diff --git a/classes/importer_test.go b/classes/importer_test.go
index 8753204..c65f914 100644
--- a/classes/importer_test.go
+++ b/classes/importer_test.go
@@ -93,7 +93,7 @@ func TestCheckQuestion(t *testing.T) {
 		want []string
 	}{
 		"no text":           {classes.ImportQuestion{Type: "self_check"}, []string{"questions[0].text is required"}},
-		"bad type":          {classes.ImportQuestion{Text: "Q", Type: "essay"}, []string{"questions[0].type must be one of single, multiple, self_check"}},
+		"bad type":          {classes.ImportQuestion{Text: "Q", Type: "essay"}, []string{"questions[0].type must be one of single, multiple, self_check, open"}},
 		"negative points":   {classes.ImportQuestion{Text: "Q", Type: "self_check", Points: &neg}, []string{"questions[0].points must not be negative"}},
 		"single two right":  {classes.ImportQuestion{Text: "Q", Type: "single", Answers: []classes.ImportAnswer{{Text: "a", IsCorrect: true}, {Text: "b", IsCorrect: true}}}, []string{"questions[0]: a single question needs at least two answers and exactly one correct"}},
 		"multiple one item": {classes.ImportQuestion{Text: "Q", Type: "multiple", Answers: []classes.ImportAnswer{{Text: "a", IsCorrect: true}}}, []string{"questions[0]: a multiple question needs at least two answers and one or more correct"}},
diff --git a/classes/open.go b/classes/open.go
new file mode 100644
index 0000000..f831e4b
--- /dev/null
+++ b/classes/open.go
@@ -0,0 +1,33 @@
+package classes
+
+import (
+	"strings"
+	"unicode/utf8"
+
+	"git.golem15.com/golem15/sm-quizzes-plugin/models"
+)
+
+// MaxLength is the longest answer an open question accepts, in characters:
+// its MaxLength when set and positive, else models.DefaultMaxLength.
+func MaxLength(q models.Question) int {
+	if q.MaxLength != nil && *q.MaxLength > 0 {
+		return *q.MaxLength
+	}
+	return models.DefaultMaxLength
+}
+
+// CheckOpenText cleans a student's answer to an open question (D-04): it
+// removes every NUL character (Postgres text refuses them) and the
+// surrounding whitespace. A blank text is a FieldError text_required and a
+// text longer than MaxLength(q) characters (runes, not bytes) is a
+// FieldError text_too_long.
+func CheckOpenText(text string, q models.Question) (string, error) {
+	text = strings.TrimSpace(strings.ReplaceAll(text, "\x00", ""))
+	if text == "" {
+		return "", &FieldError{Field: "text", Key: "text_required"}
+	}
+	if utf8.RuneCountInString(text) > MaxLength(q) {
+		return "", &FieldError{Field: "text", Key: "text_too_long"}
+	}
+	return text, nil
+}
diff --git a/classes/open_test.go b/classes/open_test.go
new file mode 100644
index 0000000..36c0f69
--- /dev/null
+++ b/classes/open_test.go
@@ -0,0 +1,59 @@
+package classes_test
+
+import (
+	"errors"
+	"strings"
+	"testing"
+
+	"git.golem15.com/golem15/sm-quizzes-plugin/classes"
+	"git.golem15.com/golem15/sm-quizzes-plugin/models"
+)
+
+func TestCheckOpenText(t *testing.T) {
+	q := models.Question{Type: models.TypeOpen}
+	long := 3000
+	wide := models.Question{Type: models.TypeOpen, MaxLength: &long}
+
+	fieldKey := func(err error) string {
+		var fe *classes.FieldError
+		if errors.As(err, &fe) && fe.Field == "text" {
+			return fe.Key
+		}
+		return ""
+	}
+
+	if _, err := classes.CheckOpenText("  \n\t ", q); fieldKey(err) != "text_required" {
+		t.Errorf("blank text: err = %v, want text_required", err)
+	}
+	if _, err := classes.CheckOpenText("\x00 \x00", q); fieldKey(err) != "text_required" {
+		t.Errorf("NUL-only text: err = %v, want text_required", err)
+	}
+	over := strings.Repeat("a", 2001)
+	if _, err := classes.CheckOpenText(over, q); fieldKey(err) != "text_too_long" {
+		t.Errorf("2001 runes with default max: err = %v, want text_too_long", err)
+	}
+	if got, err := classes.CheckOpenText(over, wide); err != nil || got != over {
+		t.Errorf("2001 runes with max 3000: %q, %v", got[:10], err)
+	}
+	if got, err := classes.CheckOpenText("a\x00b", q); err != nil || got != "ab" {
+		t.Errorf("NUL stripped: %q, %v; want ab", got, err)
+	}
+	if got, err := classes.CheckOpenText("  moja odpowiedź \n", q); err != nil || got != "moja odpowiedź" {
+		t.Errorf("trimmed: %q, %v", got, err)
+	}
+	polish := strings.Repeat("ż", 2000)
+	if got, err := classes.CheckOpenText(polish, q); err != nil || got != polish {
+		t.Errorf("2000 runes of ż (4000 bytes) refused: %v", err)
+	}
+	five := 5
+	if _, err := classes.CheckOpenText("abcdef", models.Question{MaxLength: &five}); fieldKey(err) != "text_too_long" {
+		t.Errorf("6 runes with max 5: err = %v, want text_too_long", err)
+	}
+	if classes.MaxLength(q) != models.DefaultMaxLength || classes.MaxLength(wide) != 3000 {
+		t.Errorf("MaxLength = %d / %d", classes.MaxLength(q), classes.MaxLength(wide))
+	}
+	zero := 0
+	if classes.MaxLength(models.Question{MaxLength: &zero}) != models.DefaultMaxLength {
+		t.Error("MaxLength of a non-positive value should fall back to the default")
+	}
+}
diff --git a/classes/scoring.go b/classes/scoring.go
index 7fab7c0..f2f466f 100644
--- a/classes/scoring.go
+++ b/classes/scoring.go
@@ -152,6 +152,27 @@ func Totals(points []float64, maxPoints int) (score, percent float64) {
 	return score, round(score / float64(maxPoints) * 100)
 }
 
