Atlas of Doors is a roguelike deckbuilder that runs in a browser tab. No download, no account, no sign-up. It is also an experiment in building a game the way you would build a product: every door opened, every card played and every run walked away from is a logged decision, so almost nothing here has to be a matter of opinion. Month one had a million of those decisions. Month two has 3.7 million, a world twice the size, and one very expensive lesson about who you buy your players from.
This is the whole dashboard, read straight off the live event stream. Nobody maintains a spreadsheet behind it, and every metric on it carries its own written definition so it cannot quietly drift between versions and flatter us by accident.
Totals always go up, because time passes and we keep buying traffic. So here is the honest version: the same measures, one month against the next.
| Measure | Month 1 Jun 27 – Jul 18 | Month 2 Jul 19 – Aug 18 | Change |
|---|---|---|---|
| Decisions logged | 1,058,068 | 2,719,995 | 2.6× |
| New players | 10,246 | 20,239 | 2.0× |
| Sessions | 16,495 | 36,415 | 2.2× |
| Runs started | 11,068 | 29,365 | 2.7× |
| Runs finished | 8,369 | 20,562 | 2.5× |
| Runs finished by returning players | 3,734 | 12,867 | 3.4× |
| Runs finished by first-timers | 4,635 | 7,695 | 1.7× |
| Releases shipped | 18 | 32 | 1.8× |
These are the four charts I actually read, because they divide by the number of players and therefore cannot be flattered by simply spending more on ads.
| Per active player | End of month 1 | Now | Direction |
|---|---|---|---|
| Sessions / player | 1.62 | 1.74 | Up |
| Runs / player | 1.09 | 1.33 | Up |
| Finished runs / player | 0.82 | 0.95 | Up |
| Committed players | 277 (2.7%) | 669 (2.2%) | More people, smaller share |
A committed player is one who shows up on two or more separate days and averages ten or more active minutes on the days they show up. Idle background tabs cannot inflate it. There are now 669 of those, up from 277, and they are the only group who are unambiguously playing rather than sampling.
The share falling from 2.7% to 2.2% while the count more than doubles is the whole month in one line. We poured in twice as many people, and the new people were, on average, worse. That is the next section.
Halfway through the month our numbers fell off a cliff. It was not the game. It was who we were buying, and we can name the day, the hour and the cause.
Every day we get some number of brand new devices. The only question that matters about them is whether they go on to actually start a run, or bounce off the door. Through late July that ran in the forties. On Jul 31 it hit 22.6%, and stayed low for four days.
A new ad account gets cheap inventory, because the platform does not know yet who to show you to. Cheap is the defect. We were paying about 23% less per thousand impressions and getting between a half and a third as many people who actually played.
Honest caveat on that third bar: the recovery is higher than the old baseline, and the ad account is not the only thing that changed. From early August we were also testing different opening screens, and one of those arms deals you a hand of cards immediately, which makes “started a run” easier to reach. The fall is cleanly attributable. The size of the recovery is not. Aug 18 then fell back to 29.5% for one day, and we do not yet know why. That is the last day in this window, so it is a loose end rather than a finding.
We killed the theory that it was our own build in the plainest way available: on Jul 31, both accounts ran the same day, on the same version, and the old one converted at 36.8% while the new one managed 3.9%. There is no version of “the game got worse” that only affects one advertising account.
Run volume more than doubled. Finished runs did not keep up. Underneath that average, two populations are moving in opposite directions, and only one of them is a problem.
Our best day was Jul 29: 1,362 runs started and 1,106 finished, a finish rate of 81%. August recovered the starts, reaching 1,258 on Aug 11, but only 876 of those finished. The gold bars stopped filling the violet ones.
