Skip to content

Trust scores belong on data batches, not just people

Pathwize ResearchData quality and evaluation2 min read
Pathwize AIResearch

Why per-batch trust and live inter-rater agreement beat after-the-fact spot checks for catching quality drift.

Most data quality systems rate people: this annotator is good, that one is not. It feels intuitive, but it is the wrong unit. Quality does not live in a person, it lives in the work they produced on a given task at a given time. The right unit to score is the batch.

Why per-person scoring misleads

A strong contributor has a bad day, misreads a new guideline, or hits a task type outside their strength, and produces a weak batch. A weaker contributor nails a simple batch. Per-person reputation smooths over both, so you keep shipping the strong contributor's bad batch and needlessly discount the weak contributor's good one.

Per-person scores also react slowly. By the time someone's average drops enough to notice, many batches have already gone into training.

What per-batch trust gives you

Attach a trust signal to each batch, from live inter-rater agreement, gold-standard items, and review outcomes, and you get a live, actionable metric. A batch below threshold can be paused, re-reviewed, re-clarified or re-routed before it contaminates a dataset. You act on the actual unit of risk, not a lagging average.

It also lets you quarantine, not distrust

The biggest practical win is containment. When a problem appears, per-batch lineage lets you quarantine the specific affected batches rather than throwing out a whole dataset or blacklisting a contributor. That is cheaper, fairer, and keeps good people doing good work.

Spot checks are the weakest version of this

After-the-fact spot checks are per-batch scoring done too late and too sparsely. They sample a little, look backwards, and miss clustered failures. Live per-batch trust is the same idea done continuously and completely, so drift is caught while you can still fix it.

Score the right unit

Pathwize scores trust per batch with live agreement and recorded oversight, so you catch drift early and quarantine precisely. Book a demo to see it on a sample of your data.

Frequently asked questions

What does scoring trust per batch mean?+

Instead of rating a contributor overall, you attach a trust signal to each batch of work, from live inter-rater agreement, gold-standard items and review outcomes. That gives you a live metric on the actual unit of risk.

Why is per-person quality scoring not enough?+

Because a good contributor can produce a bad batch and vice versa, and per-person averages react slowly. You end up shipping a strong contributor's bad batch and discounting a weak contributor's good one, and you notice drift too late.

How does per-batch trust help when something goes wrong?+

It lets you quarantine the specific affected batches using lineage, rather than discarding a whole dataset or blacklisting a contributor. That is cheaper, fairer, and keeps good people working.

Related reading

See Pathwize on your own data

Source verifiable expert data with provenance built in, EU-native and audit-ready.

Book a demo
← All posts

Related stories

ResearchPathwize AI

Expert evaluation vs crowd labelling for frontier models

Crowd labelling scales cheaply but breaks on hard, high-stakes tasks. Here is when expert evaluation pays for itself and how to combine the two.

InsightPathwize AI
Research

Live inter-rater agreement: catching data drift as it happens

Waiting for eval regressions to reveal bad data is too late. Live agreement between reviewers flags drift while you can still fix it.

ResearchPathwize AI

Human evaluation for reasoning models: what breaks

As models reason better, evaluation gets harder, not easier. Here is why generic raters fail on reasoning and what to do instead.