Migrating from commercial platforms to open source

Research code locked to a commercial platform has a structural problem: every user, collaborator, student, reviewer and CI runner needs a license, and reproducibility acquires an expiry date tied to a vendor. Migration to open alternatives removes the toll booth and makes the work FAIR in practice (rseng-fair-software) - but a naive "translate the syntax" migration produces subtly wrong science. The discipline: pin behavior first, choose the target honestly, migrate incrementally with numerical parity as the acceptance test. The general safety net (characterization tests, strangler migration) lives in rseng-legacy-code; this skill adds the platform-specific knowledge.

Before migrating: pin the baseline

While the commercial license still works, capture everything the migration must reproduce (this window may not stay open):

• Run the original on representative inputs and save inputs and outputs as the golden baseline (rseng-legacy-code's characterization tests).
• Record the platform version, toolbox versions and settings that produced the baseline.
• Export data OUT of proprietary formats now: .mat to HDF5-based or open formats, SAS/SPSS/Stata files to open tabular formats (rseng-scientific-file-formats) - data freedom precedes code freedom, and readers for proprietary formats are better used while a licensed installation can verify the export.

Choosing the target

• MATLAB, two honest paths: GNU Octave runs most plain MATLAB with minimal changes - the low-cost exit when the goal is simply license freedom; Python/NumPy/SciPy is the larger move that buys ecosystem, packaging (rseng-project-scaffolding) and hiring, at the cost of a real port. Choose Octave for frozen-but-must-run code, Python for code with a future. (Simulink has no clean open equivalent - flag it as the hard part early.)
• IDL: Python is the community-standard destination (astronomy's stack - astropy and friends - exists precisely from this migration).
• SAS/SPSS/Stata: R (statistical depth, native model objects) or pandas/statsmodels; verify statistical defaults match - the same named procedure can use different degrees-of-freedom or contrast conventions, which is a science difference, not a bug.
• Mathematica: SymPy for symbolic work, Julia for symbolic-numeric blends.
• LabVIEW and instrument control: Python instrument stacks are the destination, but hardware interfaces make this a re-engineering project, not a translation.
• Whatever the target, check whether the field already has an open reimplementation of the domain workflow before porting line by line (rseng-software-reuse) - many "migrations" should be adoptions.

Translation pitfalls (where silent wrongness lives)

The famous ones an agent must actively check, not discover:

• Indexing: MATLAB/Octave/R/Julia are 1-based, column-major; NumPy is 0-based, row-major. Every hand-translated index and every reshape/flatten is a suspect until tested.
• Copy semantics: MATLAB copies on assignment; NumPy slices are VIEWS - in-place modification after translation corrupts data that MATLAB code safely mutated.
• Broadcasting and implicit expansion rules differ in edges; element-wise vs matrix operators (.* vs *) invert their defaultness between MATLAB and NumPy.
• Toolbox calls map to ecosystems, not functions: Signal Processing Toolbox to scipy.signal, Statistics Toolbox to statsmodels/scipy. stats, Image Processing to scikit-image - map the WORKFLOW to the library's idiom instead of reimplementing the MATLAB function signature.
• Numerical defaults: solvers, tolerances, RNG algorithms and seeds differ across platforms; identical seeds do NOT give identical streams, so statistical results match in distribution, not bitwise (rseng-numerical-accuracy sets the tolerance discipline, rseng-reproducibility the seed bookkeeping).
• Automatic translators (MATLAB-to-Python converters and similar) produce a starting draft at best: un-idiomatic, license-check the output, and every translated function still needs its parity test. Interop bridges (oct2py running Octave from Python) are better used as migration scaffolding - call the old implementation module by module while the new one grows - than as a destination.

Migration as a project

1. Inventory and rank: which scripts/functions matter, which are dead; migrate the load-bearing path first.
2. Module-by-module with parity gates: port a module, run both implementations against the baseline inputs, assert agreement within scientifically justified tolerances, only then delete or bypass the original (strangler pattern - rseng-legacy-code).
3. Adopt the target's ecosystem hygiene as you go: environments and lockfiles (rseng-reproducible-environments), tests (rseng-testing), CI without license servers - the ability to run tests on every commit is a migration dividend (rseng-ci-cd).
4. Keep a translation log: mapping decisions, tolerance justifications, known behavior differences - the scientific record of the migration (rseng-documentation), and cite the original code's authors (rseng-citation-metadata).
5. Announce the migration to users with a compatibility note (rseng-science-communication) and license the freed code properly (rseng-licensing) - opening the platform without opening the license wastes the trip.

Working with this skill

This skill is source-independent: it encodes community migration practice between scientific computing platforms. The generic legacy-code safety net is rseng-legacy-code; this skill is its commercial-platform specialization.

Learn more (verified):

• https://octave.org - GNU Octave
• https://numpy.org/doc/stable/user/numpy-for-matlab-users.html - NumPy for MATLAB users (official mapping guide)
• https://scipy.org - SciPy ecosystem
• https://github.com/blink1073/oct2py - oct2py Octave-Python bridge
• https://www.sympy.org - SymPy symbolic mathematics
• https://julialang.org - Julia
• https://pandas.pydata.org - pandas
• https://docs.astropy.org/en/stable/ - Astropy (the IDL-exodus ecosystem)

Related skills

Check whether any of these applies before moving on:

• rseng-legacy-code - characterization tests and strangler pattern
• rseng-licensing - license the freed code
• rseng-numerical-accuracy - parity tolerances across platforms
• rseng-reproducible-environments - target-ecosystem pinning as you go
• rseng-scientific-file-formats - exporting proprietary data formats first
• rseng-software-reuse - adopt an existing open reimplementation