Mid-Course Map

Page Maps

graph LR
  family["Reproducible Research"]
  program["Deep Dive DVC"]
  section["Orientation and Study Practice"]
  page["Mid-Course Map"]
  bridge["State truth to shared trust"]

  family --> program --> section --> page
  page -.bridges.-> bridge

flowchart LR
  state["identified and declared state"] --> comparison["comparable measurements"]
  comparison --> experiments["controlled candidates"]
  experiments --> handoff["independent execution"]
  handoff --> recovery["surviving evidence"]
  recovery --> promotion["consumer trust"]

Use this page after Module 04, after a long interruption, or when a later lesson feels like an isolated DVC feature. The middle of the course changes the unit of trust:

• early modules ask whether one result has an explainable origin;
• middle modules ask whether results can be compared, shared, and recovered;
• late modules ask which result may cross a downstream boundary.

The bridge is conceptual, not chronological. Do not continue until the earlier state claims needed by your next comparison are stable.

Diagnose readiness

Without opening the earlier lessons, answer:

1. Why can two files at the same path represent different data identity?
2. Which runtime facts can influence a result without appearing in dvc.lock?
3. Why can dvc status be empty while an output is semantically stale?
4. What evidence distinguishes an unnecessary rerun from a hidden influence?
5. How would you prove that a repaired dependency works on the next change?

Use the result:

Pattern	Return before continuing
data identity answer relies on path names	Module 02
runtime is treated as automatically captured	Module 03
status is treated as universal freshness proof	Module 04
repair relies on --force or output deletion	Module 04 worked investigation
all five answers are evidence-based	begin Module 05

This is not a memory quiz. These distinctions are prerequisites for interpreting metrics and experiments honestly.

Understand the middle-course dependency

A metric value is derived from a population, controls, code, and runtime. An experiment compares such values against an authority. A reviewer handoff must reconstruct those inputs and decisions. Recovery must preserve the state needed to do so.

flowchart TD
  identity["population identity"] --> metric["metric claim"]
  controls["declared controls"] --> metric
  runtime["bounded runtime"] --> metric
  metric --> experiment["candidate comparison"]
  baseline["baseline authority"] --> experiment
  experiment --> handoff["submitted review state"]
  handoff --> recovery["protected recoverable state"]
  recovery --> promotion["eligible promotion evidence"]

Skipping a boundary does not merely leave a gap in course coverage. It weakens the claims available downstream.

Module 05: decide whether numbers are comparable

Start Module 05 with this deliberately incomplete statement:

F1 increased from 0.7368 to 0.7826.

Ask what must remain stable before “increased” can mean improvement in the same measurement:

• population identity and size;
• metric name and definition;
• unit and aggregation;
• selected threshold;
• schema contract;
• evidence presence.

Run the focused audit only after predicting accepted and rejected cases:

make PROGRAM=reproducible-research/deep-dive-dvc capstone-metric-contract-audit

Exit Module 05 able to explain why DVC can report a numerically correct diff for a semantically invalid comparison.

Module 06: decide whether candidates share an experiment

Module 06 adds baseline authority and allowed-change boundaries.

Before ranking candidates, establish:

Boundary	Review question
baseline	which revision and evidence define the reference?
intent	what question does the experiment answer?
allowed change	which controls may differ?
preserved state	which data, metric, runtime, and pipeline boundaries must match?
decision	is the candidate comparable, selectable, and promotable?

These are different decisions. A comparable result need not be selected; a selected candidate need not be ready for promotion.

Focused route:

make PROGRAM=reproducible-research/deep-dive-dvc capstone-experiment-comparability-audit

If the available target name differs, use make ... help and the module's current command block. Preserve the actual receipt rather than inventing an expected one.

Module 07: move from author to reviewer

Local evidence becomes a handoff claim. The decisive test is whether a cold reviewer or CI executor can reconstruct the submitted state without undocumented local context.

Review:

• Git revision and dirty-worktree state;
• DVC metadata and required objects;
• environment construction;
• exact command route;
• preserved status before automatic repair;
• result and policy decisions.

Focused route:

make PROGRAM=reproducible-research/deep-dive-dvc capstone-reviewer-handoff-audit

Exit Module 07 able to reject a handoff that succeeds only with the author's warm cache.

Module 08: separate availability from recoverability

An artifact available in one workspace is not necessarily recoverable after loss. Recovery review asks which protected revisions and objects remain authoritative.

Use Module 08 when:

• local cache was lost;
• retention or garbage collection is being changed;
• a remote is being migrated;
• historical revisions matter;
• incident evidence must survive repair.

Focused route:

make PROGRAM=reproducible-research/deep-dive-dvc capstone-recovery-readiness-audit

Exit able to state which exact losses the recovery evidence covers.

Module 09: cross the promotion boundary

Promotion creates a smaller consumer-facing contract than the whole working repository. It should not smuggle internal state, undeclared thresholds, or unverified artifacts into a release.

Review:

• authorized candidate and baseline;
• promotion policy and threshold;
• artifact identities and manifest;
• integrity verification;
• downstream instructions;
• rollback or rejection evidence.

Focused route:

make PROGRAM=reproducible-research/deep-dive-dvc capstone-promotion-integrity-audit

Exit able to distinguish “repository evidence exists” from “consumer may trust this promoted bundle.”

Resume after a missed class

Do not reread the whole course by default. Reconstruct the last completed boundary:

1. Open your artifacts/learning/deep-dive-dvc/module-decisions.md.
2. Find the last claim with prediction, observation, decision, and limit.
3. Re-run only its focused audit if the receipt is unavailable.
4. Use the readiness questions above.
5. Continue at the earliest unsupported boundary.

If you have no learning packet, use the module's worked investigation and final exercise as the fastest honest re-entry test.

Mid-course evidence ledger

Keep one row per boundary:

Boundary	Authority	Preserved receipt	Decision	Limit
metric comparison
experiment candidate
reviewer handoff
recovery target
promotion bundle

The authority column matters. A receipt without a named baseline, revision, remote, or policy can be operationally detailed and still ambiguous.

Leave this map when

Move to the Mastery Map when you can:

• reject semantically invalid metric movement;
• name experiment baseline and allowed changes;
• reconstruct a submitted state independently;
• explain protected recovery scope;
• verify a promotion boundary as a consumer would.

If one capability is missing, use the corresponding module rather than treating Module 10 as a general summary.