01
Opens and sets up
It picks the language, presses PLAY and creates a new save slot in an isolated profile, never touching the player's saves.
No cost, no AI at the base
The game tests itself
An automatic player opens the valley, follows the story and records every step. AI steps in only when the algorithm gets stuck.
A test player
What the Tests are
It opens the game, picks the language, creates a save slot and follows the story up to the arrival at the farm.
02
Plays like a person
It uses the normal keys to walk, run, talk and interact. It never teleports, injects items or advances quests through code.
03
Follows the story
It chases the current quest up to the arrival at the farm. Chapters 6 and 7 are already playable, but the robot does not walk through them yet.
04
Records everything
It keeps the decisions, timings, routes, blockages and screenshots of every session.
Level 1 · always on
Decides by rules
The local robot reads the game state and picks the next action by rules. No AI, no cost: the same observations give the same choice.
Physics, clock and villagers vary between runs: what repeats is the policy, not every frame.
Quest
The current requirement and who awaits it.
Target
The villager, object or spot to reach.
Inventory
What it carries and the tool in hand.
Position
Where it is and which directions are clear.
Screens and lines
What is open and who is speaking.
Results
What changed after each action.
Levels 2 and 3 · only when needed
When the algorithm cannot solve it
With no progress, in a loop or facing an unreachable target, the robot asks for help: first Jev, then GPT.
Level 2
Own mode
Jev
- Called when
- the deterministic level stalls, repeats itself or does not know the next step.
- Receives
- the game state as text and the latest actions, no images.
- Budget
- a US$ cap per session: default 0.10, authorized maximum 0.50.
Level 3
Available with a key
GPT
- Called when
- Jev cannot unblock the session either.
- Available
- with the OpenAI key configured (default gpt-5.6-luna); the key never shows on screen.
- Afterwards
- once unblocked, the decision returns to the deterministic level.
The access key never enters the game, the arguments, the reports or Git.
The design
The decision ladder
AI steps in only where the algorithm cannot solve it; once unblocked, the decision drops back to the lower level.
- Zero cost the base never calls the API.
- Reproducible same observation, same choice.
- AI as the exception pays only for the stretch that got stuck.
In the development project
A button in the menu, three keys at hand
TEST opens the session; F6, F7 and F8 follow it, take over and end it.
01
The TEST button
In the game menu, it opens the window where you tick the assists (Deterministic, Jev, GPT) and start the session.
02
Budget and time
With Jev or GPT ticked, spending is capped: default US$ 0.10, at most 0.50. The duration is optional.
03
F6 · the panel
The TESTING panel shows who decided, the reason, the quest and the progress bar toward finishing the game. F6 minimizes it.
04
F7 · take over
Takes the game into your hands without ending the session. F7 again hands it back to the robot, which recalculates from the new state.
05
F8 · end
Ends the session in any state, even with a screen open, and reopens the menu with your own profile.
06
Resume
The window shows the last session’s progress (percentage, chapter and actions) and lets you resume it where it stopped.
TEST only appears in the development project, with Python and Godot: the builds for playing do not include the button.
Every session leaves a trail
What each session produces
Report, events, summary and screenshots: timings, routes, decisions and blockages, always in a fresh folder.
01
relatorio.md
Timings, actions, stages, blockages and stretches without progress; updated every 30 seconds and at the end.
02
eventos.jsonl
Every observation, choice, reason and result: the original evidence, from which the report can be rebuilt.
03
resumo.json
Reason for stopping, number of actions, estimated cost and script errors detected.
04
quadro_*.jpg
Periodic screenshots, kept only on the tester's machine.
05
perfil/
The session's disposable saves: the player's own profile is never touched.
06
site panel
Since Oct 15, with a game token, the session summary (time, decisions, calls and cost) goes to the site’s admin panel.
Moves, screenshots and logs stay only on the tester’s machine. Without a token nothing is sent, and a network failure does not interrupt the session.
Session V21 · Oct 7, 2026
A real session
Excerpts from the report of the session that reached the farm courtyard, as far as the robot goes today.
resumo.json
- stop
- implemented_story_completed
- mode
- robot
- decisions
- 110
- estimated_usd
- 0.0
- budget_usd
- 0.1
- gameplay_seconds
- 417.3
- implemented_story_completed
- true
relatorio.md
| Stage | Results (s) | Actions | Advances |
|---|---|---|---|
| abertura | 7.7 / 16.1 | 4 | 0 |
| pedro_fazenda_ida | 69.3 / 418.3 | 88 | 2 |
| zefa_zefa_ervas | 117 / 427.9 | 7 | 2 |
| pedro_fazenda_chegada | 432.6 / 483.4 | 11 | 1 |
110 actions in 483.4 s recorded, zero cost. The report also lists 17 blockages while approaching Pedro, each with its moment and reason.
The panel during a session
Screenshots from the session
In the lower corner, the panel shows how many actions are done and which one is under way.
Before the human playtest
What it is for
It finds early what would block a person and leaves the human playtest for what only a person can notice.
01
Navigation bugs
Blocked paths, doors that will not open and unreachable targets become blockages with a moment and a position.
02
Quests that stall
Stages with many actions and no advance show where the story stops.
03
Regressions
On every build the same policy walks the path again and flags what stopped working.
A session with no suspicions does not approve the game: it complements the automated tests and the human playtest, it does not replace them.