This guide changes what Redact’s decision model looks for. It assumes the code at lesson 06 or later (course/redact-qvac). For the current questions and threshold, see Decision questions.

​
Add a question for a new kind of field

  1. Open qvac-service/reasoning.ts.
  2. Add an entry to privateQuestions (asked at both levels) or confidentialQuestions (Confidential only). Write it as a statement that is true or false about one line of text:
  3. If the ID isn’t one of Redact’s labels, add the label too: follow How to add a sensitive-field label. Otherwise the suggestion is dropped by the evidence rules.
  4. Run npm test, then measure the change.

​
Reword a question that misfires

If Laya flags ordinary lines for a field, or misses obvious ones:
  • Make the statement concrete: “This text contains a street address” works better than “This is location information”.
  • Ask one thing per question. Two fields in one statement blur the probability.
  • Don’t turn the questions into one multiple-choice question with a “none” option. In testing, Laya then answered “none” for real names, emails and addresses.

​
Change how sure Laya must be

DECISION_THRESHOLD in reasoning.ts is 0.5.
  • To catch more fields at the cost of more suggestions to reject, lower it (for example 0.4).
  • To show fewer, more certain suggestions, raise it (for example 0.7).
To see the raw probabilities before choosing, log result.answers inside classify() and run one page through the app.

​
Measure the change

Stop the service, then run the labelled evaluation:
Compare each case’s f1, precision and recall with your previous run. The gate requires all three at 0.70 or above. See Evaluation.