Design is where the case study earns its keep — the interface is simple; the decisions behind it are the point.

Three decisions in particular are the ones a reviewer should remember — each is a place I chose the harder-to-explain option because it was the correct one.

The core mechanism

ReconLayer ingests two datasets — the internal valuation and the broker mark-to-market — and matches them trade by trade on a shared key. For each pair it computes a variance; for each unmatched trade it flags a one-sided break. The output is a single ranked worklist, not a spreadsheet dump: the exceptions that matter, largest first, each with a plain-language cause and a recommended action.

Decision 1 — Rules engine, not an LLM, for classification

The obvious "AI product" move is to hand each break to a model and ask why it broke. I didn't. Classification stays a deterministic rules engine.

The reasoning is threefold, and it's the sophisticated part of the story:

• It's the moat. A domain-specific root-cause taxonomy — valuation difference vs. booking break vs. fee/tax offset vs. stale feed — is tacit knowledge a generalist can't fake and an LLM wrapper doesn't demonstrate. If the classifier were just a prompt, the whole tool would be a ten-second dismissal for anyone who knows the space.
• It's auditable. A rules engine can show which signals fired. An analyst trusts "custodian rejected + broker confirms a different account," not "92% confidence."
• It's safe. Classification is a control decision. A hallucinated "cause: fee booking, auto-close" would silently suppress a real break with a dollar cost attached. Deterministic logic can't do that.

Decision 2 — Population breaks are first-class, not a footnote

Most simple reconciliations sort matched trades by variance size. That silently drops the trades that exist on only one side — and a trade one party booked and the other didn't is frequently the single largest driver of a real dispute, yet it can't be expressed as a variance because one number is blank.

ReconLayer surfaces one-sided population / booking breaks as their own pinned group, above the value breaks, flagged for booking confirmation before anything is agreed. This is the fix for the flaw that was in the very first version of the concept, and it's the detail that signals genuine domain fluency.

Decision 3 — The AI writes; the engine decides

The tool is explicitly AI-forward — but the AI's job is bounded. The rules engine decides what broke and by how much; Gemini writes the human-facing artifact— the counterparty dispute email — from that already-classified output.

This keeps the LLM out of every decision that carries consequences, and it produces a defensible one-line architecture: the model drafts, it doesn't decide. That boundary is what separates a credible tool from an "AI wrapper," and it's the direct application of the "keep the human in control" principle from Define.