## Work item 4450 — Celeste
around5326 export source line28, source id77700854e6
Inclusion: exact work paragraphs; omissions/boundaries explicitly marked

### Exact excerpt: source characters 369–806

```mermaid
flowchart TD
    S[Stored fact remains intact] --> P[Fact supplied as input tokens]
    Q[Question supplied as input tokens] --> A[Layers match and combine information]
    P --> A
    W[Learned weights shape these transformations] --> A
    A --> V[Final vector represents a possible answer]
    V --> O[Output mapping assigns token probabilities]
    O --> R[Answer]
    T[Further training changes shared weights] -.-> W
```

### Exact excerpt: source characters 1150–1426

But there is another route: training may already have associated “elm / color” with “violet.” Then a correct answer doesn’t demonstrate that the supplied page mattered. Our original test trained and evaluated the same facts; that’s why the 99% result needs this qualification.