The architecture.

Next.js App Router screens: landing, ingest, auth, projects, workspace, and static product surfaces.

Typed frontend clients already mark where ingestion, metrics, files, reports, and assistant calls will later switch to fetch().

Quick extraction, file sniffing, and Ethos review handle obvious work before deeper reasoning.

Only ambiguous summaries, methodology judgement, and full project sweeps need model routing.

Metric strings, known log markers, basic file metadata, and simple status checks should cost nothing.

Short notes, ambiguous stage routing, and small summaries can go through a cheaper model route.

Methodology soundness, leakage risk, and project-level critique require a stronger reasoning path.

The durable container: title, objective, domain, status, tags, stage order, report state.

Large phases hold smaller workflow blocks. They are separate from iterations so the science stays readable.

Every run captures what changed, files, metrics, warnings, notes, status, and next action.

Files, metrics, assumptions, decisions, checks, and report sections reference the relevant stage or iteration.

This pass keeps the frontend self-contained and deployable without environment variables.

A future service owns parsing, routing, metrics extraction, report rendering, and model calls.

Enterprise and university deployments can isolate storage, routing, models, and audit controls.