Collecting · 10,000,000 simulated collections

How many tries to collect all 20 items, one at a time?

Each try gives you one of 20 equally common items at random. How many tries until you have them all?

72.0tries on average, exactly (to 0.1)
simulated: 71.96 ± 0.01 (95%) over 10,000,000 collections

On average it takes 72.0 tries to collect all 20 items one at a time. Half of collectors finish within 67 tries, 90% within 103, and 99% within 149. Worked out exactly and confirmed by 10,000,000 simulated collections that you can replay on this page.

What the computer did

Games: 10,000,000 (1,000 chunks of 10,000, in 1 segment)
Compute: 0.14 CPU-hours; 0.01 hours on arm64, 16 cores, Darwin 25.5.0
Engines: Python 3.14.8; every segment is spot-checked by replaying chunks in the other language
Ran: 2026-10-31T15:00:35Z to 2026-10-31T15:01:09Z
Seed: 1867922717 (from this page's address); chunk k uses xoshiro128** seeded by splitmix32(seed ⊕ k·0x9E3779B1)
Exact method: Markov chain on how many items you own, in double precision
Code: sims/rng.py sha256 b81126f5f114cdf1…
sims/kinds/collector.py sha256 0365392a9406cda1…
site_src/sims.js sha256 5cd91ae0e3f8246e…

All 1 segments of this run

#	Chunks	Chunk size	Games	Engine	Finished	Code (sha256)
0	0–999	10,000	10,000,000	Python	2026-10-31	b81126f5 0365392a 5cd91ae0

Three checks your browser runs

The totals add up. Adding up every published chunk…

It agrees with the exact answer. Your browser is working out the exact answer by a separate method…

The chunks are real. Pick chunks below and your browser replays them from the published seed. Each must match to the last game.

Replay any chunk

The run is split into 1,000 chunks of 10,000 games. Each chunk has its own random stream, built from the seed and the chunk number, so any one of them can be replayed alone. Your browser picks which ones at random, so the run could not have been prepared for your choice.

replayed by youreplayed by other readersnot replayed yet

or chunk

No chunks replayed yet. Each cell above is a chunk; replayed ones turn solid.

What your replays rule out

Suppose someone had made up a share of the chunks instead of running them. These are the chances your replays so far would have landed on at least one made-up chunk.

–if half the chunks were made up

–if 1 in 10 were

–if 1 in 100 were

Breakdown against the exact answer

Outcome	Simulated	Exact	Off by (std. errors)
Average packs	71.96	71.95	+1.09
Done within 67 packs	49.99%	50.01%	-1.34
Done within 103 packs	90.21%	90.21%	-0.07
Done within 149 packs	99.05%	99.04%	+0.88
Fewest packs in any game	21
Most packs in any game	369

Independent simulations by local models

Each model below was given only the rules of this family of questions, in plain English, and asked to write its own simulation. It never saw our code. Its program was run in a sandbox on two test cases and compared with the exact answers.

gemma4:26b agrees z = +0.17; z = -1.31
gpt-oss:20b agrees z = -0.28; z = -0.30
qwen2.5-coder:7b agrees z = -0.71; z = -0.62
qwen3.6:27b-mlx agrees z = -0.28; z = -0.30

See the code each model wrote

What a match proves, and what it doesn't

A matching chunk proves those games come from the published code and seed. Nobody could guess every count exactly without playing the games.
It does not prove the code plays the game described. That is what the exact answer and the independent simulations are for.
The run (Python) and the replay (JavaScript) were written by the same author, so they are two copies of one program. Their agreement shows the numbers were computed, not invented.
One reader replaying a few chunks rules out large fakes, not small ones. Every reader's replays are added up, so together they cover the ledger.