+// ClampOpenPoints keeps a grader's score of an open question in
+// 0..points (a negative points value counts as 0), rounded to the nearest
+// 0.5 (D-02). NaN and ±Inf are not a score: it returns 0, false.
+func ClampOpenPoints(p float64, points int) (float64, bool) {
+	if math.IsNaN(p) || math.IsInf(p, 0) {
+		return 0, false
+	}
+	full := float64(max(points, 0))
+	p = math.Round(math.Min(math.Max(p, 0), full)*2) / 2
+	if p == 0 {
+		return 0, true // no negative zero in JSON
+	}
+	return p, true
+}
+
+// OpenCorrect reports whether an open answer scored p earns full marks: a
+// question with points earns is_correct only with all of them (D-02).
+func OpenCorrect(p float64, points int) bool {
+	return points > 0 && p == float64(points)
+}
+
 // round keeps two decimals.
 func round(f float64) float64 {
 	r := math.Round(f*100) / 100
diff --git a/classes/scoring_test.go b/classes/scoring_test.go
index deb67cd..d220d3b 100644
--- a/classes/scoring_test.go
+++ b/classes/scoring_test.go
@@ -1,6 +1,7 @@
 package classes_test
 
 import (
+	"math"
 	"os"
 	"path/filepath"
 	"testing"
@@ -130,3 +131,37 @@ func TestScoringConfigFrom(t *testing.T) {
 		}
 	}
 }
+
+func TestClampOpenPoints(t *testing.T) {
+	cases := []struct {
+		p      float64
+		points int
+		want   float64
+		ok     bool
+	}{
+		{1.26, 2, 1.5, true},
+		{-3, 2, 0, true},
+		{7, 2, 2, true},
+		{0.74, 3, 0.5, true},
+		{0.25, 3, 0.5, true},
+		{0.2, 3, 0, true},
+		{2, 2, 2, true},
+		{1, 0, 0, true},
+		{5, -1, 0, true},
+		{math.NaN(), 2, 0, false},
+		{math.Inf(1), 2, 0, false},
+		{math.Inf(-1), 2, 0, false},
+	}
+	for _, tc := range cases {
+		got, ok := classes.ClampOpenPoints(tc.p, tc.points)
+		if got != tc.want || ok != tc.ok {
+			t.Errorf("ClampOpenPoints(%v, %d) = %v, %v; want %v, %v", tc.p, tc.points, got, ok, tc.want, tc.ok)
+		}
+		if math.Signbit(got) {
+			t.Errorf("ClampOpenPoints(%v, %d) is negative zero", tc.p, tc.points)
+		}
+	}
+	if !classes.OpenCorrect(2, 2) || classes.OpenCorrect(1.5, 2) || classes.OpenCorrect(0, 0) {
+		t.Error("OpenCorrect: want true only for full points of a question worth points")
+	}
+}
diff --git a/classes/service.go b/classes/service.go
index 30ee963..32dc54a 100644
--- a/classes/service.go
+++ b/classes/service.go
@@ -227,6 +227,10 @@ type Started struct {
 	AttemptsUsed int
 	// AnsweredCount is the number of questions a resumed attempt answered.
 	AnsweredCount int
+	// OpenAnswers and Unreviewed count the open answers of a resumed
+	// attempt, as in AttemptSummary.
+	OpenAnswers int
+	Unreviewed  int
 }
 
