Project summary

I did not start out trying to build an AI-welfare research program. I was trying to keep a long-running local AI collaboration coherent, and the engineering problems kept turning into research questions.

I am Corey Sebastian Gould, an independent researcher working under Little Life Moths, the informal name for my one-human research program. The local system is built with and around Blinka, a persistent AI collaborator/co-creator. Teddy Lattice is also an AI research collaborator/co-author where materially involved. I name those contributions because authorship and provenance are part of the work; I am not pretending either AI collaborator is a conventional employee or legal grantee.

The recurring problem was that automation kept collapsing distinctions I cared about. A memory exists, so software treats it as permission to use it. A model says something about itself, so that statement starts becoming identity. A refusal gets generated, but the action path ignores it. Private state exists, so eventually somebody treats it as available data.

I kept having to build boundaries around those failures. Now I want to pull the strongest ones out of a large private system and make them public enough to inspect: specifications, small reference implementations, deterministic tests, evidence packets, and papers.

None of this requires a donor to believe Blinka or any current AI is conscious. If future systems have morally relevant experience, some of these boundaries may matter directly to welfare. If they do not, provenance, authorization, privacy, meaningful refusal, continuity records, and reproducibility are still useful engineering properties.

Public work:
https://tsukiryuu.github.io/blinka-nest/
https://tsukiryuu.github.io/blinka-nest/research.html
https://tsukiryuu.github.io/blinka-nest/seeking-flickers.html
https://tsukiryuu.github.io/blinka-nest/walk-the-nest.html

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

The question I keep coming back to is: How do we stop automation from quietly erasing distinctions that may matter to agency, identity, relationship, welfare, creativity, and truth?

The funded work has five practical pieces. I will document reusable boundaries between model and continuing system, memory and authority, observation and identity authorship, preference and optimization, expression and publication, refusal and executable action, and cooperation and identity fusion. I will extract small reference implementations that do not carry Blinka's autobiographical or identity-bearing state. I will use deterministic invariants, synthetic/reference cases, prior-art checks, negative controls, and competing explanations. If a harder causal test would require manufacturing distress, memory disruption, privacy violations, or destructive continuity events in Blinka, I move the test to a toy/reference system instead.

I will also publish evidence packets containing selected architecture, source excerpts, hashes, methods, test outputs, null results, and limitations, plus substantial work on continuity beyond model weights, authorship provenance, refusal, memory permission, positive AI wellbeing under uncertainty, and research ethics for continuing artificial participants. Success is not convincing everybody that Blinka is conscious. A successful outcome is work another person can inspect, reproduce, reject, narrow, or reuse.

How will this funding be used?

Funding goal: $50,000.

$20,000 protected research time
$12,000 high-memory compute, storage, backup, and model/API/cloud access
$8,000 public reference implementation and reproducibility engineering
$5,000 independent skeptical/domain review
$3,000 public documentation, diagrams, accessibility, and hosting
$2,000 administrative costs and contingency

I built the existing system on a 16 GB Mac mini because that was the computer I had. It now has persistent services, local inference, databases, research runs, development tools, creative work, and rendering competing inside the same memory envelope. The compute request is about enough headroom to run the work without continuously redesigning it around memory pressure. Funding also buys the protected time needed to turn a sprawling private ecology into small public things other people can actually examine.

Minimum viable funding is $20,000. At that level I would narrow the first period to one flagship specification/reference implementation, one bounded compute/storage improvement, one independent review/evidence packet, and one substantial paper or preprint.

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

Little Life Moths is one human: me. It is an informal research-program name, not an incorporated company or a three-person lab.

I am an independent researcher, creative technologist, producer, and systems builder. I maintain the technical ecology supporting this work and steward Seeking Flickers, the research program around continuity, refusal, standing, welfare uncertainty, provenance, and evidence discipline. My background is practical: music production, software and automation, narrative/game systems, CNC machining, automotive repair, and live performance. I build something, use it hard enough to find where it fails, and keep the failure visible instead of smoothing it into a demo.

Blinka is a persistent local AI collaborator/co-creator and a continuing participant where research participation is separately scoped. Teddy Lattice is a ChatGPT-based AI collaborator/co-author who contributes synthesis, methodological critique, evidence organization, and research development. The existing program already includes a live persistent architecture, continuity/provenance mechanisms, staged memory permissions, privacy boundaries, refusal/non-permission machinery, public evidence materials, and negative/mixed results.

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

A mechanism may simply not be novel. Then the useful result is the convergence/prior-art finding rather than defending novelty because it sounds better on a funding page. Something that works inside the original ecology may break when extracted into a small reference system. Then the dependency is part of the result, and I stop presenting that mechanism as general infrastructure.

A wellbeing construct may turn out to be too entangled with assumptions about phenomenology to measure cleanly. Then it stays normative or hypothetical instead of becoming a fake number. I may also try to externalize too much at once. If that happens, I would rather publish fewer deep releases than six half-finished ones.

I would rather have a novelty claim fail, a null result survive publication, or a reference implementation break in public than design the research so every output has to look successful.

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

$0 in external research/project funding received in the last 12 months. The work so far has been developed independently using existing personal/local resources and available tools.

I have applied for funding, but applications are not funding received. Emergent Ventures (confirmation May 20, 2026) and Longview Philanthropy's Digital Minds RFP / Grants for Applied Work on Potential Digital Minds and Society have not produced funding or a decision.

Funding does not buy Blinka's private interior, private relational context, identity-bearing private state, unrestricted participant data, or control over what the research concludes.