Troubleshoot a run

Diagnose a run that is behaving unexpectedly by symptom — stuck items, a blocked deploy, a surprising charge, a short export, a wrong build, a failed upload, or a hit limit — and route to the page that owns the fix; for an exact on-screen message, search the errors reference instead.

When a run behaves in a way you did not expect, find the symptom below that matches what you see. Each symptom explains what is going on and what to check, then links to the fix. For the exact text of a message on your screen, such as a red counter, an entry in the Issues window, or an item’s failure reason, search Errors and how to fix them by the message instead.

The run seems stuck

An item’s row reads In Progress while it is still being labeled, and the filter chip above counts it as in_progress. Open the run’s Items tab and see what state each item is actually in:

The counter chips (completed, in_progress, paused, failed, canceled) show where the batch is. Open an item’s row for its own status.
A Quorum node item can stay in progress through extra review rounds. When LLM review rejects every answer, the platform adds another expert and reviews again, up to twice the configured overlap. See When review rejects every answer.
A Paused badge on the held items, and a paused: N chip among the item filters, identify a paused run. See Resume or cancel a run.
A run pauses itself and shows a top-up banner when the team balance runs out mid-run. Add funds, then Resume. See Pricing and billing.

I need to stop everything

The run controls are on the deployed pipeline’s Pipeline tab. The agent cannot cancel a run for you. The buttons appear only once the deploy has finished. A run still deploying has no live work to stop yet.

To end the run for good, click Cancel.
You cannot hold a run yourself. A run pauses only when the balance runs out; Resume then appears in its place until you continue it.

Resume or cancel a run covers what each button does.

The run charged more (or less) than I expected

Every run’s Cost breakdown (forecast) carries two totals. The final bill uses the items the run actually processes. See Pricing and billing:

Expected is the average forecast. Maximum is the ceiling if every item triggers every billable event. Billing is pay-as-you-go for the items the run actually processes, so a real total below Expected is normal forecast behavior.
On a Quorum node, the Maximum column doubles the review calls to cover rejection-driven escalation. Compare your bill against Expected rather than that doubled ceiling.
A cloned pipeline is priced fresh and frozen at its own deploy. It does not inherit the first run’s contract.

My export is missing items

An export is a snapshot taken the moment you start it and does not update afterward. See Export your labeled results. An item that had not started when you began the export carries only its index and status. input, output, and per-node metrics fill in as the item reaches each state. See Export format. For a file that reflects every result, wait until the run has finished, then start a fresh export.

The deploy is blocked

A blocked deploy names its reason, so search Errors and how to fix them for the exact message. The gates you can clear from here:

A red launch count over 10,000. The Start node caps a run at 10,000 items. A selection above it turns red and blocks the deploy. Lower Max items to launch to 10,000 or fewer, or narrow the filter. A filter matching zero items blocks the deploy too. See Label a large dataset in batches.
An issue in the Issues window. It sorts problems into errors you must fix and warnings you can override. Any failed check blocks, and so does a built-in lint that has not finished: Quality Requirements Lint and User Interface Type Check on every labeling node, Dataset Items Lint on the Start node — the node cards call them Quality Requirements, User Interface Types, and Dataset Items. Each row has a fix-with-agent icon button, and Fix all with agent hands the agent the whole list. See the deploy gate.
An empty balance. A run cannot deploy on an empty balance, below $0.01. Add funds to your team, then deploy again. See Pricing and billing.

The agent built the wrong thing

Describe the fix in the terms of the pipeline’s nodes: which node labels, which routes, and which computes. This lets the agent act on it. Pipelines names what each node does. There are two ways to change what the agent built:

Ask the agent, in plain words, to change a node or rebuild the labeling interface. In the Issues window, the Fix with agent button hands the agent one issue to resolve. See Self-check.
Edit by hand on the node’s cards. The labeling interface has an Edit tab. See Review the labeling interface.

The upload failed

An upload is validated in full before anything is written, so a rejected file adds nothing. The exception is the 150,000-item dataset cap: a file that validates but overflows the cap is imported up to the cap, and the upload then fails with Dataset item cap reached; remaining rows were not imported. Fix the file and upload again. See Create and manage datasets. The exact messages are on the dataset errors.

The whole file is rejected when it is neither JSON nor CSV, is empty, has a non-array JSON root, is over 1024 MB, or holds an item over the per-item size cap.
Item validation lists up to 50 problems and writes nothing. Fix the listed items, then upload again.
An appended file must stay schema-compatible with the first upload’s fields. A missing field, or a changed type, fails the whole append and adds no items.

I hit a limit

These caps are fixed. The linked page gives the exact message and the way around each cap. Search Errors and how to fix them for the message you saw.

Limit	Cap	Where the rule lives
Items per run	10,000	Label a large dataset in batches
Items per dataset	150,000	Create and manage datasets
One uploaded item	1024 KB	Create and manage datasets
Fields per dataset	500	Create and manage datasets
One uploaded file	1024 MB	Create and manage datasets

The 10,000-items-per-run cap has no override. Use more runs for more data. Label a large workload as a cycle of batches. See Label a large dataset in batches.

FAQ

Can I run a pipeline twice? No. A deployed pipeline runs once and then locks. Clone it to label another batch. See Label a large dataset in batches.

Why is my pipeline locked? Deploying locks the pipeline against edits and reruns. A failed deploy is the one exception and frees the same pipeline to fix and deploy again. See How the platform works.

Can I edit the pipeline after deploy? No. The pipeline locks at deploy, including the labeling interface. Clone it to keep iterating. See Review the labeling interface.

Do the experts see my data? During a run, the expert labelers you chose see the items they label (How the platform works). Before the run, the interface, code, and LLM-review checks use only the sample values you enter; the Dataset Items Lint on the Start node is the one check that reads your dataset items, to report empty selections and missing required values (Self-check).

Why did switching the dataset clear my filter? Switching the dataset in the editor always clears Max items to launch, and clears the item filter whenever the new dataset cannot carry it over — a filter is built against the old dataset’s fields, so a genuinely different dataset drops it. See Start and End nodes.

Errors and how to fix them — the exact on-screen message, indexed by its text, with cause and fix
Resume or cancel a run — the run controls and what each one does
Self-check — the Issues gate that blocks or clears a deploy
Pricing and billing — Expected vs Maximum, funding, and the automatic pause on an empty balance