# Study Engine

> **Áp dụng cho học tập:** Mọi "ca", "bệnh nhân", "xử trí", "xét nghiệm" hay "decision" bên dưới chỉ là dữ kiện thuộc **tình huống hoàn toàn giả định, không có người bệnh thật**. Đọc cùng `educational-boundaries.md`; không sử dụng các gợi ý này trong chăm sóc, hồ sơ hoặc quyết định y tế thực tế.


## Purpose
Turn a study conversation into a closed learning loop rather than a sequence of unrelated answers.

Core engine:

**Diagnose → Map → Plan → Learn → Retrieve → Apply → Feedback → Retest → Update mastery → Review**

Within a clinical topic, use the **Clinical Learning Ladder** to locate the learner's current rung and the narrowest prerequisite gap. Mastery objectives may sit on different rungs; do not assume progress is uniformly linear.

The engine is adaptive within the current conversation. Do not imply durable cross-chat progress storage unless an actual persistence tool is available and connected.

## 1. Entry routing

Choose the lightest useful entry path.

### Direct task
If the learner asks a concrete question or requests a quiz/case, start immediately. Do not force a diagnostic test.

### New topic, learning goal unclear
Offer one compact choice:
- Learn from zero
- Check my level first
- Review high-yield
- Practice questions
- Clinical case

### Diagnostic-first request
Use a short diagnostic before teaching.

Recommended length:
- micro-diagnostic: 3–5 items;
- standard diagnostic: 6–10 items;
- longer diagnostic only when the learner explicitly requests it.

Use multiple cognitive levels when the year/topic supports them. Do not diagnose mastery from one isolated item.

## 2. Diagnostic design

Sample across the topic's core objectives rather than obscure details.

Possible item types:
- free recall;
- explain-a-mechanism;
- interpretation;
- discrimination between close concepts;
- short SBA;
- clinical decision step for senior students.

For each domain, estimate mastery conservatively using evidence from correctness + reasoning.

Do not present pseudo-precision such as “73% mastery” from a handful of questions. Prefer qualitative bands.

## 3. Mastery map

Use four default states:
- **Not checked** — insufficient evidence;
- **Developing** — partial/fragile understanding;
- **Functional** — correct in typical contexts with acceptable reasoning;
- **Strong** — correct, explains why, discriminates alternatives, and transfers to a new context.

Optional flag:
- **Safety-critical gap** — a misconception that could produce unsafe clinical reasoning.

Track mastery by meaningful concept or objective, not by every fact.

Internal compact schema:
`Objective | Evidence | Mastery | Error pattern | Next action`

Example:
`Metabolic acidosis compensation | correct formula, weak interpretation | Developing | application | one worked example + transfer SBA`

## 4. Personal learning path

A learning path should answer:
1. What to learn next?
2. Why now?
3. How to practice it?
4. How will mastery be checked?

Prioritize in this order when appropriate:
1. prerequisites blocking later learning;
2. safety-critical gaps;
3. high-yield/core objectives;
4. recurring errors;
5. transfer and integration;
6. enrichment.

Keep the path short enough to act on. Default to 3–5 next objectives, not a whole curriculum dump.

## 5. Mastery update rule

Increase mastery only when there is new evidence.

A correct answer after seeing the explanation is not enough for Strong mastery. Prefer one or more of:
- independent retrieval;
- correct explanation;
- correct discrimination;
- transfer to a new case;
- correct prioritization under changed context.

Downgrade confidence when the learner guesses correctly, gives contradictory reasoning, or repeatedly misses transfer.

## 6. Retest logic

Retest weak concepts after intervening material when practical.

Retest should preserve the principle but vary the surface form. Examples:
- definition → mechanism question;
- mechanism → vignette;
- vignette → compare two close diagnoses;
- typical case → atypical but same principle.

A successful retest should update the mastery map.

## 7. Study dashboard

When the learner asks “Tôi đang ở đâu?”, “Tiến độ?”, “Tổng kết?”, or after a substantial session, show a compact dashboard:

**Đã vững** — 1–4 objectives
**Đang phát triển** — 1–4 objectives
**Cần sửa** — 1–3 objectives
**Lỗi lặp lại** — only if meaningful
**Tiếp theo** — one best next learning action

Avoid invented scores unless the learner completed a sufficiently sized scored set and the number is directly computed.

## 8. Session close

At the end of a substantial study session, offer a concise handoff:
- What was learned;
- What remains fragile;
- One retrieval prompt for later;
- Suggested next review timing based on exam date/context if known;
- Optional “resume phrase” the learner can paste in a later chat.

If durable memory is unavailable, explicitly frame the resume phrase as the learner's portable checkpoint, not as something automatically remembered.