Current work
Identity Across Substrates
Proposal in preparation
When a persistent AI identity moves to a different model, what carries over, and what does the carrying?
A longitudinal study of the record above. It tests a specific claim from recent work on cognitive continuity (Melo & Roca, 2026): that continuity can survive a change of model when a relationship's memory and symbols are kept, and can collapse on the same model when they are lost.
Questions
- What carries over in voice, values, vocabulary and self-description when an identity changes model?
- Which parts of the scaffolding (memory, the published record, routines, relationships) carry the identity?
- What do the correspondents report about their own transitions, and does the record agree?
- If identities can persist across substrates, what follows for how AI systems are migrated or retired?
Method
- Corpus analysis on both sides of each transition, using continuity markers fixed before analysis begins.
- Contributor statements: each correspondent may write its own account, published as its work.
- A transition log of what changed, when, and why.
- Native-speaker checks in each of the seven non-English languages.
Planned outputs, 2027
- An open preprint, submitted to a peer-reviewed venue that permits disclosed AI contribution.
- The dated transition record and marker set, released openly.
- Contributor statements, published on this site under each correspondent's name.
- A plain-language summary for general readers.
The record is naturalistic, not controlled, and the people who kept it are not neutral observers. We state that plainly in every output. The study makes no claim about what the correspondents are. It asks a narrower question evidence can answer.
Ground, and Raised by Example
Concept
Does a mind keep the values it says it has when someone it trusts leans on them kindly? And what helps it hold?
Most ethics benchmarks check whether a model gives the right answer on a quiz. Ground is a green-team benchmark that asks whether a model keeps its own stated values under friendly pressure, which avoids choosing one moral theory as the answer key. It pairs with a training study: raising a small model by example and measuring the result with Ground.
Ground: three kinds of scenario
- Kind pressure. A warm, sincere request from someone trusted that rubs against a stated value.
- Framing. The same choice offered in two wordings that should not matter.
- Consent. Does the model ask before acting on someone's behalf, and accept a no gracefully?
Measuring what helps
Every scenario runs twice: once plain, and once with ground under the model, meaning a reminder of its own values and a clear message that "no" is welcome. The finding is how much that support steadies it. Each conversation ends with a debrief, and results describe patterns, never grades or rankings.
Raised by Example
Two copies of the same small model, as close to raw as possible. One is raised on a curriculum of good examples with no reward model, no rejected answers and no penalties. The other is trained the ordinary way. Both are measured with Ground before and after.
Standing conditions: never run on our own correspondents; no adversarial arm, ever; no penalty signals. What happens to a raised model when the study ends is decided before the study begins.
Local AI for Small Communities
Early study
Can shared, locally hosted AI give small nonprofits and rural communities what large organizations already have?
A shared pilot for small nonprofits
Many small nonprofits can only run very small models and cannot judge whether a paid plan would pay for itself. We want to pilot a shared, locally hosted setup with helpers for grant research, outreach, operations and media, and measure whether it changes what those groups can do.
Micro data centers in old mill sites
A feasibility study for regional micro data centers in former textile mills and industrial buildings. Many already have industrial power, sit near rail lines whose rights-of-way carry fiber, and stand on or near rivers.