ground-mechanismlisted
Install: claude install-skill attac-t/the-foundry
# Skill: Mechanism
> "The model decides. Code enforces. Never the reverse."
## The Standard
- **Model decides, code enforces**: if a model's output *is* the enforcement, there is no enforcement.
- **Never let a model report its own oracle result**: run the command in code, read the exit code in code.
- **One answer means code**: a single correct answer is computable. A space of acceptable answers is judgment.
- **Replay means code**: billing, audit, security, migrations. A model is not reproducible across versions, and versions change underneath you.
## The Check
Ask yourself:
- One correct answer, or a space of acceptable ones?
- Must this replay identically?
- Is the input space enumerable? A model choosing between four known branches is an expensive, non-deterministic switch statement.
- If this is wrong, will anyone notice?
Silent wrongness with a wide blast radius is code, or a model behind a gate.
## The Surface
Frequency × ambiguity. Not a binary.
| | Low ambiguity | High ambiguity |
|---------------------|-----------------------------------|--------------------------------------|
| **High frequency** | Code. The economics compound. | Model — and promote aggressively. |
| **Low frequency** | Either. Don't over-engineer. | Model. Encoding never pays back. |
## The Migration
The line moves. Both ways.
- **Promote** judgment → code when it recurs identically. It gets cheaper forev