Game State Demos: Check Inputs Before Claiming Vision
A small, inspectable experiment with its inputs, decisions and limitations recorded. This guide is not a production guarantee.
Use caseLimit browser or computer actions to reviewable choices
Learn to inspect exactly what information a decision model receives.
Source · Psyho · @FakePsyho · X companion
2026-09-21 · Editorial update; no live model call
Problem
Section titled “Problem”A discussion of a game demo points out the importance of its text state. This source-informed reading exercise asks what the application exposes before judging a model’s capabilities.
Inspect the observation boundary Visible video → Program observation → Available state → Allowed actions → Next-state check
Inputs determine what a result establishes.
Before you begin
Section titled “Before you begin”-
Use a small fictional or authorized dataset. Do not paste private customer data into an unfamiliar service.
-
A current browser for reading. Check the linked project for its own runtime, access and cost requirements.
1. List the available observations
Section titled “1. List the available observations”Does the model receive pixels, transcribed text, object positions or the full game state? These are different tasks. Do not infer the input from a video of the output.
2. Check what code already knows
Section titled “2. Check what code already knows”Document preprocessing, allowed actions and any hidden helpers. A game engine may provide more information than a human player sees.
3. Change one assumption
Section titled “3. Change one assumption”In a toy exercise, remove one field from a state summary. Compare the decision and note whether the task remains answerable. Do not treat missing information as zero.
4. Keep the conclusion narrow
Section titled “4. Keep the conclusion narrow”A successful game action shows something about that state and action space, not autonomous driving, general vision or broad reliability.
The visible video is not necessarily the model input
List the observations available to the model: pixels, transcribed text, object positions, legal moves or complete state. A video of a game is an output presentation, not proof of the input modality. Mark missing documentation as unknown instead of inferring visual understanding.
Record what the surrounding program already knows. Navigation grids, legal-action lists and objective selection can make the decision much more constrained than human gameplay. These are valid system design choices, but they should be described rather than attributed to the model alone.
In a fictional toy example, remove one state field and ask whether a safe decision remains possible. Missing obstacle information is not the same as an empty path. Waiting or requesting a new observation may be the appropriate outcome.
Keep the conclusion scoped to the provided state, action space and observed result. A successful move does not establish general vision, autonomous driving or broad operational reliability.
| Input | Remaining task |
|---|---|
| Pixels only | Perception + decision |
| OCR text | Incomplete geometry |
| Objects + positions | Choice over known objects |
| Full state + legal moves | Constrained decision |
Conceptual comparison, not a game benchmark.
Result and cautions
Section titled “Result and cautions”A small, inspectable experiment with its inputs, decisions and limitations recorded. This guide is not a production guarantee.
- A valid label can still be wrong. Keep a review path and keep irreversible actions outside the exercise.
- A community demonstration is not an independent benchmark. This collection does not reproduce the linked video or repository.
- You can identify the input, its source, and the fields that leave your system.
- Original identities, failed rows, uncertain cases and human corrections remain visible.
- You distinguish an offline fixture, an author demo and a live evaluation you ran yourself.
Source boundary: compiled from the linked public sources; not reproduced here. Review classifications before acting; they do not run actions automatically.
Related guides
Section titled “Related guides”Does a video prove visual input?
Section titled “Does a video prove visual input?”No. Inspect the actual request.
Is structured state inherently invalid?
Section titled “Is structured state inherently invalid?”No. It defines a different task that should be disclosed.