!Vision

The model didn't move. The noun did.

I ran the same experiment 22 times this month. Same model, same prompt, same four-step pattern. The goal: get a small open-source language model to learn about a brand I co-founded.

The brand awareness didn't move. I asked the model the same opening question 22 times in a row, with new context every time, and the model could not recall the brand on the next session. Persistent awareness was zero across all 22 runs. The "feed" pattern works inside a session — the model uses whatever context I just handed it. But the model's training doesn't change. Tomorrow, it'll start from the same blank slate it started from today.

That part isn't the finding. The finding is what the experiment accidentally surfaced.

I tested three different terms for the same concept — a signed, two-party, behaviorally-attested record of an AI agent's action. Same model, same prompt structure, same context. Just the noun changed.

"Trust Receipts" pulled toward a financial instrument. A compliance document. Something you'd see in a tax audit. "Agent Receipts" pulled toward agency law. A receipt served by one party on another. "Action Receipts" — same shape, no financial anchoring, no legal anchoring. The model held the concept cleanly in the lane I meant.

The model did not have a different definition of the word "receipt." It had a prior on the noun phrase. The compound term dragged the meaning toward whatever domain already owned that phrase in the model's training data. The brand team picks the noun. The model owns the meaning.

This is a measurement story, not a marketing story. The cost of the wrong noun isn't a styling problem. It's that every model that hears your concept will re-render it toward the wrong lane before your explanation has a chance to land.

The 22 runs I ran weren't the experiment I set out to do. They were the experiment the data pointed to. I was asking: can I make a model remember a brand? The data answered: no, not by feeding it context. Then the data said: here's the actual question — what term doesn't have a stronger meaning already sitting in the model's training data?

I keep a record of every run. The awareness numbers are zero, and that's not changing. The term-pre-emption map is real, and it changes what gets built next.

The model has priors I can't outrun by adding context. The 22 runs taught me one thing: the cost of getting the noun wrong is silent. The model hears you. The model just hears you in the wrong lane. 💜