Skip to content

Offshoot v1.0.0

Choose a tag to compare

@ved-cmpt ved-cmpt released this 08 Nov 20:39
· 12 commits to main since this release

Offshoot: AI-assisted COBOL → C# modernization, with proof.

Live demo: https://offshoot-go2y.onrender.com. Sign in as viewer, engineer, architect or admin, all with the password demo1234. It is a free instance, so the first visit after 15 idle minutes takes about 45 seconds.

What is in v1.0.0

  • Deterministic inventory of a synthetic 1990s COBOL billing system (BillRite): copybooks, call graph, data layouts, dead code.
  • AI rule extraction with every cited span and data name checked by the server. A person verifies each rule; the model can never approve anything.
  • A modernization plan, then migration to C# in a network-less sandbox that forbids processes, network, file system and reflection, with a bounded repair loop.
  • Golden-master equivalence. The real COBOL and the new C# run on the same 500 seeded inputs, compared field by field. Legacy paragraph coverage and Stryker.NET mutation testing back it up, and server-side gates decide.
  • Human approval and a strangler-fig facade (YARP) with legacy, shadow and modern routing per module.
  • A dashboard, an assessment report and an append-only audit log.
  • A public demo that never runs code: it shows stored runs of the real sandbox, and resets itself daily.

The headline story

The first C# translation of the rate calculator rounded half a cent to the even cent; COBOL rounds it up. Its own unit tests passed, but the harness found 16 of 500 cases off by exactly one cent and blocked the module. One repair later, 500 of 500 matched, and the module went through shadow mode (20 requests, 0 mismatches) to modern.

Measured

  • Mock pipeline: rule precision 95.65% and recall 88.00%; 4 of 4 modules compiled; 3 of 4 were equivalent on the first attempt and 4 of 4 after one repair.
  • Live Gemini sample (gemini-3.5-flash, 20 calls under a hard cap of 30):
    • Rule precision 93.33% and recall 63.64% on two programs.
    • The live RATECALC migration compiled and matched 500 of 500 cases on the first attempt.
    • The first 17 calls failed, and that is recorded. Gemini refused the structured response schema, and the adapter's retries then hit the rate limit.

Details are in docs/EVALUATION.md.

Quality

pytest 787, Angular 163, .NET 26, browser and accessibility suite 18, smoke checklist 14 (also against the live demo). CI checks every push; the slow suites run weekly.

See CHANGELOG.md for the full history and docs/RUN_REPORTS/ for what each of the six runs did and proved.