Notes from the work.
Long-form writing from our team on modernising real legacy systems — tri-language analysis, honest assembler comprehension, governed AI autonomy, and why so much of this comes down to refusing to simulate.
Reading
What we have been thinking about.
Language coverage
Why one analysis surface for three languages changes the work
Most estates are not written in one language. Reading COBOL, PL/I and assembler on the same surface turns cross-language dependencies from a blind spot into something you can see.
Assembler
Honest assembler analysis: coverage, and provenance
Assembler is where migrations stall. Complete coverage matters — but so does telling the engineer what was parsed and what was inferred, rather than hiding it in a black box.
Methodology
Governed autonomy: letting AI deliver without handing over the keys
A deterministic controller driving an AI maker, at graded autonomy, with a hard denylist and an audit log. Why the Epic Loop is a safety mechanism, not a shortcut.
Testing
Simulation is forbidden: the case for real execution
A test that runs against a mock is measuring the mock. On a legacy estate, the only tests worth trusting are the ones that run against the real runtime.
Prefer to talk it through?
Bring the hard question.
Our writing is a starting point. The real answers come from putting your estate in front of the engineer who would do the work.