Snapshots and versions

After this page runs stay comparable while the dataset keeps changing, you can see which rows and fields moved between two versions before trusting a pass-rate change, and a run is nailed to a version so its number means one thing.

Create a snapshot

The Snapshots tab shows a line above the table — rows changed since v4, or v4 is current — computed from the live hash against the latest snapshot. Create snapshot (disabled while the live hash equals the latest) takes a name and note and freezes the non-archived rows into an immutable version with its row count and hash. The table lists version · name · rows · hash · created · note · runs pinned; “latest” is a pointer, not a version. Starting a run on an edited dataset offers snapshot first inline, so you never pin to a moving target.

Pin a run

Run evaluation on v4 opens New run with the dataset locked to that snapshot. The run reads the frozen rows — editing live rows while it is queued is safe — stores the snapshot id, and its Compare with… picker offers only runs pinned to the same version. The dataset’s Runs tab groups runs by version and shades its pass-rate chart by snapshot, so a drop that coincides with a data change is obvious.

Compare two versions

Compare with v3 on the v4 row opens the diff: counts of added · removed · changed · unchanged, and for each bucket the first 50 rows. Rows are matched on the identity the snapshot froze, so a row deleted and re-created with the same text reads as one removal and one addition. A changed row names the fields that differ (target · output · status) — the same field list the row tag hashes — with no before-and-after values, so the dialog can say what moved but not what it says now. Empty compares equal to absent.

Limitations

• The diff carries field names, not before-and-after values; to read a changed value, export the version.
• The table offers only adjacent versions; any pair can be compared through the API.
• No screen opens a version’s frozen rows; they are readable through the API.
• There is no restore-rows-from-snapshot.
• runs pinned counts evaluation runs, not playground grids or simulations.