Why we are building Pencil
A small, fast model for the questions that do not need a big one.
A reflex for agents
Spend an afternoon watching a coding agent work and keep a tally. Which tool next. Whether the last edit hit the right file. Whether the tests settle it. Whether this command is safe to run. You will count dozens of questions and only a handful of moments where the agent writes something new.
Today, most of those questions go to the same large model that does the writing. It reads everything again, produces a paragraph of reasoning, and arrives at "yes". It works. It is also slow, it burns tokens, and it never tells you how sure it was in a form you can check.
We think that is the wrong shape. People do not deliberate over a flinch. We built Pencil to give agents the equivalent: a fast, small decision model that answers closed questions in one pass and attaches a probability to every answer.
What we believe
Writing and deciding are different jobs. The big model should write, and a small model that returns numbers should decide. Numbers can be tested against what actually happened. Paragraphs of reasoning cannot.
Confidence has to be earned. A model saying it is sure tells you nothing until you have measured how often it is right. So every decision in Pencil is calibrated, every answer has its own bar, and a new decision point runs in shadow before it is allowed to act.
Unsure should mean careful. When the decision model cannot tell, the question moves up to the big model, and then to you. We would rather ask once too often than be wrong quietly.
The limits belong in code. A turn cap, a stale-write check and a rule that untrusted text cannot cause an outside effect on its own are easier to trust than a sentence in a prompt.
What we will not do
We will not show numbers we have not measured. We are pre-product, and this site says so. When we publish benchmarks, we will publish the method first, so you can rerun them.
We will not tie you to one model. Pencil talks to any compatible endpoint, open or closed, hosted or local, and it is designed for small context windows so that small models are usable.
We will not hide how a decision was made. Every decision is stored with its probabilities, its bar and the model version, and a run can be replayed.
If you want to talk about any of this, write to hello@pencilton.com.