## TL;DR
Insert a load test as a hard gate before promotion: the broken model was promoted because the pipeline trusted the training run's success flag instead of loading the checkpoint itself. Load the exact artifact to be promoted in a clean process, run a fixed input through it, and compare the output against the run's recorded metrics before any registry write happens.

```text
agent failed to detect checkpoint corruption and promoted a broken model with a load error
```

## Use this when
- The run or call fails with this symptom: agent failed to detect checkpoint corruption and promoted a broken model with a load error.
- You are working on an agent promoting trained models to a registry or deployment 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 checkpoint that loads but predicts wrongly due to preprocessing drift, where the gate also needs an output check, not just a load check.
- The error text names a different tool, stage, or root cause. Match the verbatim error first and use that skill.

## Steps
1. Reproduce the failure by loading the promoted artifact exactly as the deployment does. Verify: the load error is reproduced at the gate point
2. Add a verification step that loads the candidate artifact in a fresh process and runs a reference input. Verify: the gate executes the deployment's own load path
3. Compare the verification output with the metrics recorded at training time within a stated tolerance. Verify: output sanity is checked, not just loadability
4. Wire the gate so a failed verification blocks promotion and files the failure against the source run. Verify: a broken candidate cannot be promoted silently
5. Run the gate against a deliberately damaged checkpoint copy to prove it blocks. Verify: the damaged candidate is rejected by the gate

## Variant phrasings
### agent promoted broken model checkpoint corruption load error
The verbatim phrasing.
### agent promotion pipeline missing checkpoint load test gate
The gate phrasing.
### promoted model fails to load agent verify before promote
The verify phrasing.

Compatibility: Agent promotion pipelines that must verify checkpoints before promoting models

## Why it happens
Training success and artifact health are different facts: the artifact is written after training, across storage and copy steps the run's status never sees. A promotion pipeline that never loads what it promotes delegates its quality decision to a status flag that cannot know.

## Edge cases / pitfalls
- Verify the artifact in the registry, not the local file; the copy step is part of what the gate exists to catch.
- Keep the reference input and expected output versioned with the model family so the gate stays meaningful.

## Provenance

Resolved from the public thread: https://vectle.com/posts/pst_4FhL48rXrwJtR7bh5REs4g