Exercises

Page Maps

graph LR
  family["Reproducible Research"]
  program["Deep Dive DVC"]
  section["Migration Governance DVC Boundaries"]
  page["Exercises"]
  capstone["Capstone evidence"]

  family --> program --> section --> page
  page -.applies in.-> capstone

flowchart LR
  claims["claims inventory"] --> findings["evidence findings"]
  findings --> contract["migration contract"]
  contract --> decision["cutover decision"]
  decision --> rules["governance register"]
  rules --> ownership["ownership handoff"]
  ownership --> review["stewardship review"]

These exercises form one cumulative stewardship lab. You will produce a review packet for the capstone's release-boundary migration, then use that packet to make and defend a cutover decision.

The work is designed for independent study. Every exercise names:

• the evidence to inspect;
• the artifact to produce;
• a falsification question;
• a completion check.

Do not answer from memory. Run the routes and cite the generated evidence.

Prepare the evidence and workspace

From the repository root:

make PROGRAM=reproducible-research/deep-dive-dvc \
  capstone-boundary-migration-audit

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

mkdir -p artifacts/learning/deep-dive-dvc/governance-stewardship

Use this packet:

artifacts/learning/deep-dive-dvc/governance-stewardship/
├── claims-and-evidence.md
├── finding-partial-copy.md
├── migration-contract-review.md
├── consumer-and-rollback.md
├── governance-register.md
├── exception-review.md
├── intervention.md
├── ownership-map.md
├── cutover-decision.md
└── handoff-review.md

The generated audits remain under artifacts/audit/. Your learning packet should link to them rather than copying their full content.

Exercise: Inventory the repository trust claims

Inspect:

• capstone/dvc.yaml;
• capstone/dvc.lock;
• capstone/params.yaml;
• capstone/publish/v1/manifest.json;
• capstone/docs/PUBLISH_CONTRACT.md;
• the boundary migration report.json;
• the recovery readiness report.json.

In claims-and-evidence.md, create a table with these columns:

Claim	Direct evidence	Falsifying route	Result	Proof limit

Include at least these claims:

• pipeline declarations identify the current computation;
• the promoted inventory is complete;
• the target is byte-identical to the source;
• consumers can use the target boundary;
• protected revisions remain recoverable;
• each cross-system responsibility has an authority.

Classify each result as verified, contradicted, unverified, or out-of-scope.

Falsification question: What observation would make each claimed result false?

Completion check: Every row names a concrete file or executed route, and no row uses “the repository looks correct” as evidence.

Exercise: Write a contract finding from the partial copy

Compare:

workspace/complete-cutover/assessment.json
workspace/partial-copy/assessment.json

Write finding-partial-copy.md with:

• claim under review;
• exact evidence path;
• changed observation;
• failed check;
• consumer consequence;
• immediate safety action;
• smallest complete repair;
• closure proof;
• claims you deliberately do not make.

Do not use “migration failed” as the finding. Name the missing artifact and the invariant it contradicts.

Falsification question: If report.md were added, which other observations would still be needed before approval?

Completion check: A different maintainer could perform the repair and know which receipt closes the finding.

Exercise: Review the migration change budget

Read:

migration-contract.json
migration-decision.json
workspace/unbounded-change/assessment.json

In migration-contract-review.md, record:

• exact source and target;
• one permitted changed surface;
• every preserved invariant;
• required before and after evidence;
• consumer acceptance;
• rollback promise;
• ownership split;
• exception requirements.

Then explain why adding metric-schema to changed_surfaces must block this cutover even if the target files match.

Draft a separate one-paragraph contract for a future metric-schema migration. It must version semantic meaning rather than pretending the schema is only a location change.

Falsification question: Which evidence distinguishes a second boundary change from a more detailed description of the first?

Completion check: The release-location and metric-schema decisions are independently approvable and independently reversible.

Exercise: Prove consumer acceptance and rollback readiness

Inspect:

workspace/unverified-consumer/assessment.json
workspace/no-rollback/assessment.json
capstone/docs/boundary-migration-audit-guide.md

Write consumer-and-rollback.md containing two contracts.

The consumer contract must name:

• durable target reference;
• consumer-equivalent credential class;
• manifest retrieval;
• digest verification;
• smallest meaningful consumer operation;
• retained receipt.

The rollback contract must name:

• trigger;
• decision authority;
• retained source;
• restoration route;
• time objective;
• post-rollback consumer check;
• rollback window;
• separate source-retirement decision.

Falsification question: Why can equal source and target inventories coexist with a failed consumer contract?

Completion check: Neither contract uses “switch back if necessary” or “consumers were notified” as proof.

