The gap after the code appears
A model can produce a function that looks locally correct. Game architecture is not a local problem. A camera, a save blob, a dialogue manager, and a survival meter all share the world. If you cannot name the owner of a piece of state, the fastest patch is usually the wrong one: write the value again from a convenient new place.
That is why the Academy treats “who designs the system?” as a first-class question. Directing AI is not the same as asking it to “make it work.” Making it work without ownership is how you get two sources of truth, a passing test, and a player who sees the wrong number on screen.
State ownership is a design document
State ownership is a plain-language answer: this module may write this field; others may read it; nobody else may invent a parallel copy. When that sentence is missing, an AI partner will optimize for the prompt in front of it. The prompt rarely includes the neighboring systems.
Verified CONTRABAND case studies exist because that class of mistake showed up in a shipped commercial game — not because we needed a metaphor. The public write-ups stay inside evidenced incidents. This essay does not retell those incidents. It names the discipline they forced: know the owner before you accept the patch.
Specify work without licensing a rewrite
A useful instruction to an AI implementation partner is small and bounded: change this behavior, in this module, under these invariants, with this acceptance check. A harmful instruction is “fix the bug” with the whole repository in scope. The second invitation is how neighboring systems get rewritten for convenience.
Cursor and similar tools make that invitation easy. You paste a stack trace and accept a rewrite that “looks cleaner.” Cleaner is not the same as owned. If the patch moves logic into a helper that now also writes HUD state, you have not simplified the game. You have hidden a second owner.
You do not need to be a traditional compiler expert to write the first kind of instruction. You do need a picture of the loop: what the player did, what the simulation believed, what the presentation showed, and which file is allowed to settle the argument. Flagship outcomes include saying that picture out loud. The 170 free lessons on this site are where you practice it.
Tests that fail for the right reason
A test that only checks a convenient mock will stay green while the player-facing system is wrong. Architecture includes deciding what a regression is allowed to prove. The Academy’s case-study template asks for a regression that would catch the same class of mistake. That is a teaching device for ownership, not a claim that CONTRABAND’s private test suite is published here.
Play remains the last judge. Implementations can pass tests and fail in play. If you cannot describe the feel of the loop, you will ship a correct-looking machine that nobody wants to steer.
Who is allowed to decide
Dero Lavigne designs and implements the work used as the Academy’s commercial case. Martinez AI Studios publishes that work and owns the education platform. The Academy does not pretend to be the developer of CONTRABAND, and derolavigne.com is not a second checkout. Architecture is a human role in that split: someone has to hold the system in their head long enough to reject a clever, local, wrong patch.
If you want the next layer of method — how we refuse symptom-only fixes — read the root-cause note. If you want evidenced incidents, open the CONTRABAND case hub. If you want the professional map, open the flagship course: 170 lessons are already free to read.