Know what your program does without reading it.

Yichus is a series of experiments around that one idea.

Your coding agent wrote most of your app, and now you have to trust it. How is it put together? What happens when two requests overlap? Can one user reach another’s data? Today you find out by reading the code, or by asking the agent to summarize its own work.

Yichus answers these questions from the compiled program instead. Its explanations, diagrams and checks are computed, not written by a model, so you can repeat each one and inspect its evidence. Each check says what it established and what it could not decide.

Where it started

Press Why on a number in a data app

Questions like these come up with any program you rely on. Where did this value come from? Which rules produced it? What would change if one input were different? The program already worked out the answers when it ran. What is usually missing is a way to ask it.

That is why the first button above opens a data app. Data apps are one example, and we show them first because we think they are valuable and approachable. When an agent writes a data app, the people using it don’t care about most of the code, but they care a lot about how each result was calculated. They are looking at the output, a number on the screen, and before they rely on it they want to know where it came from. Spreadsheets already let you do this: you can trace any cell back to the formula and cells it uses.

The first experiment gave a data app the same habit: press Why on a result to see the line of the program that computed it, the same line with this run’s values, and the inputs it depends on, taken from the compiled program rather than from anyone’s description of it.

The Why panel for Amount left over, current value $200.00: the inputs it can depend on (Guests 40, Ticket price 25, Venue and fixed costs 600, Cost per guest 5), the source line balance = revenue - total_cost, the same line with values, balance = $1,000.00 - $800.00, and where revenue is Ticket income $1,000.00 and total_cost is Total cost $800.00. Try a data app

The same idea, applied to the app your agent wrote

Each later experiment answers a question you would otherwise answer by reading the code, or by asking the agent.

Could it leak data or overspend? Three experiments look for these problems in the program itself, including on paths your tests never run. Can one user reach another’s data?Can concurrent transfers overspend?What can a model-written program access?
Did the last change break something? Replay a saved example and compare what it read, wrote and returned before and after the change, even when the answer looks the same. Compare saved runs
Can it do less repeated work? Run a prepared calculation in the browser or a worker without preparing it again, and update only the parts of a page that a changed value feeds. Move a calculation without rewriting itUpdate only what changed

Apps made with it

The Problems API, a Bosatsu program on the Yichus engine, serves streams of practice problems. Each stream is a small program, and a model can write new ones. Times-tables and git flashcard apps use it. The time-travel puzzle is a programming game in a few versions. More demonstration apps are on Built with Yichus.

What makes it possible

What your app needs for this

To know your app without reading it, Yichus needs its logic written in Bosatsu, a small typed functional programming language that tools can analyze thoroughly. Your coding agent can port an existing Python, JavaScript or TypeScript app into it, or write a new one in it. That logic stays apart from reusable engines that run the browser, database and network, so tools can read it as well as run it.

Already using an agent? Connect it to this site and ask whether Yichus fits your project.

Ask your agent about anything here; it has the detail.