MODULE 08 · THREE LESSONS, SIX STEPS
Browser and sensitive applications
Distinguish candidate selection from professional execution and supervision.
Adjust reading and appearance
Choose the element by the text
What is it?
The documented Jev modality consulted for this course is text. Therefore, a browser flow needs to prepare a textual representation: DOM elements, accessible names, or a list of candidates. Sending a screenshot directly doesn’t match this contract.
A Choice can select the identifier of a candidate. The code queries the DOM again before acting because the page may have changed since extraction. If the element doesn’t exist or its function is different, the flow must stop or update candidates, without clicking an old approximate position.
Include a none option and handle ambiguity between similar elements. Measuring only the classifier’s speed omits extraction, navigation, and execution. The relevant performance is completing the task correctly, with action limits. In this course, the example only chooses a textual ID and doesn’t control an external browser.
Why learn
Distinguish candidate selection from professional execution and supervision.
Key concepts
Use the example below to distinguish the available data, the judgment requested, and what still needs evidence.
Apply: Choose an element by text
Your turn
The button changed after classification. Is it acceptable to reuse the old coordinate?
Check commented answer
No. Validate the current element by its identifier and function, or extract again. The decision isn’t authorization to click anywhere.
Organize a fictional clinical review
What is it?
This case is an administrative exercise using invented data. The task is to recognize whether a required field is filled in and to organize a queue for human review. We are not classifying diseases, recommending treatment, or prioritizing real patients.
A form with an empty allergies field does not demonstrate that the person has no allergies. It shows a lack of information. The exercise policy requires that this case be routed to the responsible professional to complete the form. Even a very confident answer does not eliminate the need for that review.
In a sensitive domain, it is especially important to state the activity boundary. The same software used for an exercise does not become suitable for real clinical use just because it has a nice interface. A real deployment would require a specific project, responsible professionals, and validation in the appropriate context. Here we learn to keep to the administrative scope and the option of insufficiency.
Why learn
Distinguish candidate selection from professional execution and supervision.
Key concepts
Use the example below to distinguish the available data, the judgment requested, and what still needs evidence.
Apply: Organize a fictitious clinical review
Your turn
Why doesn’t a higher confidence threshold turn this exercise into a diagnostic system?
Check commented answer
Because the task type, data, validation, and responsibility are different. A high number does not create clinical capability or authorization to use it.
Organize fictional alerts
What is it?
Consider an invented workflow for following up financial documents. The exercise policy instructs you to review a correction, follow up on a relevant update, and archive news unrelated to the topic. The goal is to organize information, not to decide investments.
A classification should rely on the policy and the document provided. “The price went down” alone does not determine which action to take; the system needs to know the rule it is applying. For an educational prototype, we keep supervised routing and do not connect execution of transactions.
The most important error can be the improper discard of something relevant. That’s why evaluation must look at false negatives, not only the overall accuracy. A queue that appears very efficient because it archives almost everything might be hiding information. The review should include samples of the archived items, in addition to the highlighted alerts.
Deepen the 1.2.0 version
In the browser, choose by DOM, validate that the element still exists, perform the permitted action, and test the effect. A filled field may use already-known data; it does not mean free generation by Jev. Animation time, inference time, and final success should be separated.
Why learn
Distinguish candidate selection from professional execution and supervision.
Key concepts
Use the example below to distinguish the available data, the judgment requested, and what still needs evidence.
Apply: Organize fictitious alerts
Your turn
An evaluation reviews only the selected alerts. What problem does it fail to observe?
Check commented answer
The relevant alerts that were discarded. It’s necessary to audit the archived cases too in order to measure false negatives.
Module wrap-up
- Retrieve the chosen decision from the start of the course.
- Compare your answer with the examples from this module.
- Record a change to the criteria and the test needed to accept it.
Quick check
One example chooses a button between text descriptions. What does it demonstrate?