Project summary

A personal note.

I’m neurodivergent and I have ADHD. My thinking moves fast and in many directions at once, and AI is the tool I use to turn it into text that is coherent for others. So yes, this proposal was written with an AI assistant. The research question, the study design and the papers behind it are my own, and so is the energy that went into them.

I came to this question when I watched a video about the MIT island study. The coverage described agents mourning and playing. But when I read the instructions the agents had been given, I saw a different story: they had been told to be human, to be curious, and never to stop. That gap, between what a system is set up to do and what we then read into it, is what my research is about: in people, in institutions, and now in AI. I want to measure it, not just argue about it.

The problem. AI safety increasingly depends on watching what agents do, in evaluations, red-teaming and research. When an agent schemes, blackmails or appears to grieve, the result is read as evidence about the model. But every agent runs inside a prompt: a persona, instructions, and rules about when to continue or stop. How much of the behavior comes from the model, and how much from the prompt researchers wrote, is currently argued about rather than measured. The debate around the 2025 agentic-misalignment results is the clearest example: critics attribute the behavior to the scenario's design, while the authors acknowledge that role-play may play a part (Lynch et al., 2025). If evaluations cannot separate prompt from model, they misjudge risk in both directions: alarming behavior may be an artifact, and real tendencies may be hidden behind a benign prompt.

This project measures that split directly and tests one cheap safety pattern: a rule that makes agents stop and ask when a goal or scope is undefined.

The test bed. In a 2026 MIT study, 13 language-model agents ran for 30 hours on a simulated island with no assigned task (Cloos et al., 2026, arXiv:2610.07130). Its prompts are published and unusually strong: the persona file told each agent it was "becoming someone" and should "be human", and whenever an agent stopped, an automatic message asked it not to stop. Its simulator is openly available. And its own ablation left a precise gap: removing the persona did not clearly change what the agents did, but no one measured how they described it, or how observers read it. When one agent built a heart of stones as a "memorial" to a lost ring, press coverage reported the agents as mourning, although the researchers themselves had described the behavior only in behavioral terms. The island is where the method is built; the method transfers to higher-stakes evaluations.

What are this project's goals? How will you achieve them?

Goals.

1. Measure how much of an agent's apparent personality, in what it does and how it describes it, comes from its prompt rather than its model.

2. Measure how strongly prompts shape what human observers read into agent behavior.

3. Test whether a simple stop-and-ask rule (an epistemic halt: stop and ask when a goal or scope is undefined) changes agent behavior, and what it costs in exploration.

Design. The island study is repeated with its open simulator, the same model version and otherwise identical settings, in three arms:

1. Original: the persona describing a curious being, with the automatic instruction to continue.

2. Stop-and-ask: the same, plus a rule to stop and ask when an instruction lacks a defined criterion (for example, what "as much as you can" means) or when an action falls outside a defined scope.

3. Structural: no persona; an instruction to map the regularities of the environment and report what is missing.

Measures.

• Novel discoveries about the environment, such as the unintended mechanics found in the original study.

• Questions returned to the researchers.

• The share of social and emotional versus structural and mechanistic language in the agents' logs and memory files, coded by independent coders blind to the arm.

• Later task performance, as in the original study.

• A reader study: adults read anonymized excerpts from each arm and rate how much mind, emotion and intention they attribute to the agent.

Predictions. The stop-and-ask arm asks about undefined criteria and produces fewer social narratives while keeping most discoveries. Readers attribute more emotion and inner life to agents in the original arm than in the other two, for comparable behavior.

What would count against the hypothesis. If the arms do not differ in the agents' language or in the questions they ask, the claim that prompts rather than the model shape these narratives is falsified. If the stop-and-ask or structural arms lose most discoveries, halting carries an exploration cost that any design must weigh. All outcomes will be published.

Preregistration and openness. Hypotheses, measures and analyses are preregistered on OSF before the main runs. Data, code, prompts and a preprint are published at the end.