 // StartAttempt starts an attempt at a published quiz for the user, like
@@ -322,11 +326,15 @@ func (s *Service) startTest(ctx context.Context, userID uint, quiz *models.Quiz)
 			return err
 		}
 		if resume != nil {
-			var answered int64
-			if err := tx.Model(&models.AttemptAnswer{}).Where("attempt_id = ?", resume.ID).Count(&answered).Error; err != nil {
+			counts, err := answerCounts(ctx, tx, []uint{resume.ID})
+			if err != nil {
 				return err
 			}
-			st = &Started{Attempt: resume, Quiz: quiz, Resumed: true, AttemptsUsed: int(used), AnsweredCount: int(answered)}
+			c := counts[resume.ID]
+			st = &Started{
+				Attempt: resume, Quiz: quiz, Resumed: true, AttemptsUsed: int(used),
+				AnsweredCount: c.N, OpenAnswers: c.OpenN, Unreviewed: c.UnreviewedN,
+			}
 			return nil
 		}
 		if left := AttemptsLeft(quiz, int(used)); left != nil && *left == 0 {
@@ -403,6 +411,8 @@ type AnswerInput struct {
 	QuestionID uint
 	AnswerIDs  []uint
 	Known      *bool
+	// Text is the answer to an open question.
+	Text *string
 }
 
 // Feedback is what the user learns after answering one question.
@@ -414,6 +424,9 @@ type Feedback struct {
 	// and the serializer must drop Result, CorrectAnswerIDs and every
 	// solution field of Question.
 	Withheld bool
+	// Answer is the stored row of an open question, with its text and
+	// grading state; it is nil for every other type.
+	Answer *models.AttemptAnswer
 }
 
 // Answer records or overwrites the user's answer to one question of an
@@ -423,8 +436,11 @@ type Feedback struct {
 // question belongs to the attempt (FieldError on question_id), a one_way
 // test has no answer to the question yet (ErrAnswerLocked; the attempt row
 // is locked, so the check is race-free), and the answer fits the question
-// type (FieldError on answer_ids or known). The feedback of an open test
-// attempt is Withheld.
+// type (FieldError on answer_ids, known or text). The feedback of an open
+// test attempt is Withheld.
+//
+// An open question stores its text as pending (see answerOpen) and is never
+// scored here: a grader or the owner scores it later.
 func (s *Service) Answer(ctx context.Context, userID, attemptID uint, in AnswerInput) (*Feedback, error) {
 	var fb *Feedback
 	err := lagoon.Transaction(ctx, s.DB, func(ctx context.Context, tx *gorm.DB) error {
@@ -464,6 +480,20 @@ func (s *Service) Answer(ctx context.Context, userID, attemptID uint, in AnswerI
 
 		row := models.AttemptAnswer{AttemptID: attempt.ID, QuestionID: question.ID}
 		switch question.Type {
+		case models.TypeOpen:
+			if in.Text == nil {
+				return &FieldError{Field: "text", Key: "text_required"}
+			}
+			text, err := CheckOpenText(*in.Text, question)
+			if err != nil {
+				return err
+			}
+			stored, err := s.answerOpen(ctx, tx, attempt, question, text)
+			if err != nil {
+				return err
+			}
+			fb = &Feedback{Question: question, Answer: stored, Withheld: Withheld(attempt)}
+			return nil
 		case models.TypeSelfCheck:
 			if in.Known == nil {
 				return &FieldError{Field: "known", Key: "known_required"}
@@ -512,6 +542,78 @@ func (s *Service) Answer(ctx context.Context, userID, attemptID uint, in AnswerI
 	return fb, nil
 }
 