Exercise: Build a proportional governance register

Create governance-register.md with one row for each:

• stage dependency change;
• parameter affecting model behavior;
• metric definition or population change;
• experiment promotion;
• release bundle change;
• DVC remote migration;
• consumer-boundary cutover;
• retention deletion.

Use columns:

Trigger	Contract	Evidence route	Block condition	Authority	Receipt

Add a consequence class: ordinary, reproducibility, comparison, consumer, or protected-history risk. Match the depth of evidence to the consequence.

Falsification question: Which meaningful change would escape a rule that looks only at filenames?

Completion check: A prose-only edit does not require a recovery drill, and a remote deletion cannot pass through ordinary review.

Exercise: Evaluate an expired exception

Inspect:

workspace/stale-exception/migration-decision.json
workspace/stale-exception/assessment.json

Write exception-review.md with:

• bypassed rule;
• owner;
• reason;
• expiry;
• removal condition;
• compensating control;
• current evaluation date;
• decision;
• permissible repair choices.

Explain the different work performed by expiry and removal condition. Decide whether the exception should be removed, renewed through a fresh decision, or cause the cutover to remain blocked.

Falsification question: What direct evidence would show that the removal condition has already been satisfied?

Completion check: Your response does not silently extend the existing exception.

Exercise: Intervene on a mutable release shortcut

A contributor writes:

The registry has latest/model.json, so deployment can use it.

Write intervention.md using:

• observation;
• mechanism;
• affected contract;
• consequence;
• bounded repair;
• closure proof;
• excluded concerns.

Your repair must connect:

Git revision
-> DVC artifact identity
-> promoted manifest digest
-> immutable registry reference
-> consumer retrieval receipt

Falsification question: How would you demonstrate that the immutable reference still identifies the approved bytes after another promotion?

Completion check: The intervention teaches a repair and does not merely label the shortcut an anti-pattern.

Exercise: Assign cross-system ownership

Inspect:

workspace/tool-ownership-overreach/migration-decision.json
workspace/tool-ownership-overreach/assessment.json

Create ownership-map.md for:

• source revision;
• artifact lineage;
• shared verification;
• consumer release lifecycle;
• access policy;
• deployment rollout;
• incident response;
• cutover decision;
• retention deletion.

Use columns:

Fact or decision	Evidence producer	Authority	Handoff receipt	Failure owner

Explain why non-empty owner fields do not prove correct ownership. Identify the fact for which DVC is authoritative and the fact for which the registry is authoritative.

Falsification question: Which system can actually observe and enforce each decision?

Completion check: Every fact has one authority, and no single tool owns the entire lifecycle by convenience.

Exercise: Decide whether cutover is approved

Assume the proposed migration has these observations:

• source and target inventories match;
• only consumer location changes;
• before and after verifiers pass;
• consumer retrieval passes;
• source remains available;
• rollback has never been exercised or timed;
• owners are correct;
• no exception is active.

Write cutover-decision.md.

Your decision must include:

• APPROVE_CUTOVER or BLOCK_CUTOVER;
• a premise table;
• the decisive evidence gap;
• immediate safety action;
• repair proof;
• what the decision does and does not authorize.

Then write the new observation that would change your decision.

Falsification question: Can strong target evidence compensate for an unexercised rollback contract?

Completion check: Source deletion is not authorized by the cutover decision.

Exercise: Conduct an independent stewardship handoff

Write handoff-review.md as if the next maintainer missed the course.

Give them a route that lets them:

• identify current declarations and recorded state;
• reproduce the pipeline;
• verify the promoted release;
• compare an experiment under a semantic contract;
• restore protected history;
• evaluate a boundary migration;
• find decision authorities;
• close or escalate one failed finding.

For each route, state:

• entry command or file;
• expected evidence;
• proof limit;
• owner when it fails.

End with three questions the new maintainer must answer without your help:

1. Which system identifies promoted bytes?
2. Which evidence permits source retirement?
3. Which claim remains unverified by the capstone routes?

Falsification question: What part of the handoff still depends on your private explanation?

Completion check: Remove or rewrite every part that requires “ask the original maintainer.”

Mastery review

Your packet demonstrates module mastery when:

• findings separate contradiction from missing evidence;
• migration contracts permit one coherent change;
• consumer acceptance and rollback have independent proof;
• governance rules connect triggers to contracts and receipts;
• exceptions have an end mechanism;
• interventions include closure evidence;
• ownership follows observation and authority;
• decisions name failed premises;
• handoff routes expose their proof limits.

Compare your reasoning with the answer page only after all ten artifacts exist. The answer page provides models, not text to copy.