Bob's Twin

THE PROBLEM When AI upgrades Python libraries (Pydantic v1→v2, Flask 2→3, SQLAlchemy 1→2), the official bump-pydantic tool mechanically fixes ~80% of syntax. The remaining 20% — Enum serialization shifts, Decimal JSON encoding flips, validators that silently no-op, __root__ model shape changes — passes type checks and unit tests but breaks the wire contract. Stack Overflow 2025 reports 96% of developers don't fully trust AI-generated code, yet only 48% verify before commit. Production regressions ship invisibly. THE SOLUTION Bob's Twin wraps any migration in a deterministic Capture → Migrate → Replay → Audit pipeline: - CAPTURE — VCR.py records HTTP behavior of the legacy app into hash-stamped golden cassettes (append-only evidence). - MIGRATE — IBM Bob's "modernize-with-twin" Custom Mode walks Bob through migration in commit-sized steps with Checkpoints. bump-pydantic runs first; Bob handles the long-tail patterns it can't. - REPLAY — Same recorded inputs against migrated code. DeepDiff with tolerance rules surfaces every diff: which endpoint, which field, what changed. - AUDIT — Hash-chained JSONL trail + signed PDF report. The same evidence shape EU AI Act Article 12 (effective Aug 2, 2026) requires for AI-modified production code. WHY IBM BOB Four Bob features are load-bearing, not optional: Custom Modes gate the 4-phase pipeline; Skills (SKILL.md) package the workflow as a redistributable team asset; the MCP server exposes 6 tools natively callable from Bob; Checkpoints auto-rollback when equivalence drops below 0.95. Replicating this on Cursor or Claude Code requires manually scripting all four layers. THE PROOF In our demo, v2_naive scores 0.595 equivalence — 15 endpoints silently broken. Bob in modernize-with-twin mode reads the diff report, regenerates the broken patterns, and brings the score to 1.000. Open-source Apache 2.0.