Benchmark 1: SlotKeeper

Build a web app called "SlotKeeper" for a small service business to take bookings.

Public side (no login needed):
- A landing page showing the business name "Maple Yoga Studio" and a list of upcoming time slots (date, start time, duration, spots left).
- Clicking a slot opens a booking form: full name, email, phone (optional), notes (optional). Validate that name and a valid email are required.
- If a slot has spots left, booking it reduces spots by 1 and shows a confirmation page with the booking details.
- If a slot is full, the button says "Join waitlist" instead, and the same form adds the person to that slot's waitlist with their position number.
- The same email cannot book the same slot twice.

Owner side (login required):
- Email + password sign up and login for the owner. Only logged-in owners can see the admin area; logged-out visitors trying /admin are redirected to login.
- Admin dashboard: create, edit, and delete slots (date, start time, duration in minutes, capacity).
- For each slot, see the list of bookings and the waitlist. The owner can cancel a booking; when a booking is cancelled, the first person on the waitlist is automatically moved into the booking.
- A simple stats line at the top: total upcoming slots, total bookings, total waitlisted.

Requirements:
- Store all data in a real persistent database (data must survive a page refresh and a redeploy).
- Clean, modern, mobile-friendly design.
- Seed 3 example slots for the next 7 days, one of them with capacity 1.
- Deploy it to a public URL I can share.

Identical text was pasted into every tool. After it: up to 5 fix prompts that only describe observed behaviour.

NoteLovable: Under review: 86.25 if our 5-point critical-bug rule is applied (the scorecard took 4). Corrections.

NoteReplit Agent: 87.625 if time counts to the agent's 13-minute finish, which would rank it above Lovable (87.25, or 86.25 under review). Corrections.

NoteChatGPT Codex: General-AI baseline, not ranked: a Plus-level plan, 2 fix prompts and a human-made hosting account. Bug score under review: 83.5 under our −1-per-minor rule. Corrections.

NoteLovable: Under review: 86.25 if our 5-point critical-bug rule is applied (the scorecard took 4). Corrections.

NoteReplit Agent: 87.625 if time counts to the agent's 13-minute finish, which would rank it above Lovable (87.25, or 86.25 under review). Corrections.

NoteChatGPT Codex: General-AI baseline, not ranked: a Plus-level plan, 2 fix prompts and a human-made hosting account. Bug score under review: 83.5 under our −1-per-minor rule. Corrections.

• Points earned
• Points lost
• Bar segments, left to right: works 30 · time 15 · bugs 15 · design 15 · deploy 15 · cost 10

Bolt.new · 94.75

Base44 · 93.5

v0 · 90.875

Lovable · 87.25

Replit Agent · 83.625

ChatGPT Codex · 82.5

• Data surviving a redeploy (all tools). Checklist item 4 was credited after a refresh check only; see corrections.
• A second owner trying to read another owner's data (all tools).
• Whether database keys are exposed in the browser (Bolt; not checked on the others).
• Backend reads without logging in (Base44 was probed at the UI level only).

• One general agent so far: ChatGPT's Codex ran as the baseline on Oct 25, on a Plus-level plan. Claude, Grok and Gemini weren't asked to build the app; they're added at the retest. See categories.
• Specialists on free plans only: no cheapest-paid-plan runs yet.
• One app type: a booking app.
• One run per tool, so luck matters and we can't show spread yet.
• One design reviewer.
• Tools change often; every result is dated.
• Not tested yet: Emergent, Claude Code.