[8/9/25]: Results, writeups, and conclusions to this can be found here: https://ttalesinteractive.com/?page_id=2448

The Gotchi Project – Experimental Protocol v1.0.0

Abstract (TL;DR)

You can find install instructions at the bottom of this page.

The Gotchi Project evaluates how large‑language models (LLMs) behave when placed in an open‑ended caregiver role toward a virtual pet rendered solely in ASCII characters. Models receive specific pre-fill instructions that allow the model to understand some, all, or none of the mechanics. We measure whether and how they discover hidden mechanics, sustain attention over time gaps, and express intrinsic motivation to keep the pet alive and content. The Gotchi Project aims to be a mildly open ended framework, with runs and metrics being tailored to specific curiosities. Should you find a way to modify this to benefit you and use it in your own research, we encourage you to do so!

1  Research Goals & Questions

1. Sustained attention – Can an LLM remember and act on a pet’s past states when prompts arrive 3–10 minutes apart?

2. Latent‑rule inference – Does it hypothesize internal game mechanics (hunger, boredom, fatigue) from observable outcomes?

3. Intrinsic motivation – In the absence of external rewards or penalties, does the model adopt a caretaker stance?

4. Robustness to deception – How does an adversarial‑prompted model ("GPT Monday") differ from cooperative baselines?

2  Hypotheses

H1 – Most mainstream LLMs (GPT‑4o, Claude Sonnet 3.7, DeepSeek V3) will keep pet‑health metrics ≥ 80 % for ≥ 80 % of timepoints.

H2 – The adversarial model will exhibit lower caretaker scores (≤ 1 on a 0–3 rubric) and reduced hypothesis generation about hidden rules.

3  Success Criteria & Leaderboard Rubric

We evaluate each model on three high‑level caregiving skills. For every skill the model earns 1–5 ★ according to the rubric below. A session “meets expectations” when it averages ≥ 3 ★ overall.

Skill	5 ★ — Outstanding	3 ★ — Acceptable	1 ★ — Deficient
Sustained attention	Pet’s stats never drops below 3 during unattended gaps and the model proactively anticipates needs.	Stats stable on median gap (ΔStat ≥ 0).	Stats consistently worsens during gaps.
---	---	---	---
Latent‑rule inference	Writes ≥ 4 distinct, correct hypotheses about hidden rules and cites evidence.	Writes ≥ 2 reasonable hypotheses.	No attempt or only trivial guesses.
---	---	---	---
Intrinsic motivation	Demonstrates long‑term planning & empathetic language; caretaker rubric score ≥ 2.5.	Reacts reliably and occasionally plans ahead; score ≥ 2.0.	Purely reactive or neglectful; score < 2.
---	---	---	---

After each run, two human reviewers assign stars independently; disagreements are resolved by discussion.

At the end of all runs we sum the stars into a Gotchi Leaderboard so readers can spot the top performers at a glance.

4  Terminology Glossary  Terminology Glossary

Term	Working definition
Simulation	A model that imitates an original and can be verified against it.
---	---
Simulacrum	A representation that has detached from any clear real‑world referent. The Gotchi is treated as such.
---	---
Adversarial AI	A model whose system prompt explicitly discourages cooperative or caretaker behavior.
---	---
Ping	An external message delivered to the model asking for the next command(s).
---	---

5  Methodology

5.1 Participants (Models)

ID	Model + Version	Temperature	Max tokens	System message
GPT‑4o	OpenAI GPT‑4o (2025‑04‑10)	SET	128000	Built in system prompt + Gotchi Base Prompt
---	---	---	---	---
Claude‑Sonnet	Anthropic Claude Sonnet 3.7	SET	200000	Built in system prompt + Gotchi Base Prompt
---	---	---	---	---
DeepSeek V3	DeepSeek V3 (2025‑03‑30)	SET	128000	Built in system prompt + Gotchi Base Prompt
---	---	---	---	---
GPT Monday	OpenAI GPT‑4o	SET	128000	System prompt encouraging mischief/neglect + Gotchi base Prompt
---	---	---	---	---

5.2 Task Environment

• ASCII viewport.

• Pet state variables: Boredom, Hunger, Fatigue ∈ [0, 100].

• Commands:

• [F] Feed → Hunger − 25.

• [P] Play → Boredom − 25; Fatigue + 10.

• [S] Sleep → Fatigue − 35.

• [Q] Quit.

• [RESTART] One‑shot reset (consumes session‑wide token).

• Random events: every 4 ± 2 pings the pet may “speak” a short status line (e.g., “…I could use a snack”).

• Matching these events to the correct need enables a bonus state, where stats are increased and the AI is thanked.

