pathwize

Quality & security

Here is exactly what we check, what we record and how we protect your data. And what we do not claim.

Quality

Six checks, before and during the work.

All of them run in the product today. You configure the thresholds per project.

Before

Qualification test on your items

Typically 50 items. You set the pass mark (default 90 percent, IoU 0.7 for boxes) or grade the answers yourself. Unpaid for experts, free for you.

During

Hidden reference items

A share of production items (default 5 percent) has a known answer. Anyone who falls below your threshold is paused automatically.

During

Second and third reviews

Choose a review sample or two to three independent raters on every item. Disagreements go to a review queue.

Always

A record for every label

Who made it, when, how long it took, which test they passed and how it was reviewed. Stored with every label and available to you.

Before

Screened in an AI interview

A short recorded interview with an integrity check. A person reviews flagged cases, nobody is rejected automatically.

During

Practice and calibration

An optional practice round before the test and a calibration round with item-by-item feedback after it, so experts align with your guidelines before production.

Human or machine

How we make sure a human expert did the work.

At the door

Every expert does a recorded AI interview. The system watches for tab switches and long silences before polished answers, then a person reviews anything flagged. No one is rejected by the AI alone.

On your data

The qualification test uses your own items and reference answers. Copying a generic chatbot answer rarely passes a domain test with your specific rules.

During production

Time per item is measured, reference items keep scoring each expert, and second reviews compare experts with each other. Outliers stand out quickly.

Records

What we store for every label.

Each judgment is linked to a person, a time and a review. This is what you can show when someone asks how your data was made.

Example label record
{
  "item_id": "img_1305940",
  "label": "170302 Bituminous mixtures",
  "labeled_by": "ex_4b81",
  "labeled_at": "2026-10-23T09:41:12Z",
  "seconds_spent": 19,
  "expert_test_score": 96,
  "is_reference_item": false,
  "second_review": { "by": "ex_2a07", "outcome": "agree" }
}

Security

How your data is protected.

Hosted in the EU

Your data and the expert workspace run on EU infrastructure.

No downloads for experts

Images and files are shown inside the platform through signed, short-lived links, not offered as downloads.

Your data stays yours

Results become yours on payment. Pathwize does not reuse your data or train its own models on it.

Need-to-know access

Only experts approved for your project see its items, after accepting confidentiality terms for it.

Export when you leave

After the contract ends you can export your data and results for 30 days.

DPA included

A data processing agreement is part of the contract. Sub-processor list on request.

Security questions or a review? Write to security@gopathwize.com.

EU AI Act

Documentation that supports Article 10, without overselling it.

High-risk AI systems need documented training, validation and test data. Pathwize gives you a record of who made each judgment, how they were tested and how their work was reviewed. That supports your documentation. It is not a compliance certificate, and it is not legal advice.

Straight answers

What we do not claim.

No certifications yet

We are not ISO 27001 or SOC 2 certified today. We share our technical and organizational measures on request.

No compliance guarantee

No vendor can make your system AI Act compliant. We provide records that support your own documentation.

No perfect labels

Experts make mistakes. That is why there are tests, reference items, second reviews and rework within 14 days.

FAQ

Quality and security questions.

How do you check quality during the work?

Hidden reference items are mixed into production. If an expert's accuracy on them drops below your threshold, their work pauses automatically. On top, you can have a sample or every item rated by a second or third expert, with disagreements going to a review queue.

How do you know a human did the work, not ChatGPT?

Experts are screened in a recorded AI interview with an integrity check (tab switching, unusual pauses before long answers). A person reviews flagged cases, nobody is rejected automatically. During work, time per item is recorded and reference items catch copy-paste answers that miss domain details.

Where is my data stored?

In the EU. Images and files are shown inside the platform through signed links and are not offered as downloads to experts.

Who owns the results?

You do. Your data stays yours, results become yours on payment, and Pathwize does not train its own models on your data. After the contract ends you can export everything for 30 days.

Do you sign a DPA and NDAs?

Yes, a data processing agreement is part of the contract. Experts accept confidentiality terms for your project before they get access. A sub-processor list is available on request.

Are you certified (ISO 27001, SOC 2)?

Not today. We say so openly and share our technical and organizational measures on request.

Does this make my system EU AI Act compliant?

No vendor can promise that. What you get is documentation of how your data was made (who, when, which test, which review), which supports your Article 10 records. It is not legal advice.

What if quality is not good enough?

Report it within 14 days and the affected items are reworked.

See the checkson your own data.

Run a free pilot on 50 items and look at every test score yourself.

Request a free pilot