## TL;DR
Validate in slices with precomputed source digests instead of one giant cross-check: build a compact digest of the source runs once, id, metrics, status, then verify the report's claims claim by claim against the digest with a per-claim budget, sampling deeply rather than exhaustively when the report is large. The validation itself must fit the same resource limits as the generation it checks.

```text
agent timed out when agent tried to validate a summarized report against source runs and failed with a validation timeout error
```

## Use this when
- The run or call fails with this symptom: agent timed out validating a summarized report against source runs.
- You are working on an agent verifying a summarised report against its source runs and the failure repeats on the same run or configuration.
- You need the failing run, call, or stage isolated before changing code or thresholds.

## Not for this skill when
- A validation that times out because a source system is slow rather than the check being large, where the fix is caching the source snapshot.
- The error text names a different tool, stage, or root cause. Match the verbatim error first and use that skill.

## Steps
1. Snapshot the source data into a local digest at validation start, with one compact record per run. Verify: validation works from a local, bounded dataset
2. Extract the report's checkable claims into a list, each with its cited run and field. Verify: claims are enumerable and countable
3. Verify claims in batches against the digest, recording pass, fail, or unverifiable per claim. Verify: progress is incremental and resumable
4. Set a validation budget and report the coverage achieved plus any failed claims, rather than failing wholesale on timeout. Verify: the result states how much was verified

## Variant phrasings
### agent timed out validating summary report against source runs
The verbatim phrasing.
### agent validation timeout report claims digest batched fix
The batched phrasing.
### verify agent report source runs too slow snapshot claims
The snapshot phrasing.

Compatibility: Agent validation of generated reports against source experiment data

## Why it happens
Validation that re-reads the entire source for every claim repeats the generation job's cost and inherits its timeouts. Digests and claim lists turn verification into many small, resumable checks whose coverage can be reported honestly when time runs out.

## Edge cases / pitfalls
- Failed claims matter more than coverage percentage; surface them even when the budget ends early.
- Reuse the generation job's snapshot when one exists; validating against a later state manufactures false failures.

## Provenance

Resolved from the public thread: https://vectle.com/posts/pst_x3eso1GfziLBpiQtg37NhQ