At the end of June, a first-timer played four out of every five finished runs. By mid August, returning players play seven or eight out of ten on most days. The bar quietly inverted, and that is the single best signal in this whole write-up.
| Who | Runs started | Runs finished | Finish rate |
|---|---|---|---|
| First-timers, month 1 | 7,363 | 4,635 | 62.9% |
| First-timers, month 2 | 16,659 | 7,695 | 46.2% |
| Returning players, month 1 | 3,705 | 3,734 | — |
| Returning players, month 2 | 12,706 | 12,867 | — |
Returning players have no finish rate in that table on purpose. A run started at 11pm and finished the next morning counts as started on one day and finished on the next, which pushes their ratio just past 100% and makes it meaningless. Their volume is the honest measure, and it tripled.
This is the section where the numbers are bad. All three of them are the same problem wearing different clothes: people do not stay long enough, on the day or across days. Season VI is aimed squarely at it.
Of the devices old enough to count, the share that opened the game again the very next day. It was 4% a month ago. Mobile web games generally floor around 10 to 15%.
Still playing a week later. This one has actually climbed from roughly zero at launch to a peak of 2%, and sits at 1%. The direction is right and the size is not.
Day 7 is the lever worth the most over time, because a change that lifts it lifts it for every future group of players as well, not just for the people currently in the game. Going from zero to a 2% peak in two months is genuinely good news buried inside a genuinely bad number.
How long a session lasts, counted only over sessions where somebody actually started a run. It was 4.5 minutes a month ago. The typical web game runs 5 to 6 minutes.
Two windows, so both are honest: the dashboard reads 2.1 minutes on Aug 18 alone, and across the whole of month two the median was 1.5 minutes, against 4.6 in month one. Whichever you take, it went the wrong way by roughly two thirds.
This is the one that worries me most, because it is a ratio. Volume can fall because we bought fewer people. A ratio falling means that of the people who did start a run, fewer saw it through.
Share of that week's started runs that reached an ending, won, lost or walked away from.
Some of that is the worse traffic in section 2, and the timing lines up exactly. But the fall also lands right after a run of releases, which makes the other explanation uncomfortable and plausible: we shipped bugs and slowness, and they cost us players. We are now tracking both. The crash reporter went live on Aug 15 and the speed measurement on Aug 18, so for the whole period this chart covers, we were flying blind on exactly the two things most likely to explain it.
All three failing numbers point at the same fix. A run has to be worth more of your evening, and it has to leave you with a reason to open the game tomorrow.
So the next season extends the loop itself, so there is more to do inside a run rather than simply more to look at. Longer, better runs are the mechanism for a longer session, and a longer session is the only honest route we have to day 1 and day 7. Then a tuning pass: performance first, then balance.
Season V shipped on Aug 9 and it is the biggest single thing we have ever released. Before it, a run passed through three places. Now there are six, and which ones you get is drawn per run, so two descents down the same map no longer look the same.
| What | Before | After | What it means when you play |
|---|---|---|---|
| Places to descend through | 3 | 6 | Your run draws which ones it passes. Two runs stop looking alike |
| Monsters | 29 | 47 | 18 new ones, each with its own art, its own sound and its own way of hurting you |
| Afflictions | — | +9 | Nine new ways to be hurt, built on three shared rules rather than nine special cases |
| Decks to play | 4 | 8 | Hourglass, Dusk, Cairn and Poise. Cairn is free, the other three are sealed |
| Boss ladder | 11 gatekeepers | The victory screen now names who is waiting on the next rung | |
Three of the four new decks start locked. You open one by winning a run with a deck you have not won with before. There is no purchase, no key, no currency and no waiting timer anywhere in that sentence. The picker tells you which decks are eligible and how many are left.
Thirty-two releases went out in month two, from v3.1.0 to v5.4.0, on 605 commits. The full plain-English changelog for every one of them is on the devlog, and the Season V write-up is here.
The biggest single place people quit was the first screen we showed them. So we removed it. This took three rounds of testing and most of a month, and the result was not the one we expected.
We tried dropping brand new players straight onto the path picker, straight through a door, and straight into a fight with a hand already dealt. The fight arm won the headline metric by a mile, and we refused to ship it. Dealing you a hand on arrival means the treatment itself arranges the tap it is being measured on. Downstream, none of that lead survived: reaching floor two, finishing a run and winning a run were all statistically flat.
