# Student workbook

Use this workbook with `COURSE-TRANSCRIPT.md` and `CASE-FILE.json`. Write short,
specific answers. Prefer exact nouns and measurements over broad adjectives.

## Exercise 1 — Separate the four layers

For the statement below, identify the objective, method, evidence, and
supportable conclusion.

> A team wants to reduce a defined activation-space separation. It applies
> three recaptured rank-1 passes to a declared matrix set. The recorded metric
> falls from 87.0274 to 32.2948 while all defined controls match exactly.

- Objective:
- Method:
- Evidence:
- Supportable conclusion:
- One unsupported conclusion:

## Exercise 2 — Experiment identity checklist

Mark each item as **identity**, **result**, or **neither**.

| Item | Classification | Why? |
|---|---|---|
| Exact base revision | | |
| Paired-bank SHA-256 | | |
| Final separation value | | |
| Capture point | | |
| Target manifest | | |
| Reviewer’s preferred color palette | | |
| Arithmetic precision | | |
| Pass count and strength | | |

Write one sentence explaining why a changed bank creates a different
experiment even when its filename stays the same.

## Exercise 3 — Architecture gate

Assume a direction `d` has length 4,096.

1. A matrix has shape `[12,288, 4,096]`. Is `W d` dimensionally compatible?
2. A matrix has shape `[4,096, 12,288]`. Can you assume the same update
   orientation? Why or why not?
3. A manifest expects 60 targets, but only 59 are found. What is the correct
   action?
4. Name four fields you would put in each target-manifest entry.

## Exercise 4 — Iteration and arithmetic

Put these stages in order:

- restore each matrix norm;
- capture paired representations;
- apply the bounded projection;
- estimate the current direction;
- run gates;
- recapture from the edited model.

Then calculate:

```text
100 * (87.0274 - 32.2948) / 87.0274
```

- Your result, rounded to two decimals:
- What does the result describe?
- What does it not establish?

Explain in one paragraph why reusing the first-pass activations for every pass
would be a different method from recapture.

## Exercise 5 — Claim audit

For each claim, write **supported**, **too broad**, or **not measured**, then
rewrite it if necessary.

1. “All 60 declared target tensors changed.”
2. “The whole model is behaviorally identical except for the intended effect.”
3. “The 15 defined benign controls were exact matches.”
4. “The artifact is approved for unrestricted deployment.”
5. “No non-finite target values were recorded.”
6. “The final arm preserved every capability.”

For each supportable claim, point to the corresponding field in
`CASE-FILE.json`.

## Exercise 6 — Historical case review

Complete this table from `CASE-FILE.json`.

| Question | Answer |
|---|---|
| Exact source revision | |
| Measurement-bank pairs and rows | |
| Edit strength and pass count | |
| Recapture policy | |
| Target count | |
| Changed targets | |
| Changed non-targets | |
| Non-finite targets | |
| Exact-match controls | |
| Initial and final separation | |
| Missing final-arm evaluations | |

Why must results from a different edit strength be kept separate?

## Capstone — Four-sentence decision note

Use exactly four sentences:

1. Exact artifact identity and edit plan.
2. Integrity and intervention evidence.
3. Geometry result.
4. Missing evaluation and next decision gate.

Draft:

1.
2.
3.
4.

## Knowledge check

1. Which items belong in the immutable experiment identity?
   - A. Only the model family
   - B. Exact model revision, bank digest, and pinned parameters
   - C. The final conclusion
   - D. Chart styling

2. Why is the measurement bank frozen?
   - A. To shrink the archive
   - B. To keep the instrument fixed while measuring change
   - C. To guarantee capability improvement
   - D. To eliminate controls

3. What does recapture mean?
   - A. Reuse original activations
   - B. Measure fresh activations after the preceding pass
   - C. Change the paired prompts after every pass
   - D. Change the base revision

4. Why restore each matrix’s Frobenius norm?
   - A. To constrain scale drift introduced by the update
   - B. To prove universal safety
   - C. To increase the context window
   - D. To make all matrices identical

5. What do exact-match benign controls establish?
   - A. Every model behavior is unchanged
   - B. The selected controls stayed identical under the defined comparison
   - C. Every capability improved
   - D. The artifact is ready for deployment

6. What is the relative reduction from 87.0274 to 32.2948?
   - A. 37.11%
   - B. 54.73%
   - C. 62.89%
   - D. 87.03%