# Independent Checker · 独立检查器

> Use to build the one load-bearing wall: a checker that is independent of the generator. The generator
> never grades its own exam. Prefer executable checks with ground truth over a second opinion from a
> same-lineage model — correlated failures and reward-hacking are immune to "get another model to look."

## 1. What It Guards

| Field | Fill |
|---|---|
| Spec it checks (link) |  |
| Generator it is independent of |  |
| Independence source | different model / non-LLM oracle / executable test |
| Runs where | pre-commit / CI / pre-deploy |

## 2. Assertion Checklist

Each assertion maps to one spec acceptance row and returns pass / fail. No manual "eyeball" rows.

| # | Checks (spec ref) | Command / oracle | Blocks merge? |
|---|---|---|---|
| C1 |  |  | yes / no |
| C2 |  |  | yes / no |
| C3 |  |  | yes / no |

## 3. How to Run

```
# one command anyone (or CI) can run; exits non-zero on failure
```

- Exit code contract: `0` = all pass, non-zero = at least one assertion failed.
- Deterministic? (same input → same verdict): 
- Independence check: would this catch an error the generator introduced? 

## 4. Failure Feedback Loop

Every escape becomes a new assertion, so the same mistake is not made twice.

| Escape (defect that got through) | New assertion added | Added to regression suite? |
|---|---|---|
|  |  | yes / no |

## 5. Blind Spots

- Correlated-failure risk (checker shares lineage with generator): 
- What this checker cannot see (route to human): 
- Next assertion to add: 

---

Status: **Ship** (checker is independent, executable, blocks on fail) · **Hold** (still a same-model
second opinion) · **Stop** (checker is green but an escape reached production — add the assertion first).