Llama-Unsealed
A writing model for adoption, from the adoptee's side of the record.
Built by an adoptee, trained on an adoptee's writing · To my knowledge, the first openly published language model of its kind
Llama-Unsealed writes about adoption, forced adoption, sealed records and family separation the way Thoughtless Delineation does: in Honest Adoption Language, naming the system rather than the person, crediting what an institution achieved before naming what it left undone. It is fine-tuned on my published work. Before it writes, it looks things up. After it writes, it checks itself.
Why it exists
General AI models learned about adoption from the people who run it, fund it and celebrate it. Ask them about adoption and you get the rescue story: the "forever family", the "gift", the gratitude. That is Non-Adoptee Bias built into the machine.
Llama-Unsealed is trained the other way round. It learned from the writing of an adult adoptee and forced-adoption survivor, and from the Framework Canon that names how adoption separates a person from their origin and keeps that separation in place. It says "biological", not "birth". It writes as the adoptee, never as the parent. It treats the sealed record as the subject, not the background.
How it works
It looks it up
Every request is matched against a library of more than 2,000 passages: my published essays and books, all 38 framework definitions, and openly licensed sources such as the UN Convention on the Rights of the Child, AIHW adoption statistics and Australian adoption law.
It writes from the record
It writes in the Thoughtless Delineation register from the passages it was given. My own writing always leads: official and legal sources can only ever fill a minority of the passages.
It checks itself
A fact check reads the answer and finds every figure, date, name, law and quotation the passages do not support. In public use each one is replaced with [needs checking] before anyone sees it.
What the fact check looks like in practice:
In [needs checking], the minister wrote to [needs checking] to withdraw the criticism.
Fact check: 2 items the library does not support were replaced with [needs checking].
What it's for
Adoptees
Drafting in your own structural, non-therapeutic register: a submission, a letter, an essay opening, the words for something you have felt and never had language for.
First families
Writing about the other side of the record without the vocabulary of "relinquishment" and "choice" that was used to cover it.
Writers, journalists, advocates
Precise language for harms official vocabulary was built to soften, and a check on the framing before it goes to print.
Researchers
A working example of a voice-preserving, retrieval-grounded model built for a lived-experience community, with its limits documented.
What it is not
Not a source of facts. Even with the library and the check, it can build an invented story around real facts. Treat every answer as a draft in a voice. Check it before you use it.
Not me. It writes in the register of Thoughtless Delineation, but its outputs are not my words or my views, and must not be presented as either.
Not advice. Nothing it writes is legal, clinical, financial or search advice.
Try it
The free demo runs on Hugging Face. Type a request, such as a framework explained, a paragraph on sealed records, or a short critique of an apology, and Llama-Unsealed looks it up, writes it and checks it. It runs on a free computer, so each answer takes about 3 to 5 minutes, and it answers one person at a time. If it has been idle for a couple of days it needs a few minutes to wake up.
Access to the model
The model and its library are on Hugging Face. Access is approved by hand: tell me who you are and what you will use it for. By requesting access you agree to Meta's Llama 3.1 licence and acceptable use policy, to use the fact check, and not to reuse or retrain on my writing.
Questions
What is Llama-Unsealed?
An AI writing model that writes about adoption, forced adoption, sealed records and family separation from the adoptee's side of the record, in Honest Adoption Language. It is fine-tuned on my published work and built with Meta Llama 3.1.
Why "Unsealed"?
Because it writes from the record, and the fact check flags anything the record does not support. Sealed records are where the separation is kept in place; this model works the other way.
Is it a source of facts?
No. It is a writing tool. Every answer is a draft to check, not a source.
Does it speak for Shane Bouel?
No. It writes in the register of the publication, but its outputs are not my words or my views.
What was it trained on?
Only my own published writing and framework definitions. Other people's documents, including inquiries, legislation and statistics, were never used for training; they are only looked up at answer time, and only when their licences allow.
How do I get access?
Anyone can try the free demo. To download the model itself, request access on Hugging Face; requests are approved by hand.