Whatever AI your team uses, get an independent verdict on its work
Your developers and vendors are already using AI agents to rewrite legacy code: Claude, Codex, Google's ADK or their own. Lightyear Verify lets any of them hand what they've built to an independent judge, inside your environment, and get back a signed verdict your risk committee can rely on.
AI writes the code fast. Nobody can prove it's right.
An AI agent can rewrite a COBOL batch job in an afternoon. The hard question comes after: does the new version produce exactly the same results as the system your business has run on for thirty years?
Today the answer usually comes from the same team, or the same AI, that wrote the code. That's not evidence a risk committee, an auditor or a regulator will accept.
What decision makers need to know
- Does the replacement behave exactly like the original?
- If not, where does it differ, and why?
- Who checked, and can someone else check it again?
- Did our data stay inside our walls?
Your team keeps its AI. We provide the judge.
Work as usual
Your team, or your modernization vendor, rewrites a legacy program with whichever AI agent they prefer.
The agent asks for a verdict
It connects to Lightyear Verify through MCP, the open standard most AI agents already use, and submits the new program.
The judge checks it independently
Inside your environment, it runs the program on your own test data and compares every field with what the legacy system produced.
You get signed evidence
A pass, or exactly where the results differ. Every attempt is recorded and can be re-checked by anyone, months later.
Proof you can take to the committee
Any AI, one standard of proof
Claude, Codex, Google's ADK or an in-house agent: all are checked against the same judge. Change AI vendors without changing how you sign off.
The AI can't mark its own homework
The agent never sees the expected results. It learns where it's wrong, never the right answer, and gets a fixed number of attempts.
Your data stays with you
The judge runs in your environment. On confidential data, the agent learns only pass or fail and which output differs.
Evidence that survives an audit
Every attempt is signed and kept. Your auditor can replay any verdict offline, without us in the room.
What you can use today, and what's coming
Available in the preview
- Mainframe batch (COBOL) to Java, starting with the CardDemo interest calculation
- Tested live with Claude Code
- Runs inside your environment, with confidential mode for sensitive data
- Signed, replayable evidence for every attempt
- People approve every exception; the AI can only propose
Coming next
- Live tests with Codex, Google ADK and Strands
- More programs and platforms, including Oracle-to-PostgreSQL business processes
- Qualification on output from a live mainframe
- A packaged install or a hosted judge
- Self-serve pricing
Ten tools. None of them reveal the answers.
Verify is an MCP server with two trust levels. The tools beside the agent hold no secrets. The judge runs as a separate account, with the expected results, the signing key and the attempt budget out of the agent's reach.
On non-public data, confidential mode limits what an agent can learn from each verdict, and attempts are counted against the data itself. We publish the limits, too: the operator guide sets out the threat model.
| Tool | What it does |
|---|---|
| describe_copybook | Field layout of a legacy record |
| decode_records | Reads public development records |
| read_job_log | Parses a public job log |
| lane_status | What the judge is qualified to compare |
| get_task | The work order and attempt budget |
| submit_candidate | Submits the new program |
| get_verdict | Pass, fail, and where it differs |
| get_budget | Attempts left |
| get_receipt | The signed receipt |
| propose_normalization | Drafts a request for a person to decide |
Put your AI agent in front of an independent judge
Tell us which AI your team or vendor uses and which system you're modernizing. We'll set up the judge in your environment and run your first verdicts with you.