Six choices that compound.

The keystone of the entire system is a versioned data contract — a structured ReserveStudy payload that flows through every component. Intake produces it. The generator consumes it. QA validates it. The platform mirrors it. The corpus archives it.

The contract is designed once, with the API-integrated end-state as a forward-looking constraint, so the schema we ship in the standalone tool is the same schema that flows through the platform API later. Every workstream plugs into the same shape, which means every new capability is additive — the contract grows, and the system grows with it.

When an engineer's judgment about a community — condition ratings, component sub-attributes, sub-area assignments, cost variances — is captured in structured fields at the moment the call is made, the system gets a richer, more reliable picture than it could ever reconstruct from cells later.

The Smart Intake form takes over the engineer-facing input surface — replacing the slower screens of the funding plan app and pushing structured data directly into its backend via API. The engineer captures judgment once, in context, with the right field types and inline validation. Better inputs mean better outputs — and every component decision the firm makes becomes part of a queryable record the team can search, cite, and learn from.

The renderer becomes a Python module that builds every report from clean code — every time. Cover page, headers, financial tables, signature block — all assembled programmatically from the validated ReserveStudy payload. Same code path, every community, every quarter.

The visual identity of a GF report becomes versioned, testable, and absolutely consistent. When the design needs to evolve — a new section, a refreshed cover, a Board-requested tweak — the change ships everywhere at once. No migrating templates, no per-report drift, no surprises.

Voice isn't a single prompt — it's the firm's accumulated craft. The narrative generator retrieves from multiple parallel sources of GF institutional knowledge: vocabulary patterns at the sentence level, historical exemplars from the firm's best delivered reports, component templates by community type, and benchmark cost data.

Every output is informed by GFE's best historical work. As more reports are delivered and reviewed, the corpus grows. The platform inherits the firm's craft — and gives it back, sharper, on every new project.

When the system encounters something that benefits from a human eye — an unmatched cost, a component without a clear analog, an interpretation that calls for engineering judgment — it surfaces that span visibly in the draft.

Every flagged span is highlighted in the document, indexed in the engineer-review summary, and accompanied by the reason for flagging. Engineers see exactly what needs their attention; everything else they can trust at a glance. Confidence is encoded into the document, ready for them.

The generator is a layered pipeline. Code handles what code is great at — ingestion, validation, section planning, QA, rendering — all testable, debuggable, version-controlled. The LLM handles what it's great at — narrative generation, bounded by structured Pydantic outputs and validated by deterministic QA on every side.

Each layer gets the engineering treatment it benefits from. Software practices apply where they work best; iterative prompt refinement applies where it works best. The result is a system that's both reliable and expressive — maintainable for the long haul, alive in the writing.