1. What Is Data Quality Triage?

Imagine that an M&E team receives thousands of indicator records every month.

Each record might contain:

• Indicator name
• Reported value
• Target
• Reporting period
• Data source
• Supporting evidence
• Comments

Before these records are used in dashboards, donor reports, or management decisions, they may need a data-quality check.

Triage simply means sorting records according to what needs attention.

For example:

Important: PASS does not mean that the reported number is proven to be correct. It means that the record passed the checks you defined.

2. The Example We Will Build

We will use this fictional indicator:

3. First Build Your Data-Quality Rules

Do this before using AI.

Decide what you want to check.

For our example, we will use four questions:

These are illustrative rules. Your organization should use its own approved data-quality framework.

4. Understand the Three Simple AI Decisions

TypeSafe’s workflow methodology uses narrow decisions rather than asking an AI model to solve the entire problem in one large prompt. Its published examples use three basic types: Yes/No, Choice, and Score.

For this beginner exercise, we only need the first two.

For example:

Yes/No: Is supporting evidence available?

Choice: What is the evidence strength — Strong, Moderate, Weak, or None?

5. Open TypeSafe

Open the TypeSafe Console:

Open the TypeSafe Console →

TypeSafe’s current agreement identifies the web interface at this address and also provides an API for programmatic use.

6. Create Decision #1 — Completeness

We want to know whether the record contains the information we need.

Create a Yes/No decision with this question:

Provide the indicator record as the information the AI should examine.

Our example contains all the fields, so our human reference answer is:

Run the decision and record what TypeSafe actually returns.

7. Create Decision #2 — Supporting Evidence

Now create another Yes/No decision.

Our fictional record contains a signed attendance register.

Assuming that this is an appropriate source for this indicator, our reference answer is:

Again, compare this with the actual TypeSafe result.

8. Create Decision #3 — Evidence Strength

Now we need a slightly more detailed judgment.

Create a Choice decision:

Use four choices:

• Strong — direct and appropriate evidence is available.
• Moderate — evidence exists but additional checking may be useful.
• Weak — evidence is incomplete or unclear.
• None — no supporting evidence is available.

For our fictional record, the reference answer could be Strong, assuming the signed attendance register is the appropriate source.

9. Create Decision #4 — Overall Status

Now create the final Choice decision.

Choices:

PASS   |   REVIEW   |   FAIL

This is where your organization’s actual data-quality rules become important.

For example, you might define:

These rules are examples, not universal M&E standards.

10. Create Decision #5 — Human Review

Finally, create a Yes/No decision:

Possible answers:

YES or NO.

11. Your First Workflow Is Ready

You now have a simple workflow:

TypeSafe’s published workflow examples use this same general principle: break a larger task into smaller intelligent judgments and combine their results with ordinary programmatic rules.

12. Now Test a Problematic Record

This is where the tutorial becomes more interesting.

Change only the evidence field:

Run the same decisions again.

You would expect the workflow to identify a problem with the evidence.

A possible human reference outcome might be:

*The final status depends on your organization’s approved rules.

13. Use Rules Where AI Is Not Necessary

This is one of the most important lessons in the tutorial.

Suppose your organization’s policy says:

That is a normal software rule:

You do not need AI to make that decision.

AI is more useful for questions that require interpretation, such as:

This combination of AI judgment + ordinary software rules is central to TypeSafe’s published workflow approach.

14. Record Your Results in Excel or Google Sheets

Create a simple validation sheet with these columns:

This spreadsheet is your validation record.

15. Test More Than One Example

Do not stop after one successful test.

Try cases such as:

• complete record + strong evidence;
• missing evidence;
• weak evidence;
• conflicting evidence;
• missing reporting period;
• unclear data source;
• reported value that looks unusual;
• contradictory comments.

The difficult cases are often more useful than the easy ones.

16. Compare AI With Human Decisions

Suppose you test 100 records.

Your validation table might eventually show:

Do not simply calculate an overall percentage and stop.

Look at why the AI and human reviewer disagree.

17. What Does Confidence Mean?

TypeSafe says Jev provides probabilities and confidence with its structured decisions.

For a beginner, think of confidence as information about how strongly the model supports its decision.

But there is an important warning:

You need to test whether higher confidence actually corresponds to higher accuracy on your M&E data.

18. Build a Human Review Queue

After validation, the workflow can eventually be used to sort larger numbers of records.

The goal is not to remove M&E professionals from the process.

The goal is to help them focus their time on records that need attention.

19. Start Small Before Going to Production

Use this progression:

1 record → 10 records → 50 records → 100+ validation records → larger dataset

At every stage, check performance and investigate important errors.

20. Do Not Start With Sensitive Data

For your first TypeSafe experiment, use fictional, public, or properly anonymized data.

Before sending real evaluation data to an external AI service, check:

• your organization’s privacy requirements;
• participant consent arrangements;
• contractual requirements;
• applicable data-protection law; and
• the provider’s current terms and data-processing documentation.

TypeSafe currently provides both a hosted web interface and API, and its legal documentation should be reviewed before operational use with sensitive data.

21. The Complete Workflow

22. Beginner Checklist

23. The Main Lesson

The practical formula is:

Official Resources

TypeSafe AI

TypeSafe Console

TypeSafe Workflow Evals

Introducing System One Models & Jev