Structured targets matter
Teams that had sequenced extraction targets (which module first, what depends on what) progressed faster than teams exploring freely. This shaped extraction-plan.json.
Ongoing Experimentation
ModernizeSpec is an evolving specification. We are currently running real-world experiments to validate and refine it — and we update the spec as we learn.
Active Experiment
Section titled “Active Experiment”PearlThoughts has deployed 7 intern teams working in parallel on ERPNext modernization. Each team takes a different approach: RAG-assisted code understanding, AST-based transpilation, graph-augmented retrieval, and direct extraction with parity testing.
The goal is not to produce a finished migration. The goal is to learn what information AI agents actually need when working with a large legacy codebase — and encode those needs into the specification.
What We’re Learning
Section titled “What We’re Learning”Parity needs baselines
Self-reported accuracy without baseline snapshots creates false confidence. This led to parity-tests.json requiring captured behavior, not just claims.
Progress tracking is essential
Without shared state, multiple teams converged on the same module unknowingly. This drove migration-state.json with per-context progress and blockers.
Multiple approaches coexist
RAG, AST parsing, direct extraction, and hybrid approaches all produced useful output. The spec is deliberately approach-agnostic — domains.json works with any toolchain.
How Experiments Feed the Spec
Section titled “How Experiments Feed the Spec”Every lesson gets encoded:
| What We Observe | How It Shapes ModernizeSpec |
|---|---|
| Teams need to know where to start | complexity.json scores hotspots so agents prioritize correctly |
| Teams duplicate effort without coordination | migration-state.json tracks who is working on what |
| Teams need to verify behavioral equivalence | parity-tests.json defines test structure with baseline snapshots |
| Teams need extraction order | extraction-plan.json sequences phases with dependency awareness |
| Different AI toolchains need the same context | The spec is tool-agnostic JSON — any agent reads it |
Current Status
Section titled “Current Status”The experiments are ongoing. As we gather more data, we will:
- Publish quantitative results from the experiment rounds
- Refine spec schemas based on what agents actually query
- Add new specification files if gaps emerge
- Document patterns that work across different team sizes and AI toolchains
The specification is v0.1 precisely because we believe real-world usage should shape the standard, not theoretical design. We ship, learn, and iterate.
See Also
Section titled “See Also”- Migration Experiments — The reference Go extraction approach
- Platform Analysis — Understanding ERPNext’s architecture
- Feasibility Research — Cost projections and scenario analysis