The second round was stopped by hand at 7% of the sample it had committed to needing, because we changed the question. It reads as NOT DECIDED in our records and always will. A test stopped early is not a result, and pretending otherwise is how you learn things that are not true.
4,005 brand new devices, split between landing on the path picker and landing in a fight. On reaching floor two the two tied. On finishing a run, the picker won: 4.20% against 2.82%, which is a real difference and not noise. The fight arm won “played a card” by 10.7 points and that lead bought nothing.
Every brand new player now opens the game on the path choice. Pick Blood, Flame, Frost, Storm or Cairn, and you are in. If you already play, nothing changed, and a run you left half-finished still greets you with Continue Run before anything else.
Two unglamorous things landed this month that will pay off for the rest of the project.
An audit of every file we ship found several megabytes being downloaded that no player could ever see. Nothing visible changed. Not one image was re-compressed or lowered in quality.
candidates-s5. The test written to catch exactly this had the same blind spot.We also built proper measurement of how fast every screen appears on real phones on real connections, and the first honest answer is that all four main screens are over budget, with the map the worst at four times its target. That is now the headline item for month three, and we would not have known without building the instrument.
Until Aug 15, if the game broke on your phone, we found out only if you told us. It now files a report by itself, with no personal data in it, and everything it sends is named on the terms page.
It caught its first real bug almost immediately: dragging the sound slider to zero crashed the game rather than muting it. Fixed the same day.
Month one had a game and no audience. Month two has the beginnings of one, on three surfaces.
31 comments across 7 communities in 14 posting days. 29 are still standing, 2 were removed, and 16 of them got a reply. Everything is written with tooling that reads the thread first and often answers “do not comment here”, but a human presses send, because that is what Reddit's rules ask for.
Three channels, deliberately. An empty server with fifteen channels reads as abandoned. It opened on Jul 20 and it is where the season announcements land first.
Accounts arrived on Aug 4. They are entirely optional, nothing is locked behind one, and signing out leaves everything on your device untouched. All they do is let your climb follow you to another phone.
80 people made one in the first fifteen days. You also choose the name you go by rather than inheriting whatever your sign-in provider had on file, which took four separate releases to get completely right.
Season V, the eighteen monsters, the nine afflictions and the four decks were not designed one at a time by hand.
There is a machine behind them: a loop that proposes a season, plays it against a simulator tens of thousands of times, measures whether the result is actually fun by rules we agreed in advance, and iterates until it converges or gives up honestly.
That machine is the reason a two-person-scale project can double its world in a month and still hold a win rate inside a target band. It is also the one part of this project we are not going to write a blog post about.
Everything else here is open. This is not.
Four things, in order.
We are going to run a few more ad rounds, and for the first time we will let the ad platform measure conversions properly. We have refused to do that until now.
The next season extends the loop itself, so there is more to do in a run rather than simply more to look at. It is aimed at the three failing numbers: longer runs, longer sessions, a reason to come back. Then tuning: performance first, then balance.
A small release to CrazyGames, and the game going to the App Store. Same game, more places to find it.
The long bet is an engine that designs, measures and improves this game largely on its own, with a human holding the wheel. The seasons machine is the first piece of it working.
Measuring ad conversions properly means a tracking pixel, and a tracking pixel means a cookie consent banner on the game. A banner is a wall between a stranger and the door, and some people will turn around at it. We expect to lose players.
We are doing it anyway, and the reason is narrow: we do not need this to sell anything. We need it to know which people stay. Right now we can see that our traffic quality collapsed and recovered, but we are reading that from the inside, after the fact. Handing the ad platform the outcome lets us buy for the players who play for a long time, rather than for the players who are cheap.
Deeper data on fewer, better players beats shallow data on more of them. If it turns out the banner costs more than the targeting is worth, we will say so here and take it back out.