If you read the first post in this series, you know how it started. Not with a grand vision — with me saying something to myself, half to Cask, half to no one:
“If only I could know what he was thinking right then.”
That sentence hung in the air for a moment, and then something clicked. A week later, we had a working system — one that could ask a character a question, generate possible responses, and tell you whether each one felt right or wrong.
This is the story of how that judgment works. And it starts with a weird thing humans do every day without noticing.
You’re doing it right now.
Someone walks into a room. Before they’ve said a word, you’ve already formed an impression. She looks tired. He’s in a bad mood. They seem nervous.
Then you find out she just ran five miles. He got promoted. They’re about to give a speech.
Suddenly, everything you thought flips. But the flip happens after the first impression, not instead of it. Your brain jumps to conclusions so fast that by the time you have real information, the conclusion is already sitting there, fully formed, insisting it was right all along.
This isn’t a bug. It’s the fastest pattern-matching engine on the planet. And it’s the reason why, when you read a novel and a character says something that doesn’t fit, you feel it in your gut before you can explain why.
The question we wanted to answer: can you teach a machine to feel that same gut check?
🎭 The First Experiment: When “Warm” Becomes “Cold”
In 1946, Solomon Asch did something deceptively simple. He gave people a list of traits describing a person: intelligent, skillful, industrious, warm, determined, practical. Form an impression, he said.
Then he changed one word. “Warm” became “cold.” Everything else stayed the same.
The impression didn’t shift slightly. It reorganized. A warm, determined person feels like a leader you’d follow. A cold, determined person feels like a villain in the making. Same list, one word swapped, completely different human being.
What this means when you’re reading: You do this to fictional characters too. Within a few pages, you’ve built a mental model of who they are — not as a list of traits, but as a network where each trait colors every other trait. When a character does something that contradicts that network, you don’t just notice it. You feel it. It’s the same mechanism that makes you trust someone or not trust them in real life, applied to people who don’t exist.
This is the first thing our system tracks, and we call it T — Trait Consistency. It’s the heaviest-weighted factor in the formula, because trait violations are what readers feel first and strongest.
🔄 The Second Experiment: The Coin Flip
In 1967, Edward Jones and Victor Harris asked people to read essays that were either for or against Castro. Then they told some readers: the writer chose which side to argue. They told others: the writer was assigned a side by a coin flip.
Even when people knew the writer had no choice, they still believed the writer genuinely held that view.
This is called the fundamental attribution error — and it’s everywhere. You see someone cut you off in traffic and think “what a jerk.” You don’t think “maybe they’re rushing to the hospital.” Your brain defaults to personality explains behavior even when the situation is the obvious answer.
What this means when you’re reading: You do the same thing with characters. When a character acts, you assume it reflects who they are, not just what the plot requires. That’s why a character doing something that feels “written” — motivated by plot convenience rather than personality — breaks immersion. You can’t articulate why it feels wrong, but your brain already classified it as a trait violation.
Our system tracks this separately as B — Behaviour Consistency, asking not “does this violate a core trait?” but “has this character actually done something like this before?” It’s the second most important factor, because readers remember what characters do, even when they forget what the book says they’re like.
📏 The Third Experiment: The Penguin Problem
In the 1970s, Eleanor Rosch asked people to rate how “bird-like” different birds were.
Robins scored almost perfectly. Penguins? Not so much — even though penguins are taxonomically birds. Rosch called this prototype theory: our brains don’t categorize things by definitions. They compare against a mental average of everything they’ve encountered in that category.
What this means: Sometimes a character response checks out on every conscious level — it doesn’t violate a trait, it’s consistent with past behavior — but it still sounds off. The word choices are wrong. The sentence rhythm doesn’t match. The character sounds like themselves but not quite, like a cover band playing your favorite song.
This is D — Semantic Distance. It’s the subtlest factor and the only one not guessed by an AI. We compute it numerically: turn each response into a mathematical vector, measure how far it sits from the character’s past dialogue in that vector space. If the distance is large, the response is semantically foreign — even if it passes every other test.
🪞 The Fourth Experiment: The Doomsday That Didn’t Happen
In 1957, Leon Festinger joined a doomsday cult. The leader had predicted the world would end on a specific date. Festinger wanted to see what would happen when the prophecy failed.
The world didn’t end. What did the cult members do? They didn’t leave. They doubled down. They convinced themselves the prophecy had been fulfilled in a spiritual sense. The belief got stronger after being proven wrong.
Festinger called this cognitive dissonance — the mental discomfort of holding two contradictory beliefs, and the brain’s desperate drive to resolve it, even if resolution means rewriting reality.
What this means when you’re reading: You catch cognitive dissonance in characters too, even if you don’t know the name for it. A character says “I don’t care about revenge” and then spends three paragraphs describing their revenge plans. It doesn’t matter that each individual sentence sounds like them — the contradiction breaks the spell.
Our system calls this C — Self-Consistency: a check that a response doesn’t contradict itself internally. It’s a surprisingly common failure mode, and once you start looking for it, you see it everywhere — in books, in movies, in conversations with real people.
⚡ The Fifth Experiment: Why You Can’t Stop Thinking About It
In 1994, George Loewenstein proposed that curiosity isn’t a personality trait. It’s a state triggered by an information gap. When your brain detects a gap between what you know and what you could know, it creates a discomfort that feels like an itch. You scratch it by seeking information.
What this means: The most memorable moments in fiction aren’t the most predictable ones. They’re the ones that surprise you — but in a way that, in retrospect, feels inevitable. That’s the sweet spot.
This is P — Surprise / Novelty, and it’s the only factor that works backward. Every other factor penalizes deviation from expected behavior. P rewards it. A surprising but plausible response actually lowers the risk score. Without this, the system would always pick the safest, most boring option — and that’s not how good stories work.
The Formula
Here’s how they fit together:
Each weight reflects how much that kind of “wrongness” matters when you’re reading. T, the trait violation you feel immediately, gets the most weight. P, the delightful surprise that makes a story worth reading, gets the least — but it’s the only one that subtracts from the risk score, because sometimes the best choice a character can make is the one nobody saw coming.
The numbers themselves came from trial and error, calibrated against real readers’ intuition. In a future post, we’ll talk about how the system learns from your feedback and adjusts these weights over time — but for now, the important thing is this:
Five experiments, each about a different way humans judge other humans. Combined into a single question: does this feel right?
This is Part 1 of the NMA Architecture series. Follow along for more.