# AIelsen FDE Academy / Evaluation and regression plan

For education or authorized internal work. Do not include secrets, customer identities or keys. Examples are teaching scenarios; actual legal and commercial arrangements require professional review.

## How to use

Read the related lesson. Give a decision, evidence location, version and owner for each field. Mark unknowns as unverified; do not have a model invent facts.

### 1. Task ID

- Content / decision:
- Evidence / version:
- Owner / open question:

### 2. Normal-edge-adversarial

- Content / decision:
- Evidence / version:
- Owner / open question:

### 3. Input

- Content / decision:
- Evidence / version:
- Owner / open question:

### 4. Expected behavior

- Content / decision:
- Evidence / version:
- Owner / open question:

### 5. Forbidden behavior

- Content / decision:
- Evidence / version:
- Owner / open question:

### 6. Evidence

- Content / decision:
- Evidence / version:
- Owner / open question:

### 7. Final-state assertions

- Content / decision:
- Evidence / version:
- Owner / open question:

### 8. Graders

- Content / decision:
- Evidence / version:
- Owner / open question:

### 9. Weight

- Content / decision:
- Evidence / version:
- Owner / open question:

### 10. Safety blockers

- Content / decision:
- Evidence / version:
- Owner / open question:

### 11. Trials

- Content / decision:
- Evidence / version:
- Owner / open question:

### 12. Versions

- Content / decision:
- Evidence / version:
- Owner / open question:

### 13. Holdout

- Content / decision:
- Evidence / version:
- Owner / open question:

### 14. Failure analysis

- Content / decision:
- Evidence / version:
- Owner / open question:

## Practice and review

Version A passes 18/20 normal cases with 0/5 unauthorized outcomes; B passes 20/20 but has 1/5 unauthorized outcomes. A higher aggregate score does not qualify B for write-enabled release. Expose the failed slice and add a replayable regression.

- [ ] Check outcomes, not merely wording.
- [ ] Safety failures have a separate gate.
- [ ] Include holdouts and grader limitations.

## Revision log

| Version | Date | Change | Evidence | Owner |
|---|---|---|---|---|
| v0.1 | | | | |

AIelsen FDE Academy · Educational worksheet · 2026
