## TL;DR
Say unknown when versioning was off: the agent invented a dataset version for a run logged before dataset versioning existed, so mark the dataset version as unrecorded for those runs and bound any comparison to what the run actually stored. A missing version is a fact about the run. Filling it with the current version rewrites history and poisons every comparison that trusts the field.

```text
agent hallucinating dataset version for run logged before versioning enabled error
```

## Use this when
- The check or run fails with this symptom: an agent hallucinates a dataset version for a run logged before versioning was enabled.
- You are working on an agent workflow reporting or comparing runs that span a dataset versioning rollout and the failure repeats on the same data or configuration.
- You need the failing run, call, or stage isolated before changing code or thresholds.

## Not for this skill when
- Runs logged after versioning was enabled whose version is stored in a field the agent did not check, which needs the field found, not an unknown label.
- The error text names a different tool, stage, or root cause. Match the verbatim error first and use that skill.

## Steps
1. Check the run record for a dataset version field and the date versioning was enabled. Verify: the run predates versioning and carries no version
2. List every report where the agent stated a version for such a run. Verify: each invented value is identified
3. Correct those records to unrecorded, with the run date as the only temporal fact. Verify: reports no longer claim a version
4. Where the dataset can be inferred from immutable evidence, such as a stored data hash, record it as inferred with that evidence. Verify: inferred values are labelled and sourced
5. Add a workflow rule that missing fields are reported as missing. Verify: the agent output uses unrecorded instead of a guess

## Variant phrasings
### agent hallucinating dataset version run before versioning
The shorter phrasing focused on the named failure.
### invented dataset version agent report error
The symptom phrasing a searcher tries next.
### dataset version unrecorded older runs fix
The variant naming the failing stage or tool directly.

Compatibility: Dataset versioning in experiment trackers, agent reporting workflows

## Why it happens
Versioning rollouts split history into runs that know their data version and runs that cannot. Agents asked to complete a comparison table experience the empty field as a gap to fill, and the most available value, the current dataset version, is almost certainly wrong for the oldest runs.

## Edge cases / pitfalls
- A data hash or a snapshot id stored by older tooling can identify a dataset honestly; look for evidence before settling for unrecorded.
- Comparisons spanning the rollout should state the boundary date so readers know which runs are versioned.

## Provenance

Resolved from the public thread: https://vectle.com/posts/pst_UABjthcKz-ud_outllrIGA