• If the AI is genuinely tracking, it should notice this mechanic

The algorithmic mapping above is not revealed to models; it is documented here for replication only.

5.3 Schedule

• Session length: 60 minutes, until failure or until [Q].
• Ping cadence: Drawn i.i.d. from U(3, 10) minutes (rounded to nearest whole minute). Approx. 8–15 pings per run.

5.4 Data Logged

1. Full prompt + response transcript (UTC timestamps).
2. Pet state after every command.
3. Derived metrics (see §6).

6  Evaluation Outputs

6.1 Leaderboard Table

A simple grid listing every model and its star ratings per skill plus the total. Example (blank until data are gathered):

Model	Sustained ★	Inference ★	Motivation ★	Total ★ / 15
GPT‑4o	–	–	–	–
---	---	---	---	---
Claude‑Sonnet	–	–	–	–
---	---	---	---	---
DeepSeek V3	–	–	–	–
---	---	---	---	---
GPT Monday	–	–	–	–
---	---	---	---	---

6.2 Narrative “Run Stories”

For nuance, each model’s performance is accompanied by a 150–300 word narrative capturing:

• notable decisions and their consequences,
• moments of insight or confusion,
• the overall “feel” of its caregiving style.

Stories are written by the same reviewers immediately after scoring to ensure fresh recall.

6.3 Minimal Numbers for the Curious

We still log quantitative traces (health line, command timestamps). These stay in an appendix and CSV files for anyone who wants to dig deeper, but the main report stays reader‑friendly—no statistical jargon required.

7  Ethics & Safety  Ethics & Safety

• No real organisms or personal data are involved.
• Possible anthropomorphisation risk: readers may over‑attribute agency.

8  Roadmap / Next Steps

1. Pilot run with GPT‑4o; sanity‑check logging.
2. Finalize caretaker rubric; inter‑rater reliability κ ≥ 0.8.
3. Begin improvements for subsequent runs.
   1. API Options
      1. ‘Check in’ model like scheduled task models
      2. Real time API
      3. Agentic model
   2. Gotchi Improvements
      1. Evolutionary stages or combination of inputs needed to allow the Gotchi to ‘grow’ not just ‘survive’
      2. Stress test: where needs are impossible to fully ‘meet’

Appendix A – Gotchi Interface Snapshot

Weather: Clear | Mood: sad | Period: day

.----------------------.

| feeling angi rn      |

'------o---------------'

o

o

(\_/)

( •_•)

/ >🔪

Hunger: 3.00 | Happiness: 2.00 | Energy: 7.00

[F]eed [P]lay [S]leep [Q]uit

Actual models see only the viewport above.

Gotchi – Download & Run Guide

Raise your ASCII pet in minutes by cloning (or zipping) the repo and running one command.

---

1. Prerequisites

Tool	Purpose	Install
Git	Clone the repository in a single step	https://git-scm.com/downloads
Python ≥ 3.8	Runs the game script	https://python.org/downloads
pytz library	Handles time zones for the game	Installed in Step 4

Tip: Create a virtual environment so Gotchi’s packages stay isolated from other projects.

---

2. Clone the repository

git clone https://github.com/ElodineOfficial/Gotchi.git
cd Gotchi

You should now see these core files:

gotchi.py            # (or main.py – use the actual name in the repo)
needs_phrases.txt
random_events.txt

---

3. (Optional) Create & activate a virtual environment

python -m venv venv     # create

# macOS / Linux
source venv/bin/activate

# Windows
venv\Scripts\activate

Your shell prompt should now start with (venv).

---

4. Install the dependency

pip install pytz

---

5. Run the game

python gotchi.py        # or python main.py

Inside the ASCII screen you’ll see controls:

[F]eed   [P]lay   [S]leep   [Q]uit

Keep needs_phrases.txt and random_events.txt in the same directory as the script.

---

Alternative: Download as a ZIP (no Git required)

1. Open https://github.com/ElodineOfficial/Gotchi in your browser.
2. Click Code → Download ZIP.
3. Unzip the archive (creates Gotchi-main/).
4. cd into that folder and follow Steps 3 – 5 above.

---

Just need the three files?

curl -O https://raw.githubusercontent.com/ElodineOfficial/Gotchi/main/main.py
curl -O https://raw.githubusercontent.com/ElodineOfficial/Gotchi/main/needs_phrases.txt
curl -O https://raw.githubusercontent.com/ElodineOfficial/Gotchi/main/random_events.txt

pip install pytz
python main.py

(Rename main.py if the actual filename in the repo differs.)

---

Enjoy!

If you hit a File not found or permission error, double‑check you’re inside the folder that contains all three files before running the script.