# Review model revisions before a deployment change

Preserve model repository revisions, declared license labels and artifact inventories for a small deployment watchlist.

## Quick start without code

Open https://apify.com/maydit/huggingface-model-change-monitor, paste input.json into its JSON input editor, review current pricing and set a maximum run charge of $0.05. Run manually, check SUMMARY and the result count, then export CSV. Use your own Apify account.

## Downloadable manual runner

Save config.json and run-workflow.mjs in the same new folder. Requires Node.js 22 or newer; no npm install. Supply APIFY_TOKEN through a secure environment variable using your own token. Do not put it in config.json or a URL.

Run without flags for offline instructions:

    node run-workflow.mjs

After reviewing input, scope and current pricing, start one potentially chargeable run:

    node run-workflow.mjs --run

The runner requests 256 MiB, 180 seconds and a $0.05 maximum total charge. The input time budget is 120 seconds. This cap may stop a run before all requested data is saved. It never retries the launch POST, creates a schedule, sends messages, or writes to a remote destination. It uses the Actor's current default build.

If a launch response is uncertain, inspect Apify Console before doing anything else. Resume the accepted run using its ID, with the same config and a new output folder:

    node run-workflow.mjs --resume RUN_ID

The runner checks Actor identity, original input, terminal status, SUMMARY, result count and row identities. It stops on source failures and time/charge stops; a deliberate result cap is allowed and remains a bounded sample. It polls for at most 270 seconds; timeout does not abort or relaunch the Actor.

## Destination

CSV and JSON are written into a new local results folder. Import CSV as literal text (RAW), preserving JSON for exact types. Upsert on record_key and deduplicate replay_key. Add owner/review notes in your own sheet. No sheet is automatically connected. Empty or absent rows never authorize deletion.

## Steps

1. Start with exact public model repository IDs. These two embedding models are examples, not a performance ranking.
2. Record revisionSha alongside the version pinned in your deployment. Fetching metadata does not download or execute model files.
3. Treat licenseDeclared as a publisher label; inspect the repository license and model card for your intended use.
4. Upsert the review row by modelId. Preserve old revision hashes in a separate audit trail rather than overwriting your deployed-version record.
5. Enable comparePrevious with a unique snapshotName for sequential comparisons. Review file-path changes before any download or deployment; this workflow performs neither.

## Limits

- Repository metadata does not establish model quality, security, license suitability or compatibility.
- Artifact counts describe returned paths, not model size or model capability.
- This bounded watchlist does not discover all models or infer repository deletion.

## Example pricing

At FREE-tier rates observed October 18, 2026, 2 saved rows × $0.003 plus $0.00005 Actor Start = $0.00605 in modeled event charges for this sample. This is not an owner invoice or a complete account-cost quote. Platform/API/storage and plan pricing may differ. Check https://apify.com/maydit/huggingface-model-change-monitor/pricing before running.

## Evidence

Saved observation: 2026-10-18T13:05:27.991Z; build 1.0.2; 2 records. Historical output is included in sample.json and sample.csv. Later results can differ. Product: https://mayd-it.com/data-tools/actors/huggingface-model-change-monitor/