UNSAT means proved. SAT means there's a counterexample, and we show you the witness values.
Your agents get better at optimising your codebase over time
Each run teaches the system what good looks like in your estate. The longer you run Refinery, the smarter your optimisation loop becomes.
Three panels.
Every change, governed.
Live diff, 11-check semantic feed, Z3 formal proof, and estate blast radius: one audit workspace your CRO can actually read.
The question isn’t whether AI will touch our COBOL. It’s whether we can prove to the regulator that what came out is the same as what went in.
Chief Risk Officer
Tier 1 UK Bank
Recognised by
See Refinery run
on your code.
Refinery is priced as a share of the compliance risk we remove. No fixed licence, no upfront cost. You pay when you’re verified.
Book a call →Previously at Palantir, SAP, Goldman Sachs, and Google · Palantir Winter Fellow · Goldman Sachs Spring Intern · SAP SWE