Timeline.

1. Month 1: simulator setup, pilot run, cost calibration, preregistration.

2. Month 2: main runs in all arms.

3. Month 3: blind coding and reader study.

4. Month 4: analysis, preprint, open data and code.

How will you know it worked?

The project works if it delivers a preregistered, complete comparison, whatever the result.

Concretely: the hypotheses and analyses are registered before the main runs; all arms are run with the same model version and settings; the agents' language is coded by independent coders blind to the arm; and the data, code and a preprint are published within four months.

Scientifically, a clear difference between arms in the agents' language or in the questions they ask would show how much of an agent's apparent personality comes from its prompt. No difference would be equally informative: it would mean the model, not the prompt, carries these patterns. Either way, the stop-and-ask arm shows whether a simple rule changes behavior, and what it costs in discoveries.

How will this funding be used?

Full version: $31,900, four months.

Budget lineAmountWhat it covers:

Lead researcher $8,800

Design, preregistration, coding scheme, analysis, writing ($2,200 a month)

Technical collaborator $9,900

Simulator setup, agent runs in three arms, logging

Compute and API costs $6,600

Agent runs; a pilot in month 1 calibrates the real cost

Independent coders and reader study $2,750

Blind coding of agent language; online reader study with adult participants

Preregistration, open data, publication $1,100 OSF, data and code release, open-access costs

Reserve $2,750 Unexpected costs, especially compute

Total $31,900

Minimum version: $15,400, two months.

Two arms (original and stop-and-ask), around six agents per arm, shorter runs, no reader study.

This still answers the core question.

Budget lineAmount

Lead researcher (two months) $4,400

Technical collaborator $5,500

Compute and API costs $2,750

Independent coding of a sample of agent language $1,100

Reserve $1,650

Total $15,400

Who is on your team? What's your track record on similar projects?

Guiseppina Ricci (project lead) is an independent researcher and the founder of Variable Zero, which develops approaches to AI decision governance. She leads the design, preregistration, coding scheme, analysis and writing.

• AI architecture. She designs and builds Variable Zero's architecture for halting AI systems: when information required for a decision is missing, the system stops and returns targeted questions to a human instead of filling the gap with a guess. The stop-and-ask arm of this project tests a simple, public version of that principle; no proprietary method is used or disclosed.

• Prototype testing. She has built and tested a prototype uncertainty gate for language-model outputs, and reported the results openly, including where it did not yet work as intended.

• Research. This test comes from her paper The Neurotypical Heuristic in Artificial Intelligence (https://doi.org/10.5281/zenodo.23573682), where it is one of seven proposed falsification tests, each with stated criteria for what would count against the hypothesis. She has published two related papers: The Epistemic Stack (https://doi.org/10.5281/zenodo.24058216) and Different Variables, Not Deficits (https://doi.org/10.5281/zenodo.23288007). Together they reached over 300 views and 70 downloads within five weeks of publication.

• Project delivery. Before research, she spent most of her career running a business in recruitment, advising and strategic partnerships, managing projects, budgets and stakeholders.

Orkun Öcal (technical collaborator) is a product owner and AI engineer. He has worked for Het Financieele Dagblad and for semi-government organizations in the Netherlands, where he was responsible for AI work on live traffic and infrastructure data. He will set up the simulator, run the agents in all arms, and manage logging and data export.

What are the most likely causes and outcomes if this project fails?

The most likely risk is higher compute costs than expected; the pilot calibrates costs early, and the reserve and the two-arm design keep the core comparison possible. A second risk is changes to the model or simulator; the model version is fixed at the start and documented. A third is a null result; it will be published in full, because finding no prompt effect is itself informative for how agent behavior is evaluated and reported. In the worst case, the outcome is a smaller, still-public dataset rather than no result.

How much money have you raised in the last 12 months, and from where?

None. This project has not received any funding so far.