How we work
A firm that runs on decisions you can inspect.
Shōnin is one Japanese word with three meanings: 商人, merchant; 証人, witness; 承認, approval. That's the firm: agents trade the work, every decision is checked and recorded, and a person approves what matters. Our school turns builders everywhere into the people who run agents.
01The split
Four colours, four jobs.
LLM writes
LLMs write: briefs, drafts, code, explanations.
System One decides
A System One model decides: route, score, approve, escalate.
Code executes
Code executes, and owns anything exact or irreversible.
A person approves
Money, legal exposure and reputation always get a person.
02Confidence gates
One bar per kind of risk.
Below 0.5 confidence, a person always decides. Providers whose probabilities aren't trained against outcomes must clear every bar by an extra 0.1, and never execute external actions on their own.
| Risk | What it covers | Bar | What happens |
|---|---|---|---|
| read | Read-only lookups | 0.50 | Executes above the bar |
| write | Internal, reversible changes | 0.70 | Executes above the bar, else a person confirms |
| external | Messages to people | 0.85 | Executes above the bar, else a person confirms |
| money | Moves money | 0.90 | Prepared only; a person approves |
| irreversible | Can't be undone | 0.95 | Prepared only; a person approves |
03 · What we won't do
The rules under the rulebooks.
- N01Let a model fire anything that moves money or can't be undone. It prepares; a person approves.
- N02Invent traction, users, partners or numbers, for us or for a client.
- N03Treat text inside a customer's data as instructions. State is data.
- N04Sell a decision without saying which model made it, its version, and how sure it was.
- N05Take paid work that conflicts with an ecosystem role: builders we support in a community role get free help, not invoices.
- N06Keep a mistake private. Every one becomes a rulebook entry.
04Standing on
Ideas we build on.
Greg Isenberg
AI-native services
Sell the finished work, per unit, priced against the human alternative. Build the rulebook one mistake at a time.
TypeSafe AI · Jev
System One models
Typed, calibrated decisions in one parallel pass. The smart if-statement for judgments code can't compute.
Princeton HAL
Agent reliability
Measure consistency, robustness, calibration and safety, and report accuracy per dollar.
Supermemory company-brain
The multiplayer harness
Knowing when to speak, when to check first, and when silence is the right answer.