AI modernization, proven
Lightyear’s AI factory converts legacy code at machine speed, and our independent Verdict Engine proves field by field that the new system behaves exactly like the old one. Retire legacy sooner, spend less on testing and sign-off, and cut over with evidence your auditor can replay.
COBOL · CICS · VSAM · IMS · HLASM · PL/I · JCL · Db2 · Oracle · PostgreSQL · AlloyDB · SAP ASE
Try the judge
A rewrite that rounds where COBOL truncates is one cent off on a monthly interest charge. Testers rarely notice. Across a million accounts, that’s $10,000 a month that won’t reconcile. The Verdict Engine compares every field, so it can’t slip through.
Illustrative: CardDemo monthly interest calculation (INTCALC), one account.
What changes for the business
Most programs stall in testing, reconciliation and sign-off, not in writing code. Lightyear removes that bottleneck, so the backlog moves and the legacy bill ends sooner.
AI agents convert programs in parallel, and the judge decides when each one is done. Work doesn’t sit in a test queue waiting for people.
Automated field-by-field comparison replaces long manual parallel runs and spreadsheet reconciliation.
Stop paying to run two systems. Cut over when the evidence says the new one is ready, not when the calendar runs out.
Every change arrives with signed, replayable evidence for auditors and regulators, so approval isn’t the last-minute risk.
Why it has to be independent
AI coding platforms, systems integrators and your own engineers can all produce the new system. None of them should be the one who says it’s right. Lightyear is the referee: we don’t write your code, and the AI that builds it never sees the expected answers.
Works the same whether the code came from an AI platform, an SI or your team. The judge never shares a context with the writer.
Both systems run on real engines with the same inputs. The Verdict Engine compares outputs, data changes and errors, field by field.
Every verdict ships with a signed evidence pack that your auditor or regulator can re-check without trusting us.
We asked AI agents for the rules that translate a 1.4-million-line ERP from Oracle to PostgreSQL. It returned fourteen rules in one response: same format, same confidence.
The agent dark factory
Our factory’s advantage is built in: nothing it writes counts until the independent Verdict Engine says so. The judge is a deterministic program, not a model, and the agents never see its answer key. Prefer your own builder? The same judge verifies that too.
Works from the knowledge graph: what the program reads and writes, which rules it carries, what “the same” means. Rules lock before any code is written.
SEES: source, layouts, the work orderWrites the replacement in a sealed workspace with a fixed budget of attempts.
SEES: source and interfaces, never the answersRuns old and new on the same inputs. The answer key is the legacy program itself, run for real.
HOLDS: the answers and the signing keyDifferences go back by field name only. When the budget is spent the run stops and the failure is kept as evidence.
SEES: which field differs, not whyWatch the loop · ten seconds
Both systems ran. The engine compares every field.
One field differs. The engine names it, not the answer.
The AI fixes its own code and runs it again.
Equivalent. Signed. Anyone can replay it.
The factory does the repetitive work. Decisions that carry risk stay with a named person, and each one is signed and kept with the evidence.
How we say no
When the evidence isn’t there, we say so. An empty output, a missing file or an untested branch never turns into a pass. That refusal is the reason our passes mean something.
Proof, with its receipts
Each claim carries an evidence label and links to the public receipt behind it.
Don’t trust us
Download any public evidence pack and check it with our open verifier. It passes only what it can anchor, and reports “incomplete” rather than guessing.
$ python3 -I tools/verify_public_evidence.py \
all PUBLIC_EVIDENCE --strict --json
0 PASS every check anchored and passed
1 FAIL something doesn’t match
3 INCOMPLETE nothing it could anchor; no pass
How it works
The knowledge graph maps programs, data and business rules, then the rules are frozen before anyone writes code.
Our dark factory, your AI platform or your SI. The builder never sees the hold-out answers.
Old and new run side by side in your cloud, read-only. Outputs, data and errors are compared field by field.
Named differences, a risk figure for untested logic, and an evidence pack anyone can replay.
Coverage, stated honestly
Start with one job
Send us one batch job or process you’ve already modernized, or let our factory build it during the sprint. In four weeks you get a go/no-go answer, the cost of getting to equivalent, and a business case for the rest of your backlog.
Founded by Howard Weale, who led the mainframe modernization practice at Google Cloud and brought Dual Run, Google’s parallel-run validation for mainframes, to market. Lightyear is the independent version of that idea, for every platform and every builder.