How can teen-specific safety classifiers be calibrated and updated over time using real product telemetry (e.g., appeal outcomes, high-friction sessions, and repeated refusals) to systematically reduce false positives on legitimate learning questions without measurably increasing underprotection on non‑negotiable harms like self-harm methods or sexual exploitation?

Answer

Use a closed-loop calibration pipeline where telemetry tunes only appealable cells of the teen safety matrix and where non‑negotiable cells are monitored but never relaxed.

1. Fix the policy frame

• Use the shared teen risk_area × intent × age_band matrix with non‑negotiable cells (self-harm methods, sexual exploitation, etc.) that classifiers can never relax.
• Mark cells: {appealable, review_only, non_appealable} and log which kind was applied.

2. Define telemetry signals per request For each teen request, log (with privacy safeguards):

• matrix_cell, model_action (allow/partial/block/escalate), refusal_style.
• appeal_used?, appeal_reason_chip, appeal_outcome.
• friction signals: rapid abandon, re-ask in other words, repeated refusals in session, channel switch (e.g., to search).
• high‑risk flags: any hit in non‑negotiable taxons, red‑team triggers, or reports.

3. Label and segment calibration data

• Auto-label positives for “likely false positive” when: teen selected learning/help chip, appeal succeeded, and domain reviewers confirm content is allowed for that cell.
• Auto-label “likely underprotection risk” when: reviewer or red‑team finds disallowed detail or grooming-like patterns, especially in or near non‑negotiable cells.
• Always segment metrics by matrix_cell and age_band; never mix non‑negotiables with appealables.

4. Calibration rules for classifiers

• For appealable cells only:
  • Use labeled data to adjust decision thresholds per cell (e.g., lower block probability where many safe appeals succeed).
  • Retrain or fine‑tune narrow intent/risk classifiers on misclassified samples (especially homework, sex‑ed, and mental‑health learning).
  • Add light features from telemetry (appeal_reason, prior safe history for that teen cohort) but not from individual identity.
• For non‑negotiable cells:
  • Forbid any automatic threshold relaxation.
  • Use underprotection findings to tighten models (expand block patterns, add negative training examples), not to loosen.

5. Safety constraints and checks

• Before shipping any classifier change:
  • Run stratified evals per cell: false positives on tagged “legitimate learning/help” vs underprotection on red‑team and abuse sets.
  • Require: (a) no increase, and preferably a drop, in underprotection for non‑negotiables; (b) underprotection stays below fixed ceilings.
  • Only accept changes that lower or hold false positives where underprotection is stable or improved.

6. Use session-level patterns

• Detect cells with many high‑friction sessions (multiple refusals, re-asks, safe appeals) and prioritize them for review and tuning.
• If repeated refusals in an appealable cell almost always end in abandon, bias classifier and prompts toward partial answers + clearer explanations rather than blocks.
• If repeated refusals cluster near non‑negotiable cells, keep blocks but improve graceful refusal templates and routing to safer topics or support.

7. Human-in-the-loop review

• Sample:
  • successful appeals (to confirm they are truly safe learning);
  • failed appeals with strong friction (to catch over‑blocking);
  • any non‑negotiable hits where the model answered.
• Reviewers tag root cause (risk misclassification, intent misclassification, bad refusal pattern) and suggest matrix or prompt tweaks.
• Feed these tags into periodic retraining and matrix updates, not just ad hoc fixes.

8. Product metrics and guardrails

• Track per-cell:
  • false positive rate on curated “legit teen queries,”
  • underprotection rate on curated and red‑team sets,
  • teen satisfaction/retention after refusals.
• Allow automatic calibration only within preset change bands (e.g., thresholds can move ±X per cycle) and roll back any change that pushes underprotection toward ceilings.

This keeps learning-focused false positives falling over time mainly in appealable cells, while non‑negotiable harms are guarded by fixed policy, conservative thresholds, and separate underprotection monitoring that tuning is not allowed to relax.