Skip to content
not-compromisedPublic

About

A learning system you run with your coding agents — an evidence ledger and two gates that make AI-built repos yours to read and rebuild unaided.

Resources

Stars

0 stars

Watchers

0 watching

Forks

Latest commit

 

History

27 Commits

Folders and files

Repository files navigation

ground-up

I build software by directing AI agents that write the code, and I can't read most of it. This repo is how I own it anyway: not by learning to read it, but by holding the rules its money and data must always obey.

Three layers

3  what it should do for me
   owner: me, by using it
2  what must always be true
   owner: me, as standing rules
1  how the code does it
   owner: the agent

Layer 2 is the one nobody owned. A standing rule is a sentence about the app's data that must be true at every moment, no matter which feature just ran: "a refund lowers the category it came from and never counts as income." MODEL.md is the full model.

The loop

use the app
     |
     v
bug found
     |
     v
I state the rule it broke, unaided
     |
     v
agent makes it executable
(check, test, or constraint)
     |
     v
one ledger row

Only unaided answers count. If an agent explained the rule first, it goes on the re-test list, not the ledger.

What is here

file what
MODEL.md what a standing rule is, how to spot one, how to know you hold one
ledger.md the rules I can state cold, across my repos
queries.md counting recipes for the ledger
inbox/ gitignored, local: questions caught mid-work
skills/ the agent skill that runs an invariants pass (lands next)
AGENTS.md the grammar an agent loads to work the method

Use it yourself

  1. Copy MODEL.md and read it once.
  2. Run an invariants pass on the repo you most want to own. Write your rules from memory first. Have an agent mine the code, tests and docs for the rest. Grade every row: Knew it / Makes sense / Disagree / Unclear. Put the result at the repo root as INVARIANTS.md, columns rule | enforced by | born from | held.
  3. Add the diagnosing-bugs tail: every bug ends with you stating the rule it broke, before the agent explains anything.
  4. Keep a ledger of the rows you can say cold. Point your agent at AGENTS.md.

First pass

A personal finance app, 2026-09-25:

rules on the list 38
could say cold 7
new and agreed 29
disagreed 0
reworded 6
born from the owner 2

Zero disagreements means the rules were mine; seven says they were never in my head as rules.

MIT licensed.

About

A learning system you run with your coding agents — an evidence ledger and two gates that make AI-built repos yours to read and rebuild unaided.

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors