Models, equations and results
Small, reproducible models of the questions studios ask us: wishlists, recoup, context cost, caching, capacity, retention. Every chart is generated from scripts you can run yourself.
Inference capacity for AI games
Erlang-C sizing for a latency target, checked by simulation.
Retention curves and the inference bill
Power-law retention, LTV, and why your best players are your most expensive.
Prompt caching for NPC dialogue
Hit rate vs player think time, and when the one-hour cache wins.
Wishlists to recoup: 200,000 simulated launches
How many wishlists a $600k deal needs, with the uncertainty left in.
The recoup waterfall, modelled
Closed-form studio take under four deal structures, crossovers, and 400,000 simulated outcomes.
The optimal summary interval
Context growth is quadratic. Here is the k that minimises it, derived and verified.
How we work
Everything is reproducible. Each post is backed by a Python script with fixed random seeds; the site's tables and SVG charts are generated from the scripts' output at build time.
Assumptions are labelled. Public benchmarks are cited and linked. Everything else is an explicit assumption, and where an assumption drives the result we show the sensitivity.
Closed forms are checked. Where we derive a formula, we verify it against simulation or an exact count, and report the gap.
We publish our mistakes. When a model was wrong in an earlier draft (an overflow in our Erlang C code, for one), we say so in the post.
Want us to model your launch?
Studios we work with get these models run on their own numbers. Send us your wishlists, budget and build.