Resume and reproduce

A study keeps its own methods and scientific programs. New central code does not silently change an old campaign. That makes recovery and reproduction inspectable, but it also means a new release’s fixes do not retroactively repair an old bundle.

Resume the same campaign

Read its state and run the remaining work through its exact helper:

python3 -m allagma campaign status --study work/my-first-study --campaign toy-v1
python3 -m allagma campaign run --study work/my-first-study --campaign toy-v1

Completed runs are retained. Failed or interrupted attempts remain separate from retries. A process still running under a controller must be reconciled, not duplicated. The resource guide explains reservation recovery and what the local supervisor actually bounds.

Recompute the report

First, inspect a delivered record without executing any of its programs:

python3 -B -m allagma verify-evidence --study work/my-first-study \
  --record campaigns/toy-v1/analyses/a001/paper/claims.json

This reads declared references and reports all independent missing/changed artifacts it reaches. It writes no study files or reviews. Use a quiescent copy and review the coverage and findings; a pass does not establish package completeness or scientific correctness. See limits and exit codes.

To rerun the study-owned analyzer and writer and append a new review:

python3 -m allagma campaign audit --study work/my-first-study --campaign toy-v1

An audit verifies references and regenerates analysis/writing under the pinned helper. It does not retrain a scientific model unless the study’s reproduction command explicitly does so. A saved-weight replay, raw-data reanalysis and fresh full training are three different kinds of evidence. Name the one you performed.

Restore a delivered real study

Each showcase study links to a package index and exact reproduction instructions. Archives may be multipart. The retention helper streams restoration and verifies every hash; do not concatenate a large archive into a filesystem with a small per-file cap.

python3 evals/research-v0.3/retention.py restore \
  --source evals/research-v0.3/runs/r01/package \
  --destination work/restored-culp --download-wheels

This optional path requires the retained evaluation files from a full or expanded sparse checkout and may download pinned dependency wheels. Review artifact sizes and distribution first. Read the restored REPRODUCE.md before execution. Run science through a new finite resource ledger using the standalone broker.

Change the plan deliberately

An amended hypothesis, dataset, metric or sampling plan starts a new campaign with an explanation of the change. Do not edit a frozen protocol or lock to make old results appear to follow the new plan. To change Allagma composition, use Check → Plan → Reconcile → Validate → Adopt at a campaign boundary. Rollback preserves both the prior and later bundles.

---

Canonical source · Generated from the maintained repository document at this build’s commit.