Refineryline 32A verdict in minutes, in the pipeline.
Runs next to your build and tests, on every change the AI makes.
Illustrative example
Buildpassed
Unit testspassed
Refinery verdictCheckingPassed
Backed bybased.
collective
collective
Verification infrastructure
for consequential software.
Proving mission-critical code still behaves correctly after an AI tool rewrites it, with evidence a regulator can inspect.
Product preview: Refinery flags an AI change to INTEREST-CALC that moved the overdraft threshold from 5,000 to 9,000, and asks a person to accept it or return it to the developer.
Product preview. Example change, not real customer data.
AI changes are outrunning review.
An AI tool can rewrite a COBOL program in minutes, and checking what it actually changed is still one person reading it line by line.An AI tool can rewrite COBOL in minutes. One person still reviews it line by line.
Unverified change
- AI rewrote INTCALC.cbl
- "Looks fine to me"
- Threshold moved, unnoticed
- "Who approved this?"
How it worksFollow one change
Follow one change
across your whole estate.
Product preview, in eight steps: an AI tool edits INTEREST-CALC; Refinery maps the programs, copybooks, datasets and batch jobs linked to it; it maps everything the change reaches, four programs and two batch jobs, and checks the changed program; it finds that line 32 moved the overdraft threshold from 5,000 to 9,000, so a balance of 5,000.01 is charged 25.00 before and 0.00 after; Explain answers in plain words; the change manager accepts it as intended and signs; the certificate lands on the pull request and the change ticket; the next change to INTEREST-CALC arrives with that decision, who made it and why.
The change
Your AI tool changes the code.
Any tool, or any person, edits one program. It compiles, but other programs and nightly jobs depend on it, and nobody has checked what it now does.An AI tool edits one program. Other programs and nightly jobs depend on it. Nobody has checked its new behaviour.
The estate
Refinery already knows your estate.
It reads the programs, copybooks and batch jobs in your source, and how they connect. So it sees a change in context, the way your most experienced engineer would.Refinery maps your programs, copybooks and batch jobs. Each change arrives with the connections that matter.
The reach
It shows everything the change reaches.
Not just the edited line: the programs that call it, the copybooks and data it shares, and the jobs downstream. So you know what to retest before it ships.See the programs, shared data and downstream jobs a change reaches. Know what to retest before it ships.
The finding
It finds the one line that matters.
Line 32 moved a threshold from 5,000 to 9,000. Run both versions on a balance of 5,000.01 and the charge drops from 25.00 to nothing.The threshold moved from 5,000 to 9,000. At a balance of 5,000.01, the charge falls from 25.00 to zero.
The answer
And explains it to whoever decides.
Ask in plain words and get a plain answer, so the person deciding does not need an engineer beside them.Ask about the finding in plain words. The person deciding gets an answer without needing an engineer beside them.
The decision
A named person signs.
They accept the change as intended or send it back. Either way the decision is recorded under their name, and an accepted change gets a signed certificate.A named person accepts the change or sends it back. Their decision is recorded, and an accepted change gets a signed certificate.
The record
The evidence files itself.
The certificate and the evidence pack land on the pull request and the change ticket, ready for the change board and for audit.The certificate and evidence pack attach to the pull request and change ticket. The change board can inspect both.
The memory
And it gets better with every change.
Every accepted change is remembered, with who accepted it and why. The next change to that program arrives with what was decided before.Each accepted change records who decided and why. A later change to that program can refer to that decision.
Product preview. Example change, not real customer data.
Product preview. Example change, not real customer data.
What you getOne framework for AI changes
One framework for AI changes
to legacy systems.
Engineering, risk, compliance and the business work from the same evidence, on the same change.
One change · illustrative example
PAYROLL.cblPassed
PR #479CHG0041219
- Changed by
- GitHub Copilot
- Checked by
- Refinery
- Signed by
- Jordan Lee
Evidence on the change ticket. A clean pass goes through as a standard change.
The certificate and evidence pack sit on the ticket, so the board decides on facts.
Illustrative example
CHG0041219Standard change
Certificate attached
Evidence pack attached
One signed report over every AI change.
Every change, every check and every sign-off in one place, ready for audit.
Illustrative example
ChangePR #479
Signed byJordan Lee
Signature verified
AI changes shipped safely, and measured.
See how much AI work reaches production safely, and where people still step in.
Illustrative example
AI changes shipped safely
Versus today
Manual review, against Refinery.
The same release decision, without the wait or the guesswork.
| Aspect | Manual review today | With Refinery |
|---|---|---|
| Time | Days to weeks of specialist review | Minutes |
| Coverage | A sample gets a close look | Every change checked |
| Record | A tick in a ticket | A signed certificate anyone can verify |
| Assurance | Trust the reviewer | Proof and evidence a regulator can inspect |
| Your code | Sent out to review tools | Nothing leaves the bank |
Why RefineryEveryone checks their own homework.
Everyone checks their own homework.
Now someone else can.
The platforms rewriting your COBOL also supply the evidence that the rewrite was safe. They test well, and they test themselves.
- 01 / 04
Independent of the tool that made the change
It checks code written by any AI tool or person, and has no stake in the answer.
Product names and logos are trademarks of their owners. Refinery is not affiliated with them. - 02 / 04
Context across the estate
It maps the programs and jobs a change reaches, so you retest what actually changed.
- 03 / 04
Proof, not sampling
It proves supported arithmetic matches across the model, or hands you an input that breaks it.
- 04 / 04
Better with every change
Every accepted risk is kept with its reason in your own register. With precedent recall connected, the next change to that program arrives with what was decided before, and why.
See how that compares to AI migrations, code gates and manual review →
Governance portalOne place to see
One place to see
every change.
Every change, every check, and every sign-off in a single view your risk team can read without an engineer sitting next to them.
Portfolio
5 example changesEvery change, in one view.
Verdicts and sign-offs, together with the code they cover.
ProgramChanged byVerdictSigned byDate
INTCALC.cblwatsonx Code Assistant for ZAccepted as intendedMaya Chen
PAYROLL.cblGitHub CopilotPassedJordan Lee
LOAN-RATE.cblPriya ShahFlaggedNot signed
STATEMENT.cblCursorPassedEva Brooks
BATCH-POST.cblSamir AliPassedOlivia Price
Product preview. Example changes, not real customer data.
The journey so far
Backed by the ecosystems
shaping what comes next.
Refinery has grown through programmes and communities spanning mainframe technology, cyber resilience and ambitious company-building.
Built with ecosystem support01
DiSH Accelerator
Selected for the Manchester programme helping security-focused startups build stronger, more resilient companies.
BARCLAYS EAGLE LABSPLEXALTHE UNIVERSITY OF MANCHESTERLANCASTER UNIVERSITY
Deep mainframe expertise02
IBM Z ScaleUp
Accepted into the cohort for startups building on IBM Z and enterprise-grade infrastructure.
IBM Z
Founder community03
Based Collective
Backed by the European founder collective and selected for the first OpenAI Hacker House in London.
BASEDOPENAI HACKER HOUSE