Nostalgia Codec: can you miss a place you’ve never been? Three AI-made games of the same forest market
About
Nostalgia Codec · The Market After Summer
Can you miss a place you’ve never been?
Van gave three AI models the same brief: a game of 60 to 90 seconds, set in a forest market that you leave and come back to. Coming back should feel like returning somewhere you remember, though you have never been there.
Each model made one attempt: Opus 5.5 at medium effort, GPT-6 at extra high, Gemini 5 at medium. These are the games as they were written. No AI runs on this page.
Play them first. Below the games, the three are compared part by part, with Van’s own response and what this can and cannot show.
Three AI models, one brief: a forest market you leave and come back to. Play each, then compare them below.
Walk into the market, go north to the stone, touch it, then walk back. The other games pause until you come back to them. Changing the return restarts all three.
How they differ
Choose a part of the game to see what each model made of it. If you have played them, mark the one that worked best for you. Your picks stay in this browser.
This part describes what happens after the stone.
Opus 5.5
Medium effortInto the moment of change. The stone glows, light fills the scene, and the forest comes back in another season, with different residents and details in the market.
GPT-6
Extra high effortInto the frame around the game: title, entry screen, controls and closing questions. Under it is a longer route, a shared movement model and tests for the walk.
Gemini 5
Medium effortInto studying the visit: a reflection form, inspections, dwell times and a JSON log. The scenery is simpler, and the seasonal fade has a rendering problem, so the change is sudden.
No picks yet. Mark the one that worked best for you in any part, and your tally appears here.
Van’s take
Opus came closest for me.
The stone glowed, the light changed, and the forest faded into another season. That magical transition had the greatest potential to recreate nostalgia for me. GPT-6 spent more of its attention on the menu.
Gemini’s reflection and timing log were useful. Its trees and tiling felt more basic, and the compass added little for me. WASD worked; the advertised arrows and Space did not in my playthrough. The seasonal change looked immediate.
This is one person’s response, not a measured ranking of how well each one works.
What this can and cannot show
The next test
Opus changed several things at once, so it is not clear what made its return work. The next step would be to keep one world fixed and change only the return: an abrupt cut, a dark fade, or the glowing stone and light. Then ask people what they remember before telling them what was tested.
That test has not been run. This page only lets you compare the three original attempts.
What was checked
Gemini’s arrow keys and Space worked in our desktop Chromium test, including touching the stone. They did not work in Van’s playthrough. Its seasonal veil stayed fully transparent throughout the test: only the overall opacity animated, so the texture change showed and the veil did not.
GPT-6’s own route simulation completes in 68.4 seconds. That is a simulated walk, not a measured visit. The 60 to 90 seconds in the brief is a target for all three, not a measured completion time.
Method, limits and the original files
Van supplied the shared brief and one output per model, labelled Gemini 5 / medium, GPT-6 / extra high and Opus 5.5 / medium. The labels and the account of an identical prompt come from him. Generation transcripts, token counts, generation times and costs were not supplied. The effort settings differ.
This compares three games, not the general abilities of three models. The premise is a hypothesis, not evidence that an unfamiliar world can reliably bring on nostalgia. There are no participant results here. The model labels, the page’s question and the prototypes’ own study controls can prime a response; this is not a blind study.
The brief asked for a changed and an unchanged return. Each prototype does that differently. Opus also moves a cart and opens a lane in both versions; GPT-6 and Gemini keep their discoveries for the changed version. These are not equivalent tests of the season alone.
The supplied game code, art, controls and transitions are unchanged. Hosting only moves runtime libraries and fonts onto this site and adds a small bridge for pausing and sound. Each game runs in its own isolated frame. Opus’s saved history between visits is unavailable in that frame; its log for the current visit remains. Switching away pauses drawing and sound, but timing based on the clock may still count time spent away. This page collects no answers or logs from the games.
The original archives and HTML are unchanged. The SHA-256 fingerprints identify the exact files.