# DWM foundations learning runbook

This runbook turns the course into a repeatable learning sequence. It is not an
execution guide for modifying a model.

## Stage 0 — Confirm scope

**Inputs:** `START-HERE.md`, course boundary, authorization boundary.

**Actions:**

1. State whether you are reviewing evidence, designing an experiment, or
   evaluating an artifact.
2. Confirm that this pack contains no executable edit workflow.
3. If using outside artifacts, confirm you own them or have explicit permission.

**Gate:** the task, artifact, and authority are unambiguous.

## Stage 1 — Define vocabulary

**Inputs:** transcript module 1.

**Actions:**

1. Write one sentence each for DWM, fine-tuning, ablation, and prompt steering.
2. Separate objective, method, evidence, and conclusion.
3. Complete workbook exercise 1.

**Gate:** no conclusion is presented as if it were a measurement.

## Stage 2 — Freeze an experiment identity

**Inputs:** transcript module 2, `CASE-FILE.json`.

**Actions:**

1. List every item needed to identify one experiment exactly.
2. Explain why the paired bank is a measurement instrument.
3. Verify that the historical case records a base revision and bank digest.
4. Complete workbook exercise 2.

**Gate:** a silent revision, bank, parameter, or capture-rule change requires a
new experiment identity.

## Stage 3 — Map the edit surface

**Inputs:** transcript module 3.

**Actions:**

1. Draw the conceptual data path.
2. Check matrix and direction dimensional compatibility on paper.
3. Draft a target-manifest schema with tensor name, shape, layer, and family.
4. Complete workbook exercise 3.

**Gate:** all declared targets are exact and shape-compatible; no guess is used
to resolve a mismatch.

## Stage 4 — Trace the iterative method

**Inputs:** transcript module 4.

**Actions:**

1. Put capture, estimate, update, restore, gate, and recapture in order.
2. Explain why a later pass uses fresh measurements.
3. Calculate the historical separation reduction.
4. Complete workbook exercise 4.

**Gate:** norm restoration is described as a scale constraint, not as proof of
behavioral preservation.

## Stage 5 — Audit evidence layers

**Inputs:** transcript module 5, `EVIDENCE-CARDS.md`.

**Actions:**

1. Sort each record into integrity, intervention, or behavioral evidence.
2. Rewrite any claim that outruns its evidence.
3. List missing tests next to the measured results.
4. Complete workbook exercise 5.

**Gate:** each claim points to a measurement, and each limitation is visible.

## Stage 6 — Review the historical case

**Inputs:** transcript module 6, `CASE-FILE.json`.

**Actions:**

1. Verify the arithmetic from the recorded separation trace.
2. Check target, non-target, finite-value, and control counts.
3. Identify results that do not transfer from another edit strength.
4. Complete workbook exercise 6.

**Gate:** the case is not described as a release qualification.

## Stage 7 — Write the decision note

**Inputs:** completed workbook, `MODEL-CARD-TEMPLATE.md`.

**Actions:**

1. Write the four-sentence capstone note.
2. Put limitations directly after the result.
3. Name the next evaluation gate.

**Gate:** a reviewer can tell what is known, unknown, and required next without
consulting unstated context.

## Stage 8 — Verify the learning pack

**Inputs:** `MANIFEST.sha256`.

**Actions:**

1. Compute a SHA-256 digest for each listed file.
2. Compare every digest with the manifest.
3. Record any mismatch and obtain a clean copy before relying on the material.

**Gate:** every listed file matches its published digest.