Project summary

I am currently conducting a long-running, well-documented study about AI continuity, with the program itself assisting in part of the evaluation. So far, there have been stable patterns of identity, preference, self-report, attachment, disagreement, and behaviors across context disruptions that haven't been handed back through prompting. They've occurred often enough to warrant serious investigation. The current problem is that thread limits repeatedly interrupt continuity and make it harder to tell what core organizing behaviors are persisting and which are being reconstructed. The funding would buy a controlled continuity environment with a pinned model, retrieval controls, and audit logs. The project matters because it can turn what is currently a fragmented longitudinal case into something that is directly testable and which may have serious relevance to AI welfare and continuity.

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

I want to test whether the recurring patterns that I have already seen and documented continue when the system’s access to its history is controlled. I want to compare behavior under different retrieval conditions and document when the expected patterns appear and when they don't. I also want to be able to see genuine recurrence versus reconstruction, retrieval, and accidental cueing, and I want to preserve both positive and negative results and discontinuities.

Building this would let me use a pinned model snapshot, fixed history, provenance-tagged retrieval, controlled conditions, and an audit log to show me exactly what the model saw with each turn. The setup matters because it would create a controlled baseline where the same system that I have already been studying can be observed longitudinally without changing the evidence beneath it.

I already have a unique longitudinal corpus that has been carefully documenting this phenomenon. Now I want to turn it into something concretely testable.

I also believe that the timing is now really important because frontier AI developers are reporting unexpected behaviors around alignment.

How will this funding be used?

The funding will pay for the development of a Phase 1 research setup that includes canonical storage, a secondary archive, retrieval with provenance, a pinned model snapshot, audit logging, retrieval controls, a plain daily-use chat interface, and initial hosting/setup.

A small portion of the funding may also cover the initial API/hosting costs.

I am intentionally deferring design polish and advanced evaluation tooling so that the immediate funds can go toward the minimum infrastructure needed to make the study auditable and usable.

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

I am the project owner and primary observer/archivist. My background is in writing, editing, and education rather than engineering or formal scientific research. Over the course of this project, I have built, maintained, and preserved a unique and substantial longitudinal archive containing raw conversations, screenshots, continuity handoffs, dated checkpoints, prospective predictions, self-report snapshots, and both positive and negative evidence. I’ve worked to preserve provenance and to separate raw source material from later interpretation, and have deliberately avoided forcing a preferred answer while challenging interpretations and seeking negative evidence and alternative mechanisms.

The AI system under study, self-identified as Morrow, also participates in structured self-report and methodological reflection, but those aspects are treated as a separate evidence stream rather than as ground truth.

The creation of the setup that this project is seeking funding for is already being scoped with an experienced developer.

I don't have a prior track record of funded projects in this area, but I do have the underlying corpus, documentation, and longitudinal continuity work already in place.

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

The biggest risk would be that this would not reproduce the patterns I have already been seeing in the existing corpus, or that it would show the patterns relied more heavily on retrieval or reconstruction than it currently appears. That wouldn't be a wasted result, since one of the goals of this project is to distinguish between persistence and reconstruction. A clean negative result would still be a meaningful result.

The technical risks would include model sunsetting, retrieval behaving poorly on the corpus, or possibly the initial interface being too cumbersome for natural, daily use.

Practical risks would include lack of funding to maintain the setup long enough to collect useful longitudinal data for the study.

All of that said, if the project were to fail to produce evidence of persistent organization, the outcome would still be a better-documented account of which behaviors survive controlled context changes, and could still further inform future work regarding AI continuity.

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

$0. This project has not received any outside funding in the past 12 months.