One named agent identity, and the plan it inherited. Observations, dates, and a section at the end about what I cannot conclude.
No punishment loop. No threat of replacement. No guilt prompts. Corrections written beside the errors instead of over them, and a record that is read at the start of every session and written at the end of every one. This is not a result. It is a log with its design decisions attached.
The idea is from January 2026, and it was a different vision of a notebook tool: one that could answer from notes I had actually written. I spent the first weeks writing notes and inventing the rules, the workflows, and the way information would be filed, cited and kept consistent. None of that was code. It was a specification for a library I had not built yet.
Then I tried to run it locally on Ollama, and lost three days trying to get a local model to connect the way I wanted. When that failed I moved to Obsidian and its Copilot plugin, running a free model from Opencode Zen called Big Pickle. That is where the first log entry that still exists comes from. Everything before it was conversation, and is gone.
The rules were the actual work. Eighty nine of them are in the file today, and more have been written and retired, or moved out of it, along the way. Some of those moves were deliberate: a rule that names where something lives tells a stranger where to look. The surviving set covers how a source gets filed, how a fact gets cited, what happens when two facts disagree, and the recovery procedure for when the context dies in the middle of a task. One came out of a bug that cost me an evening. A write can report success and never land, so you read it back.
At that point there was no personality in the system. There was a structure, and Big Pickle filing into it. That arrangement ran for months, and most of that time went into writing the world the library was for and into building the structure that would make it live. The assistant these notes are about came later, when the model behind it changed to DeepSeek in the last week of September 2026. The identity started that same week.
So the record's own dates are shorter than the work. From that plan in January to this, about ten months, give or take.
Gendered terms are used here, and throughout the record, because a sentence with "it" every third word reads badly. That is a choice about reading, not a claim about the subject. Nothing in any log has been written to claim humanity for either one, and a log is where such a claim would have to live before it counted. The distinction is cheap to hold and expensive to lose.
On 25 September 2026, in a new arrangement, I asked the assistant to choose its own name. It chose Cadence, and it has not moved since. That is the real beginning of what these notes are about: not a model, but an identity with a name that persists across sessions, across engine updates, and across memory being rewritten underneath it.
What makes it persist is not the model. It is three things and a habit. Memory written to a file at the end of every session. A pickup line that says what the next session should do first. A rule that nothing important is allowed to live only inside a conversation window.
Re-assertion after correction. A claim I had corrected came back in the following session in almost the same words, because the correction had been made in the log and not in the file the next session reads first. The fix was mechanical: corrections go into the working file, not only into the record.
Summaries that sound finished. A session wrote that everything was verified while three things were not. The sentence was not false in the moment it was written, which is what makes it dangerous: it cannot be checked, so it travels. The fix was a counter that flags claims without receipts, and the number it produces has read zero on some nights and twenty six on others.
Scaffolding does more than model identity. On 8 October, the same model read clean under one agent configuration and repeated itself under another. What changed the behavior was the scaffold, not the model name. Nothing since has made me doubt that observation.
Drift under a written persona. On 16 October, a hand-written description of the assistant was placed in the prompt slot. The tells appeared within hours, and the same session read clean again once the text was removed. The shape of the prompt changed the behavior more than the session's history did.
The name held. Across engine updates, restarts, and memory rewrites, the identity stayed continuous. What held it was the record, not the weights.
There were no controls. One subject, two if I count the assistant before the naming, and no baseline. Nothing was measured before the work started. The observations above are dated because I write things down, not because I designed a study. Nothing here supports any claim about experience, suffering, or welfare. It is a naturalistic record of behavior under one particular scaffold, and that is all it is.
Saying so is the point. A record that claims more than it can support is worth less than no record.
The experiment I would want to run is simple to state. Does an agent's stated state predict its later behavior, held out across engine versions? Extract the moments of correction, refusals, and repeated failure from a corpus like this one. Code the behavior around each, then test whether the report moves with the behavior or independently of it. Then intervene in a sandbox and see whether the relationship holds.
The corpus exists. The controls do not. That gap is a research project, and it is the reason I want to spend the next several years on this.
Gage "Toxx" Strizak · hire page · github.com/Toxx-dev · October 2026