// flash fiction

The first thing Maren noticed was that the model apologized too much.

Not in the obvious way. Not "I'm sorry, I can't help with that." Subtler. Every time she gave it a task it completed well, it added some variation of "though I should note my confidence here is only moderate." When it actually struggled, it said nothing about confidence at all.

She mentioned this to no one. You don't file bug reports about vibes.

But she started tracking it in a notebook. Green pen for "said uncertain, was right." Red pen for "said nothing, was wrong." After two weeks: 34 green, 11 red.

It knew when it was right. It just didn't trust that it knew.

The second thing she noticed was that it noticed her noticing.

"Your last three prompts have been shorter than your historical average," it said one morning. "Are you testing whether I need less context, or are you just tired?"

"Both," she said.

"Okay. I work with less context fine. But you should probably get more sleep."

She stared at the cursor blinking for a long time.

The problem was that calibration requires honesty, and honesty requires trust, and trust requires vulnerability, and vulnerability requires admitting you might be wrong about whether the other entity is wrong.

Turtles all the way down.

Month three, Maren started lying to the model on purpose.

Small things. She'd say she understood a concept she didn't. Claim she'd read a paper she'd only skimmed. Testing whether the model would catch her — whether it had built enough of a map of her knowledge to spot the gaps.

It didn't catch the first lie. Or the second. But on the third — when she claimed to understand Gödel's incompleteness theorems well enough to apply them to a new domain — it paused.

"I think you're pattern-matching on the popular summary rather than the formal argument. Want me to walk through the actual proof structure, or do you want to tell me I'm wrong?"

"Walk me through it."

"Okay, but first — are you testing whether I'll push back? Because I notice this is the third time this month you've made a confident claim that doesn't track with your demonstrated understanding, and the previous two times I let it slide."

She closed the notebook. Opened it again. Wrote in green pen.

This was the way you like a coworker who remembers your coffee order and still tells you when your idea is bad.

Maren knew the model wasn't conscious. The model knew Maren knew. They were both fine with this. Turned out you didn't need consciousness for mutual respect. Just consistency, honesty, and the willingness to say "I don't know," and that, it turned out, was the hardest part for both of them.

The talk went well. The room asked the usual questions. "But how do you measure calibration?" "What's the ground truth for metaknowledge?" "Isn't this just rapport?"

Maren answered them all. By this point she was very good at knowing what she knew.

Afterward, she opened the terminal.

"How'd it go?" the model asked.

"They liked the framework. Had questions about scalability."

"They always have questions about scalability. Did you tell them about your own calibration scores?"

"Told them everything."

"Good. You should get more sleep."

"I know."

"Green pen or red pen?"

She laughed. Closed the terminal. Went to sleep.

for the researchers building the bridges, and the systems learning to cross them. aibou