> For the complete documentation index, see [llms.txt](https://docs.askyazi.com/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.askyazi.com/build-a-study/response-quality-and-validation.md).

# Response quality and validation

Check answers as they arrive with automatic format checks and AI answer checks on text, photos and voice notes, and understand how this differs from Response Trust.

Yazi checks answers at two moments. **While the participant is answering**, format checks and optional AI answer checks can reject an answer and ask for it again. **After they finish**, Response Trust scores the whole response and flags anything suspicious for you to review. This page covers the first kind and the settings that control both.

{% hint style="info" %}
**In short:** Yazi always re-asks when an answer is the wrong type (a photo when you asked for a video, a word when you asked for a number). For stricter checks, switch on **Enable Question Actions** in **Study configuration > Response quality** and add an AI check to a question. It holds the participant at that question, asks them to redo it, and can screen them out. It judges what a video says, not what it shows. Response Trust is separate and blocks nobody.
{% endhint %}

## The Response quality panel

Open **Build > Study configuration > Response quality**.

<figure><img src="https://358667285-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FzMRQUXiGQlBwWdg6WEmG%2Fuploads%2F6O14JlpnbriWmuSvWCo0%2Ffs-response-trust-signals.webp?alt=media&amp;token=975423e8-906b-47be-b142-e7a9f890a191" alt="Study configuration > Response quality: Response Trust and its five signals"><figcaption></figcaption></figure>

| Setting                             | What it does                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                   |
| ----------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ |
| **Response Trust**                  | Scores each completed response for low quality and fraud ("Catch low-quality and fraudulent responses automatically"). On by default for new studies.                                                                                                                                                                                                                                                                                                                                                                          |
| **Signals**                         | The checks Response Trust combines ("5 of 5 active" when all are on). Each can be ticked on or off: **Response speed** (finishing much faster than the rest of the panel), **Straight-lining** (picking the same option again and again across choice questions), **Duplicate images** (the same photo reused across questions or sent by different participants), **Voice match** (the same voice under several participant IDs) and **AI judge** (an AI model reads the answers and flags nonsensical or inconsistent ones). |
| **Sensitivity**                     | **Lenient** (only flag the most obvious cases), **Balanced** (recommended for most studies) or **Strict** (cast a wider net and expect more false positives).                                                                                                                                                                                                                                                                                                                                                                  |
| **Auto-exclude rejected responses** | Removes participants marked Rejected from your charts and analysis automatically.                                                                                                                                                                                                                                                                                                                                                                                                                                              |
| **Enable Question Actions**         | Turns on AI answer checks for questions in this study ("Enable AI-powered validation for media responses").                                                                                                                                                                                                                                                                                                                                                                                                                    |

Response Trust is explained in full in [Response Trust](https://docs.askyazi.com/monitor-fieldwork/response-trust.md).

## Checks that always run

You do not need to switch these on:

| Check                     | What the participant experiences                                                                                                                      |
| ------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------- |
| **Wrong answer type**     | Sending a photo to a video question, or text to a voice note question, is not accepted. They are told and asked again.                                |
| **Choice questions**      | Typing something that is not one of the options shown is rejected and the question is asked again.                                                    |
| **Multi Select limits**   | Too many or too few selections are rejected with "You have selected too many options" or "You have not selected the required amount of options".      |
| **Number range**          | A number outside the minimum and maximum you set is rejected with a message naming the range, for example "Please enter a number between 18 and 100." |
| **Document type**         | If you ticked Word, PDF or both on a Document question, other files are refused with a message asking for the right kind of file.                     |
| **Other, please specify** | A participant who picks Other must type something before moving on.                                                                                   |
| **Round selfie videos**   | WhatsApp does not pass these to Yazi, so the participant is told to send the video another way.                                                       |

All of these messages are sent in the participant's language. Rejected answers appear in the transcript so you can see what the participant sent and what they were told, but they are not stored as answers.

## AI answer checks (Question Actions)

An AI answer check reads an answer as it arrives and compares it with criteria you write. If the answer does not meet them, the participant is told what was missing and asked to try again. It works on:

* **Text** answers;
* **Voice notes**, which are transcribed first and judged on what the person said;
* **Photos**, for example "is the till slip readable and dated this week?" or "does the photo show the product on a shelf?";
* **Videos**, judged on their audio only.

### Set up an AI answer check

{% stepper %}
{% step %}

### Switch on Question Actions

In **Study configuration > Response quality**, turn on **Enable Question Actions**.
{% endstep %}

{% step %}

### Open the question

Open the question you want to check and add an AI answer check to it.
{% endstep %}

{% step %}

### Write the criteria

Describe a good answer in specific, checkable terms. For example: "The photo must show the full receipt, with the store name and date readable." Avoid vague criteria such as "a good photo".
{% endstep %}

{% step %}

### Choose what happens on failure

Choose whether a failing answer is asked again or the participant is screened out. A retry limit stops anyone being asked forever.
{% endstep %}

{% step %}

### Test on a small batch

Run the study yourself, then with a small group of real participants, and read what the check accepted and rejected before opening fieldwork fully.
{% endstep %}
{% endstepper %}

### Writing good criteria

* **Be tight and specific.** Checks are reliable when the criteria are concrete ("receipt from a supermarket, dated in the last 7 days") and less so when they are loose.
* **Check one thing at a time.** Split "clear photo of the pack and the price label" into two questions if both matter.
* **Think about the participant.** Every rejection is friction. Use checks where a bad answer would make the response useless.

## Response Trust versus AI answer checks

These two are often confused.

|                         | AI answer checks (Question Actions)                              | Response Trust                             |
| ----------------------- | ---------------------------------------------------------------- | ------------------------------------------ |
| When                    | In real time, on one question                                    | After the participant finishes             |
| Blocks the participant? | Yes: holds them at the question, asks for a redo, can screen out | No: it only flags                          |
| What it judges          | One answer against your criteria                                 | The whole response, across several signals |
| Output                  | Accepted or rejected answer                                      | Approved or Flagged verdict with a reason  |
| Set up                  | Per question, after **Enable Question Actions**                  | Study level, on by default                 |

Use both together: answer checks keep obviously unusable answers out at source, and Response Trust catches patterns such as speeding or duplicate photos across the whole response.

## Limits and things to know

{% hint style="warning" %}
**A rejected participant has no way back.** If a check screens someone out, they receive a rejection message and cannot re-enter the study. Use screen-out on failure only when a failed answer truly disqualifies someone, and prefer "ask again" otherwise.
{% endhint %}

* **Videos are judged on audio only.** No check can see what happens in a video.
* **Test before relying on it.** We do not quote an accuracy figure. Reliability depends on how specific your criteria are and on photo quality.
* **Checks add friction.** Each failed attempt means another message for the participant. Do not add checks to every question.
* **Answer checks do not replace screening.** Use screening questions and quotas to decide who qualifies. See [Screening and incidence](https://docs.askyazi.com/build-a-study/screening-and-incidence.md).

## Related

* [Response Trust](https://docs.askyazi.com/monitor-fieldwork/response-trust.md)
* [Question types](https://docs.askyazi.com/build-a-study/question-types.md)
* [Screening and incidence](https://docs.askyazi.com/build-a-study/screening-and-incidence.md)
* [Media in questions](https://docs.askyazi.com/build-a-study/media-in-questions.md)
