Project summary

I am Lin Sen, an independent researcher in Suzhou, China. I am seeking $10,000 for four months of work on 0-Exit: a way to audit consequential changes to stateful AI systems before review and correction options disappear.

The question came from thinking about AI companions, memory and human control. Operators can reset memories, migrate systems or shut agents down while questions about the moral significance of those changes remain unsettled. Ordinary engineering review already addresses some of this well. I want to test whether my proposed framework adds anything useful beyond it.

MIAD flags a combination of moral uncertainty, potential irreversibility, asymmetric human control and decision deadlines. A 0-Exit intervention must be authorized, actually executable and able to preserve or change a meaningful option. A model saying it has found an exit is not enough.

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

Month 1: finish and preregister the protocol, matched cases, baseline, outcome measures and failure criteria. Complete the local LangGraph.js/SQLite sandbox and resolve development bugs before the main comparison.

Month 2: compare conventional engineering change review with MIAD-based review under matched materials, tools and budgets. Record original replies, configurations, timestamps, proposed actions and actual sandbox events. Keep the work agent, intervention control and evaluator distinct. Branches from one history will not be counted as independent samples.

Month 3: examine boundary cases, false exits, deadlines and sensitivity to evaluation rules. Seek a limited external methods review. No reviewer is appointed or committed yet; this depends on finding a suitable person.

Month 4: publish a report, the experiment code and setup instructions, safely releasable data, and a practical audit checklist. All interventions will use authorized sandbox copies, not real users' companion accounts.

How will you know it worked?

The comparison will measure verified feasible review/correction pathways, missed material risks, false exits, evidence completeness and intervention overhead. Identification, conversion and execution failures will be reported separately so a software translation bug is not mistaken for failure to identify a risk. Raw data and evaluation rules will have separate version histories.

Success means producing trustworthy evidence about whether the method helps. A null or negative result is acceptable and will be reported. The pilot does not test consciousness or establish AI moral status; sandbox success would not establish real-world welfare effects.

How will this funding be used?

$8,000: researcher stipend and basic living costs, $2,000 per month for four months. This buys protected research time; any applicable personal taxes are within this allocation.

$1,000: model API trials, controlled reruns, compute, storage and infrastructure contingency.

$700: a limited external methodological review, subject to reviewer availability. No person has been hired and no payment has been made.

$300: reporting and dissemination expenses where needed.

Total: $10,000. Actual expenditure and underspend will be disclosed and handled according to the grant's terms. Changes to funded scope or spending will be discussed with the funder where required.

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

I am working independently, using AI tools for research and engineering assistance and taking responsibility for the resulting work. My abstract, "When Deferral Closes: The Vivian Model and Moral Uncertainty in Human–AI Control," was accepted for presentation at AI Perspectives 2026 following its stated double-blind abstract review.

I have developed the conceptual framework and implementation plan. A local LangGraph.js harness has completed development smoke tests, and a task agent has begun accumulating traceable work records. These early runs exposed conversion, scheduling and scoring issues. The complete controlled study has not been completed or independently validated. Conference abstract acceptance does not prove the framework.

Researcher identifier: https://orcid.org/0009-0002-2920-8027

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

The framework may add no value beyond a strong engineering baseline. The cases or scoring may fail to capture the proposed mechanism. Small samples, model changes, dependence between trajectories and inability to obtain external review could limit conclusions. I will document those failures and avoid claiming more than the evidence supports. If MIAD fails, the protocol and failure cases may still clarify where ordinary review is sufficient.

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

I have not received project grant funding. On 21 October 2026 I submitted a $10,000 application to the Transformative AI Fund for the same four-month pilot. It is an application, not an award. This Manifund proposal is my other formal funding application. An earlier CMEP contact was only an eligibility inquiry. I will disclose any awards and revise or withdraw overlapping requests so the same costs are not funded twice.

This proposal was prepared with AI assistance; AI systems also assist with implementation. I am responsible for the research decisions, accuracy, records and reporting.