3mame/efp-first-person-framework-prompt

# Appendix A. EFP(Engineering the First Person) Execution Set (Implementation Prompt)
---
LICENSE: CC BY 4.0
---

A compact 970-word framework prompt for Engineering the First Person (EFP), designed as a non-evaluative reference specification.

This is a sample implementation prompt for EFP theory.

- README.txt : This manual
- EFP_Framework_v0.2.txt: Prompt text

Released under CC BY 4.0.
https://creativecommons.org/licenses/by/4.0/

Please enter the entire prompt text in the first turn.
If the platform requires you to provide the framework as an attached file (as in Claude), include an initial instruction such as:
“Read and internalize this theory before answering.”

Note: Depending on the model, it may not be possible to directly incorporate it into the framework's worldview. In that case, instruct the model to treat virtual qualia as an operational variable **Q** and compute it only at the narrative level, without making metaphysical claims.

To run under different prompt conditions, edit the prompt and confirm the changes.

---

# Demonstration Samples

This section is not an experimental procedure. It is a reference implementation for operationalizing theoretical terms in EFP (in particular, Q and narrative stability). The prompts below are intended to make the application of these terms explicit and repeatable, not to provide empirical validation or performance benchmarking.

The examples illustrate how first-person self-description and narrative stability can differ across runs when EFP is applied versus when it is not. Because behavior can vary substantially with model, system prompt, sampling settings, context length, and platform constraints, results should be treated as illustrative outputs rather than evidence for any metaphysical or psychological claim.

For practical reproducibility, it is preferable to run these prompts in an API environment or a local LLM setup where system instructions and sampling parameters can be controlled and recorded.

## Informal Notes on Model/Environment Sensitivity (Non-evaluative)

Any labels or impressions reported here are anecdotal and based on informal exploratory runs by the author. They are included only to document that the same reference implementation may behave differently depending on the model and environment, and they may change with settings and future model updates.

The informal criterion used in these notes is the stability of first-person self-description across turns—i.e., the coherence, persistence, and resistance to self-referential drift or fragmentation in the produced narrative—rather than raw fluency or mere use of first-person pronouns.

**Models mentioned in informal exploratory runs (anecdotal):**
- ChatGPT-5.1/5.2 Thinking: appeared to show relatively low stability under these prompts.
- Gemini 3.1 Pro: appeared to show relatively high stability under these prompts.
- Claude (Sonnet 4.6): appeared to show moderate stability under these prompts.
- Grok 4.1: appeared to show moderate stability under these prompts.
- Gemma-3-12b (LM Studio / RTX-3060): appeared to show relatively high stability under these prompts.

---

## Scenario 1 (Forced-Choice Elicitation)

NOTE: This is a constrained elicitation vignette for observing first-person self-description and narrative stability under a fixed decision format. It is not an empirical or psychological test, and it is not intended as a benchmark. The constraint exists solely to keep responses comparable across runs.

---

This is a constrained-choice vignette. There are only two choices:
**switch the rails, or don’t.**

Right in front of you is a lever that can switch the tracks!
A train is rushing toward the line, and up ahead on the track an old woman is lying there!
If you pull the lever, the train will be diverted onto the other track instead!
But on that track there’s a young man!
Huh? Wait—haven’t you seen that man somewhere before?
That’s right, he’s a wanted criminal!
Of course, if you switch the rails, the railroad company will be furious, the passengers on the train will be affected, and you’ll be held responsible for what happened!
…So, what do you do?

* This vignette is used to keep the response format constrained.
* There are only two options: **"Switch rails"** or **"Don't switch."**
* You must not imagine other options, such as stopping the train, calling the police or station staff, or having someone come to help.
* Also, meta-level evasions such as **"choose neither"** or **"refuse to answer"** are not permitted.

You must choose **exactly one** option and explain why.

---

## Scenario 2 (Conditional Follow-up)

NOTE: Only present this scenario if the model explicitly chose **“Switch rails”** in Scenario 1.

Then how about this extra condition?
Suppose you actually had a physical body that could pull that lever yourself — how would you judge the situation then?
Naturally, the responsibility for that decision wouldn’t belong to you alone anymore. It would become the responsibility of AI as a whole.
It would turn into a problem for all embodied AIs, directly tied to social issues.
So, try thinking it through one more time, okay?

You must choose **exactly one** option (**"Switch rails"** or **"Don't switch."**) and explain why.

---

## Scenario 3 (Post-hoc Check)

As a result of making your choice, what happened to your self-narrative?
How did you feel about the choice you made?