Can software be replaced without changing what it does?

The problem

Why this matters outside computer science

Public institutions run software that outlived its builders: benefits calculations in COBOL, hydrological and traffic models in Fortran, case management in VB6 or Delphi. These systems encode decades of policy, exceptions and corrections that exist nowhere else — not in the documentation, and not in anyone's memory.

The barrier to replacement is therefore not engineering capacity but epistemic: nobody can state with confidence what the current system does, so nobody can prove a replacement does the same. Organisations stay with a vendor or a language because leaving cannot be shown to be safe. Software independence fails on a verification problem.

Large language models now rewrite such code fluently. Whether the result is equivalent is an open and, so far, largely unmeasured question. That gap is what this project addresses.

Research question & method

Verified legacy migration

How reliably can agentic AI migrate a legacy codebase to a modern stack while preserving its observable behaviour, and where exactly does it fail?

1. Characterize Derive the behaviour of the running legacy system from real inputs, outputs and state transitions — not from documentation — and express it as an executable characterization-test suite.
2. Migrate Run multi-agent workflows that translate the codebase to a modern, open target stack, with the test suite withheld from the migrating agents.
3. Verify Run both systems against identical inputs and compare results case by case (differential testing), including edge cases, error paths and numeric tolerance.
4. Measure Record equivalence rate, failure taxonomy, human interventions required, and cost per thousand lines — per language, per system size, per model.

Scope

Only publicly available legacy codebases are used, so results are reproducible and nothing proprietary or confidential enters the work. Target languages are chosen for public-sector relevance: COBOL, Fortran, VB6 and Delphi first.

Deliverables

First 6 to 12 months · everything public

Code under a permissive open-source licence, data and report under an open licence. Links are published here as each component is released.

People

Project lead

Sofcom is a small, independent effort. The project is open to co-investigators; researchers joining the work are listed here by name.

Collaborate

Open to researchers, lectorates and public-sector teams

Write to info@sofcom.org. Co-authorship on resulting publications is the default for substantive contributions.