Inventory. Translate. Validate.
The same discipline every time.
Six steps. Two non-negotiables. Every step produces an artifact you can audit. Nothing irreversible happens without your signoff.
Read-only credentials on your source. The engine crawls metadata, query logs and dependencies — no production read or write. You receive an inventory artifact (Appendix A) that becomes the contract for everything that follows.
Wave plan, dependency-ordered. Effort estimate per asset, complexity tier, owners. 3-year TCO model. Risks recorded with named owners (always human). Out comes a written plan signed off by you before any conversion happens.
Translation runs against the codified rule library. Every output is classified — auto_ok, needs_review, or blocked. Anything ambiguous goes to a review queue. Senior engineers handle the queue. Nothing ships unattended.
Deterministic, LLM-free. Row counts. Schema parity. Float-tolerant checksums. Query-result diffs on a target sample. Bounded cell-diff (≤50 unmatched rows per side) to keep reports readable. The engine never grades its own work.
Per-wave reconciliation report. Each discrepancy gets blame assigned and a recommended fix. Anything that can become a reusable rule goes into the library — the moat compounds across engagements.
Signed evidence delivered to you. Inventory, transformation log, reconciliation results, fix log, runbook. You keep all of it. The artifact is auditable by your team or your auditor without us in the room.
Two rules that stay load-bearing.
Validation is deterministic. Forever.
Row counts, checksums, query-result diffs, schema parity — computed by code we wrote, not by any model. No model is ever asked to grade its own work or another model’s. The validation layer of our codebase stays free of LLM dependencies. This is the line we do not cross.
Humans approve what matters.
Scope approval, pilot acceptance, every wave start, every cutover — gated on a named engineer signing off. The engine prepares the work; humans decide whether it ships. Agents in our engine receive read-only client credentials; production writes happen via generated scripts a human executes.
See the methodology in action.
Read a sample assessment built from this exact pipeline — same inventory format, same wave plan structure, same reconciliation artifacts.