I Built a Lottery App That Refuses to Predict the Jackpot

TL;DR Hot numbers and "AI predictions" promise an edge on the jackpot that does not exist, and at the minimum jackpot a ticket still loses about 73% of its price on average. vietlott-lab computes Vietlott lottery odds exactly, picks tickets that win some prize more often (23.01% against 22.06% for ten random Mega 6/45 tickets), and prints beside them that the jackpot odds are the same for any ten.

Do this: run vlott pick mega 10 and read the jackpot line before the percentage above it.

Which ten Mega 6/45 tickets give you the best shot at the jackpot? Any ten. Ten distinct tickets win it with probability 10 in 8,145,060, whichever ten they hold.

So vietlott-lab, a Python CLI with a VS Code panel for Vietnam's Vietlott games, says that on the screen where you pick. Then it improves the one thing a pick can improve.

Built with AI, five tools built with an AI coding agent, one post each:

1. I Built a Lottery App That Refuses to Predict the Jackpot (you are here): exact odds.
2. A Spider Crawls Over My Screen and Remembers What It Read: a searchable screen log.
3. An AI Meeting Assistant That Never Hears the Audio: answers from captions.
4. One Skill Folder That Works in Every Coding Agent: rules every agent reads.
5. An LLM Proxy That Retries a 429 Before Your Client Sees It: rate limits, in the open.

Not covered: prompting and model choice.

Every tool in this series has an honesty line

You can read each of the five the same way:

1. Input: a real problem, and one honesty line: the thing the tool must say plainly it cannot do.
2. Process: an AI coding agent writes the code. A test that can fail, or the screen itself, holds the line.
3. Output: a tool that states its limit where you will see it, and a repo you can clone.

For vietlott-lab it is short. No pick changes the jackpot odds.

What ten tickets can buy

Less overlap. Two tickets that share four numbers tend to win on the same draws, so vlott pick searches for the set of N tickets most likely to win any prize:

  P(at least one ticket wins any prize): 23.01%  (exact)
  same count of random tickets:          22.06%
  P(Jackpot):  1 in 814,506  -- the same for ANY 10 distinct tickets
  expected return: 26,960 VND   expected loss: 73,040 VND

Exact means every possible draw

The engine scores your tickets against all 8.1 million Mega draws, or 29 million for Power 6/55. No simulation.

Tickets and draws are bitmasks, so matches are popcount(ticket & draw), run by numpy over every draw at once. The VS Code panel only runs vlott --json, so every number on screen comes from the tested Python code.

The hot-number test was allowed to fail

Hot numbers are the oldest promised edge. Before each of 1,274 past Mega draws, vlott stats takes the six numbers drawn most so far and counts how many the draw contained:

Across 1,374 draws the frequency test gives p = 0.88: no evidence of bias.

Try it

Python 3.10 or later:

git clone https://gitlab.com/agentic-fx/vietlott-lab.git
cd vietlott-lab
python -m venv .venv
.venv\Scripts\activate          # macOS/Linux: source .venv/bin/activate
pip install -e ".[dev]"
vlott fetch all
vlott pick mega 10

Limits: history comes from a public mirror (vietlott.vn refuses connections from outside Vietnam), and the prize tables in games.py need checking against vietlott.vn. Play only with money you can afford to lose.

Every draw checked, and the jackpot line printed beside the pick: that is the part to look for in any odds tool. The honesty line held because every number came from my own CLI. Part 2 starts from input I don't control: whatever is on my screen. When an AI reads it, what may leave the PC, and how would you know?

Read this with your AI

Read the post "I Built a Lottery App That Refuses to Predict the Jackpot".
1. Explain in 3 bullets why tickets that overlap less raise
   P(win any prize) but not P(jackpot).
2. Ask me 2 questions to check I understood expected value; wait, then correct me.
3. Name the honesty line of a tool I use, and what in it could fail.