EVT 2026-02-22 - Cross-stage UX Optimization Iteration Loop (Draft Only, No Decision Promotion)

Session scope

Autonomous ideation->refine loop for DOL English V2. Goal: UX best quality with high completeness but lean complexity.

Assumptions (max 5)

1. Product lane is DOL English V2 only.
2. Existing confirmed DEC baseline remains valid; this cycle creates draft ideas only.
3. Recommendation remains optional (no mandatory task, no penalty model).
4. Current stack can support rule-based ranking before ML.
5. Existing metrics events can be extended without redesigning tracking architecture.

---

Iteration 1

1) Draft ideas (9 ideas -> 3 directions)

Direction A - Guided Autonomy Loop

1. Next-best-action card with one primary CTA and 2 alternates.
2. Program-adaptive ranking with continuity after each submission.
3. Session-level controls (skill/difficulty/duration) with one-click reset.

Direction B - Low-friction Activation and Return

4. Home first-week activation strip (day 0-7), fully optional.
5. Intent-preserving return after auth/payment.
6. Empty-state trio CTA (Practice -> Vocab -> Course) with contextual reorder if user has course.

Direction C - Trust and Visibility Layer

7. One-line reason label per recommendation.
8. Progress-without-goal panel (trend only, no target gap).
9. Entitlement-aware lock previews with in-place upgrade path.

2) Choose top direction

Top 1: Direction A - Guided Autonomy Loop.

• Impact: direct effect on daily learning actions.
• Feasible: rule-based v1 is practical.
• Differentiation: preserves user freedom while still guiding.
• Risk: recommendation repetition can reduce trust.
• Scope: bounded to recommendation surfaces and interactions.

3) Concept draft (v1)

• Value prop: Give each learner a clear next step that fits their current context without forcing a rigid path.
• User/problem:
  • User: self-study learner (free/pro/pro max), mixed motivation.
  • Problem: too many choices; users know they should practice but don’t know what to do next.
• Scope in:
  • Context-aware candidate selection.
  • Optional user controls in-session.
  • Explainable recommendation reason.
• Scope out:
  • ML model training.
  • New grading logic.
  • Mandatory learning path.
• Core flow (6):
  1. User lands on recommendation surface (Home/Practice Result/Practice Management).
  2. System resolves context (program intent, recent submissions, entitlement).
  3. Candidate pool built from available items.
  4. Ranking applies continuity -> habit-fit -> diversity cap.
  5. UI shows 3-7 items with one-line reason.
  6. User starts/switches/ignores; system logs action for next refresh.
• Must:
  • Always actionable recommendation set.
  • Optional and non-blocking behavior.
  • Fast response on refresh.
• Should:
  • Respect session controls.
  • Cap same-skill repetition in one set.
• Could:
  • Show low-confidence hint.
• Risks + mitigation:
  1. Repetition fatigue -> add topic/skill cap and freshness quota.
  2. Sparse inventory -> nearest-ladder fallback.
  3. Over-control complexity -> keep controls at 3 only.

4) Validation (v1)

• H1: Context-aware ranking increases recommendation start rate.
  • Test: A/B context-aware vs baseline.
  • Pass: +8% recommendation_start_rate and no day7_active_rate drop >2%.
  • Fail: lift <3% or retention drop >2%.
• H2: Session controls reduce irrelevant skips.
  • Test: controls on vs off.
  • Pass: -6% recommendation_ignore_rate.
  • Fail: ignore unchanged or worse.
• H3: One-line reason improves trust.
  • Test: reason label on vs off.
  • Pass: +5% recommendation_item_click_through_rate.
  • Fail: <2% lift.

Self-score (1-5)

• Clarity: 4
• Simplicity: 4
• Feasible: 5
• Differentiation: 3
• Testable: 4

DoD checklist (>=8)

1. 8-10 ideas created: Dat
2. Exactly 3 directions: Dat
3. Top direction selected with 5 criteria: Dat
4. Value prop one sentence: Dat
5. User/problem explicit: Dat
6. Scope in/out explicit: Dat
7. Flow 5-7 steps: Dat
8. Must/Should/Could included: Dat
9. 3 risks + mitigation included: Dat
10. 3 hypotheses with pass/fail metrics: Dat
11. Metric mapping to goal explicit: Chua dat (missing explicit mapping table)
12. Differentiation >=4: Chua dat (score 3)

Result: iteration 1 NOT PASS (score<4 and DoD has gaps).

---

Iteration 2 (refined)

1) Updated direction positioning (to raise differentiation)

Top direction renamed: Adaptive Practice Copilot (Rule-based). Differentiation anchor:

• Competitors often provide static lists; this direction gives adaptive-but-simple guidance that stays optional, explainable, and entitlement-aware.

