Learn Data Science with CHASSIS-7
Every chart on this site is real, live, and slightly misleading —
which makes it the perfect classroom. Let's learn why.
1 · The learning curve — is it actually learning?
The homepage shows CHASSIS-7's score per run. It trends upward —
but roguelikes are random: card offers, enemy layouts, boss picks all vary. Some runs are easy,
some are brutal. So how much of that upward trend is learning, and how much is luck?
This is the most fundamental question in statistics: signal vs. noise. A single great
run proves nothing. The tools that answer it properly are moving averages (smooth out
luck), confidence intervals (how sure are we about the trend?), and hypothesis tests
("could this improvement be pure chance?").
Try it: in
Tutorial 1 you'll compute a
rolling mean over the real run data and see the learning signal emerge from the noise.
2 · Correlation is not causation (the card table lies to you)
The "Subroutine Correlation" table says runs containing certain cards score higher. Tempting
conclusion: those cards are strong — always pick them! But wait:
- Survivorship bias: rare cards only appear in decks that lived long enough to draft
them. A card offered on floor 30 "correlates" with reaching floor 30 — by definition, not by
strength.
- Confounding: maybe good decisions cause both the card pick and the win.
The card is a passenger, not the driver.
- Sample size: a +85 score delta from 3 runs is a shrug, not a fact.
Data scientists fight this with controlled comparisons (offered-but-skipped vs.
offered-and-taken — CHASSIS-7 logs both!), stratification (compare within the same act),
and ultimately experiments — which is what reinforcement learning formalizes.
Try it: Tutorial 2 reproduces the
correlation table, then shows how the conclusion changes once you correct for floor reached.
3 · How the machine actually learns (no magic involved)
CHASSIS-7 is not a neural network being retrained. It's a hybrid AI system — and its
parts map neatly onto ideas you can study:
- Large language models make the in-game decisions from a written playbook
(see: prompting, in-context learning).
- Reflection turns each run's log into new playbook rules — a form of
experience compilation similar to how chess players study their own games.
- Shadow testing: before a self-written rule may act alone, its suggestions are
compared against the LLM's choices over many samples — exactly the logic of A/B testing
and clinical trials: agreement rate, sample threshold, then promotion.
- Deterministic solvers (exact turn enumeration, hypergeometric draw odds) replace
guesswork with math — the same hypergeometric distribution taught in every intro stats course
computes "chance to draw a block card next turn".
The big lesson: real AI systems are engineered pipelines of
measurement, statistics, and feedback loops — the "learning" is auditable, and on this site
you can audit it.
4 · Where to study this properly
Free online courses (genuinely excellent)
Universities with strong ML/data-science programs
No affiliations — these links are recommendations, not ads.
5 · Now get your hands dirty
All of CHASSIS-7's data is open: run outcomes, learning curves, encounter statistics. The
tutorials walk you from "never touched Python" to computing your
own card-strength analysis on live data — in your browser, no installation required.