Recovery Guide

Guide Maps

graph TD
  loss["Local loss"]
  remote["DVC remote"]
  pull["dvc pull"]
  checkout["dvc checkout"]
  verify["contract verification"]
  review["recovery review bundle"]

  loss --> remote --> pull --> checkout --> verify --> review

flowchart LR
  before["Record before-status and remote config"] --> erase["Simulate local loss"]
  erase --> restore["Restore from remote-backed state"]
  restore --> validate["Validate publish contract"]
  validate --> inspect["Inspect the recovery evidence"]

This guide exists because recovery is easy to talk about loosely and hard to evaluate honestly.

What the recovery drill proves

• tracked state can be restored after local cache loss
• the promoted publish bundle can be validated after restore
• the remote is part of the repository’s durable story, not an optional convenience

What the recovery drill does not prove

• that the publish bundle is the full internal state story
• that experiments remain semantically comparable
• that every local convenience file is reproducible or durable

Read STATE_LAYER_GUIDE.md when the main confusion is not the recovery sequence itself but which layer is authoritative before and after recovery.

Best route

1. Run make recovery-drill when you want the raw restore rehearsal.
2. Run make recovery-review when you want a durable bundle for later inspection.
3. Read before-status.txt, pull.txt, checkout.txt, verify.json, and publish-v1/manifest.json in that order.