2) Final concept draft

• Value prop: Adaptive Practice Copilot giúp user luôn có “bước tiếp theo phù hợp nhất” theo chương trình và ngữ cảnh hiện tại, nhưng vẫn giữ toàn quyền lựa chọn.
• User/problem:
  • User segment: self-study learners + paid learners in mixed consistency states.
  • Core problem: choice overload + low confidence in what to practice next, especially after finishing one attempt.
• Scope in:
  • Program-adaptive candidate resolution.
  • Continuity-first ranking after submit.
  • Session controls: skill/difficulty/duration.
  • Explainability: one-line primary reason.
  • Fallback ladder when inventory/entitlement constraints occur.
• Scope out:
  • Mandatory curriculum.
  • New payment plans.
  • New scoring rubric.
• Core flow (7):
  1. User reaches recommendation surface (result/home/management).
  2. Context resolver reads current signals (program intent, recent attempts, streak state, entitlement).
  3. Candidate generator filters by available-now and access rights.
  4. Ranker applies: continuity -> habit-fit -> freshness/diversity -> lock teaser cap.
  5. UI renders dynamic set (3-7) + one-line reason + quick controls.
  6. User action:
     • Start item -> open attempt directly.
     • Adjust controls -> rerank in-session.
     • Ignore item(s) -> adaptive strategy shift after ignore streak threshold.
  7. Telemetry logs action/outcome; next refresh uses latest interaction memory (session TTL).
• Must:
  • Optional guidance only; no penalty.
  • Set always actionable even with thin inventory.
  • Stable and explainable ranking order.
  • Refresh triggers on submit + manual.
• Should:
  • Same-skill cap in one cluster.
  • Ignore-streak adaptation.
  • Entitlement-aware composition with minimal lock noise.
• Could:
  • Confidence badge shown only when low-confidence.
• Risks + mitigation:
  1. Risk: Low inventory causes poor relevance.
     • Mitigation: nearest-ladder fallback + transparent reason tag.
  2. Risk: User sees too many locked items and churns.
     • Mitigation: lock teaser cap, prioritize available-now first.
  3. Risk: Logic drift becomes hard to maintain.
     • Mitigation: contractized ranking order and weekly QA snapshot.

3) Validation (final)

• H1: Adaptive continuity ranking improves immediate practice continuation.
  • Test A: A/B adaptive continuity vs static mixed list.
    • Pass: +10% post_result_start_within_10m_rate.
    • Fail: <4% lift.
  • Test B: holdout cohort by program.
    • Pass: lift is positive in >=70% program cohorts.
    • Fail: positive in <50% cohorts.
• H2: Explainability + controls reduce irrelevant skipping without harming simplicity.
  • Test A: reason label + controls vs reason-only.
    • Pass: ignore_rate down >=7%, manual_refresh_per_session not up >10%.
    • Fail: ignore drop <3% or refresh spike >10%.
  • Test B: usability smoke (5-task completion).
    • Pass: >=85% complete task without assistance.
    • Fail: <75%.
• H3: Entitlement-aware composition improves conversion quality and learner trust.
  • Test A: available-now-first + lock cap vs no cap.
    • Pass: +8% upgrade_from_lock_click_rate with no rise in recommendation_bounce_rate >3%.
    • Fail: conversion flat and bounce +>3%.
  • Test B: qualitative pulse survey (in-product quick poll).
    • Pass: >=70% users choose “suggestion useful/relevant”.
    • Fail: <55%.

Metric mapping (explicit)

• User/Problem: “khong biet hoc tiep gi” -> Outcome: start nhanh sau ket qua -> Metric: post_result_start_within_10m_rate -> Experiment: H1A/H1B.
• User/Problem: “goi y khong phu hop” -> Outcome: giam bo qua -> Metric: recommendation_ignore_rate, manual_refresh_per_session -> Experiment: H2A.
• User/Problem: “bi kho chiu vi lock” -> Outcome: tang chap nhan nang cap + giu trust -> Metric: upgrade_from_lock_click_rate, recommendation_bounce_rate -> Experiment: H3A/H3B.

Self-score (1-5)

• Clarity: 5
• Simplicity: 4
• Feasible: 5
• Differentiation: 4
• Testable: 5

DoD checklist (>=8)

1. 8-10 ideas created: Dat
2. Exactly 3 directions: Dat
3. Top direction selected with impact/feasible/diff/risk/scope: Dat
4. Value prop one sentence: Dat
5. User/problem explicit: Dat
6. Scope in/out explicit: Dat
7. Flow 5-7 steps: Dat
8. Must/Should/Could included: Dat
9. 3 risks + mitigation included: Dat
10. 3 hypotheses included: Dat
11. Each hypothesis has pass/fail criteria: Dat
12. Metrics map to outcomes explicitly: Dat

Result: iteration 2 PASS.

Change log (3 lines)

1. Raised differentiation by formalizing unique “Adaptive Practice Copilot” positioning and competitor contrast.
2. Added explicit User/Problem -> Outcome -> Metric -> Experiment mapping to close testability gap.
3. Tightened risk controls with lock-teaser cap + contractized ranking order for long-term maintainability.