PRD — End-to-end Mac app demo with A1 specialist

Goal

The product-shaped artifact that the platform’s pitch hinges on: a Mac app the user can open, point at their data, pick a base model, train a LoRA, and run inference on the specialist they just trained — all without leaving the app. Today the Mac app ships training controls + an eval tab but no “specialist factory” wiring; B6 closes that loop end-to-end.

Visible-to-user product story: “open the app → drop in some data → click train → 30 minutes later, here’s your specialist running on your machine.”

Why now

• A1 (CLI specialist) ships first. B6 is the GUI wrapper around the recipe A1 proves works. Order: A1 validates the science; B6 packages the experience.
• Existing app shells (app-eval-tab.md, app-train-controls-thermal.md) already cover the supporting tabs. The missing piece is the “factory” tab that chains: data import → recipe pick → train → eval → deploy.
• The platform’s “Mac specialist factory” framing in PLAN.md needs a product-shape artifact, not just a CLI.

Scope — in

• New tab in the existing app: Factory — wizard-style flow:
  1. Pick your task (preset: tool-calling, shell, SQL, custom)
  2. Drop in your data (drag JSONL/JSON/CSV; the app validates against the recipe’s expected schema; shows a preview)
  3. Pick your base (from the model picker; shows size + capability tags; default = the curated A1 base)
  4. Recipe summary (read-only view of the SFT recipe; “Start training” CTA)
  5. Live training view (loss + grad-norm + LR — embeds the C10 train viewer; cancel + checkpoint controls)
  6. Eval gate (auto-runs the recipe’s domain eval; passes or fails the ship gate; ships an adapter file on pass)
  7. Try it now (opens a chat with the new specialist; before/after compare against the base 0-shot)
• The wizard is a single SwiftUI view tree calling the existing tinygpt train / eval-* subprocesses through the existing ServerController + ProcessRunner.
• A “Save recipe” button so a working recipe becomes a sharable .tinygpt-recipe file users can ship to others.

Scope — out

• Custom recipe authoring in-app (multi-stage pipelines, conditional steps). V1 = presets + load .tinygpt-recipe files.
• Cloud training as a button. The Mac is the training surface by design.
• App Store distribution. TestFlight + direct download for V1.

Files to touch

Don’t touch

• The existing Train + Eval tabs — Factory is a new sibling, not a refactor.
• tinygpt train / eval subcommands — V1 calls them as subprocesses through the existing ServerController.

Acceptance criteria

• User can open the app, run the tool-call preset against a test JSONL, train + eval + deploy a specialist in one session on M5 Pro in < 1h wall-clock.
• The eval gate fires correctly (passes when the trained adapter beats the gate; fails-with-actionable-message otherwise).
• “Try it now” tab opens chat against the new specialist with a baseline side-by-side comparison.
• Save-recipe produces a .tinygpt-recipe file that another install can load to reproduce.
• User-doc walkthrough reproducible from a clean install.

Reference patterns

• app-eval-tab.md — existing tab pattern.
• app-train-controls-thermal.md — existing live-training view.
• C10 train-run-dashboard — the browser-side companion; the in-app view can share the chart components if useful.

Open questions

• Whether to support custom-base import inside Factory (drag a .tinygpt / .safetensors from disk). Recommendation: yes — the existing model picker already supports it; pipe through.