A table loads successfully and still gives the model the wrong evidence. A column can disappear, an event can change meaning, or a labeler can answer a different question from the one you intended. More rows will not repair those failures. Begin by defining what makes data wrong and tracing a suspicious result back through its processing. Convoy's incidents show why checking schemas alone is insufficient. GoCardless then moves expectations toward the producers, while a second contracts talk separates prevention from detection and helps decide where to invest. Labeling instructions bring the same discipline to human-created training data. Finish with a smaller, carefully checked starting dataset and a way to select useful examples, so improving quality does not require cleaning every source before you can begin.
Why first: An unwelcome result is not necessarily bad data. This talk separates incorrect, incomplete and misleading records from evidence that merely challenges expectations, then works backwards through the data lifecycle. Establish that diagnostic habit before the incident accounts tempt you to prescribe a tool.
Why here: Convoy's pricing incidents expose two different breaks: a removed column and machine-generated bids entering data previously produced by people. The second can preserve the table's shape while changing its meaning. That expands the diagnosis from missing fields to source behavior, setting up the need for explicit producer commitments.
Why here: Upstream teams can publish a defined data interface instead of exposing internal tables and leaving consumers to repair every change. Jones shows GoCardless's response, including ownership, versioning and migration expectations. The contract is an agreement about what downstream work may rely on, with the implementation serving that agreement.
Why here: Contracts are targeted work, so applying them everywhere is not an automatic next step. Freeman distinguishes checks on proposed changes from observation of what has already happened. Use the broader evidence to identify critical workflows, then agree on the specific changes those contracts should catch.
Why here: People producing labels also need an unambiguous agreement. Brecque's document-extraction examples show how unclear instructions and guessed answers teach unwanted behavior. The setting is fine-tuning, but the quality question is broader: do the instructions and available evidence support the answer your training record claims?
Why last: You do not have to repair the whole archive to begin. Coleman recommends a small clean source, checks after each transformation, and gradual expansion. His uncertain bowling images also show how example selection can reveal cases people would overlook. This turns the preceding quality requirements into a bounded first dataset and a deliberate way to improve it.