Check a table's data

Everything else in this console is derived from manifests. It opens no data file, touches no blob column, and costs kilobytes — and the README, the guide and the instructions an agent gets all say so.

The questions a training run actually arrives with are not in a manifest. Are these rows duplicated. Is anything missing its content. Did my split leak. Are these embeddings alive. Those are properties of the data, and the only way to answer them is to read it.

So they are a scan rather than a finding, and everything about the panel follows from that.

The bill comes first

Data tab. Nothing has run: what is on screen is what each check would read, weighed from the data-file footers without opening a page of them.

missing-content    reads 11.9 KB–43.4 KB, and none of the 2.65 GB of blob payload
                   those descriptors point at
exact-duplicates   reads 6.8 KB–43.4 KB
vector-health      reads 3.4 MB

Two things to read in that.

The range, not a single number. The columns weigh the first figure; a pass also pays footers and column metadata per data file, and Lance reads a small file whole. On a table of kilobyte files the real read can come in a little above the top of the range, and the caveats under each check say which of those reasons apply. What the quote is for is the order of magnitude, because that is the decision you are making.

What it will not read. On a media table this is the interesting half. A Blob V2 column projects to its descriptor — a position and a size — so asking whether every video is actually there reads the descriptors and none of the video.

A check that will not run says which sentence applies

class-balance    name the one column that holds the label
split-leakage    name the identity an item is known by, and the column that says
                 which split it is in
near-duplicates  'vector' has no vector index. Without one this would be a full pass
                 over every vector for each row sampled

Which column holds a label is not visible in a schema, and the first string column is usually an id. A check that guessed would produce a real-looking answer to a question nobody asked, so it asks.

near-duplicates refuses an unindexed column rather than falling back to a brute scan. The fallback is one full pass over every vector per row sampled — a bill nobody agreed to by pressing a button labelled check.

Stopping stops it

This is the one place in the console where cancel means cancel.

A Lance query cannot be interrupted — the query panel says so, and abandons the wait rather than the work. The loop over batches here is ours, so the flag is read between them, and a cancelled scan reports the checks that finished and the bytes they had spent.

What comes back

Each result carries what it actually read, drained from the same handle every other panel drains — so the quote can be checked against the outcome rather than believed. The findings themselves are the same shape as the metadata ones, with the same evidence on the same row.

The version is pinned when the scan starts. A distribution reported against "the table" is a claim with no moment attached to it, and a scan takes long enough for a write to land in the middle of one.

Measured, on somebody else's data

hf://datasets/lance-format/openvid-lance/data, pylance 11.0.0, 2026-09-05. 937,957 rows of video with their captions and embeddings, published by LanceDB.

The first line is the one to look at. Nine hundred thousand videos, and deciding whether the checks are worth running costs less than a megabyte.

The third line is the shape of the trade this panel exists to make visible: checking for duplicate paths is 69 MB, and you knew that before you spent it.

From a terminal

lancescope check moments --quote

Prints what each check would read and stops. Drop --quote to run them, add --fail-on warn to make it a gate in a pipeline. A check that could not run exits 3 rather than 1, because "this split leaks" and "we could not look at your split" are different facts and a gate that returns the same code for both has stopped being one.

An agent gets data_scan_estimate and deliberately not a way to run a scan: it can tell you what checking would cost, and spending your read budget stays your decision.

What none of this can tell you

Said here because a panel of green ticks implies otherwise.

It cannot tell you whether a label is right. It cannot find a duplicate the embedding does not encode — two photographs of the same thing under different names are one item to you and two to this. It cannot attribute a distribution to a cause. And near-duplicates are approximate twice over: a sample rather than the table, and an index that returns close neighbours rather than the closest.

Every one of those is written on the panel as well, because the place a caveat is useful is beside the number, not in a document somebody reads afterwards.