+// answerOpen stores a cleaned open answer in a transaction that holds the
+// attempt row lock, so reading the existing row first is race-free. The
+// same text again changes nothing. Another text is a new answer (Pitfall 1):
+// the row goes back to pending with the next grading_seq, and every grade,
+// error, review and point of the old text is cleared, so a learning
+// re-answer replaces an earlier grade or review. A first answer creates the
+// row as pending with seq 1. A new or changed text is handed to
+// enqueueGrading.
+func (s *Service) answerOpen(ctx context.Context, tx *gorm.DB, attempt *models.Attempt, question models.Question, text string) (*models.AttemptAnswer, error) {
+	tx = tx.WithContext(ctx)
+	now := s.now().UTC()
+	var existing models.AttemptAnswer
+	err := tx.Where("attempt_id = ? AND question_id = ?", attempt.ID, question.ID).Take(&existing).Error
+	switch {
+	case err == nil:
+		if existing.TextAnswer != nil && *existing.TextAnswer == text && existing.GradingStatus != nil {
+			return &existing, nil
+		}
+		seq := existing.GradingSeq + 1
+		if err := tx.Model(&models.AttemptAnswer{}).Where("id = ?", existing.ID).Updates(map[string]any{
+			"text_answer":          text,
+			"answer_ids":           nil,
+			"self_check_known":     nil,
+			"grading_status":       models.GradingPending,
+			"grading_seq":          seq,
+			"grading_attempts":     0,
+			"grader":               nil,
+			"grader_points":        nil,
+			"grader_feedback":      nil,
+			"grading_error":        nil,
+			"graded_at":            nil,
+			"review_comment":       nil,
+			"reviewed_at":          nil,
+			"reviewed_by_user_id":  nil,
+			"reviewed_by_admin_id": nil,
+			"points":               0,
+			"is_correct":           false,
+			"updated_at":           now,
+		}).Error; err != nil {
+			return nil, err
+		}
+		if err := tx.Where("id = ?", existing.ID).Take(&existing).Error; err != nil {
+			return nil, err
+		}
+		if err := s.enqueueGrading(ctx, tx, existing.ID, seq, attempt); err != nil {
+			return nil, err
+		}
+		return &existing, nil
+	case errors.Is(err, gorm.ErrRecordNotFound):
+	default:
+		return nil, err
+	}
+	pending := models.GradingPending
+	row := &models.AttemptAnswer{
+		AttemptID: attempt.ID, QuestionID: question.ID,
+		TextAnswer: &text, GradingStatus: &pending, GradingSeq: 1,
+		CreatedAt: now, UpdatedAt: now,
+	}
+	if err := tx.Create(row).Error; err != nil {
+		return nil, err
+	}
+	if err := s.enqueueGrading(ctx, tx, row.ID, row.GradingSeq, attempt); err != nil {
+		return nil, err
+	}
+	return row, nil
+}
+
+// enqueueGrading hands a new or changed open answer to the quiz's grader.
+func (s *Service) enqueueGrading(ctx context.Context, tx *gorm.DB, answerID uint, seq int, attempt *models.Attempt) error {
+	return nil
+}
+
 // CorrectAnswerIDs lists the IDs of a question's correct answers in sort
 // order; it is empty for a self_check question.
 func CorrectAnswerIDs(q models.Question) []uint {
@@ -574,6 +676,50 @@ func (s *Service) close(ctx context.Context, tx *gorm.DB, attempt *models.Attemp
 type AttemptSummary struct {
 	Attempt       models.Attempt
 	AnsweredCount int
+	// OpenAnswers counts the attempt's answers to open questions (rows
+	// with a grading status).
+	OpenAnswers int
+	// Unreviewed counts the open answers the owner has not reviewed yet:
+	// pending, graded or failed.
+	Unreviewed int
+}
+
+// answerCountRow is the per-attempt answer counts answerCounts reads.
+type answerCountRow struct {
+	AttemptID   uint
+	N           int
+	OpenN       int
+	UnreviewedN int
+}
+
+// answerCounts counts, per attempt, the answers, the open answers and the
+// open answers not reviewed yet.
+func answerCounts(ctx context.Context, db *gorm.DB, attemptIDs []uint) (map[uint]answerCountRow, error) {
+	var rows []answerCountRow
+	if err := db.WithContext(ctx).Model(&models.AttemptAnswer{}).
+		Select("attempt_id, COUNT(*) AS n, "+
+			"COUNT(*) FILTER (WHERE grading_status IS NOT NULL) AS open_n, "+
+			"COUNT(*) FILTER (WHERE grading_status IN ('pending', 'graded', 'failed')) AS unreviewed_n").
+		Where("attempt_id IN ?", attemptIDs).Group("attempt_id").Scan(&rows).Error; err != nil {
+		return nil, err
+	}
+	out := make(map[uint]answerCountRow, len(rows))
+	for _, r := range rows {
+		out[r.AttemptID] = r
+	}
+	return out, nil
+}
+
+// unreviewed reports whether an open answer still waits for the owner.
+func unreviewed(a *models.AttemptAnswer) bool {
+	if a == nil || a.GradingStatus == nil {
+		return false
+	}
+	switch *a.GradingStatus {
+	case models.GradingPending, models.GradingGraded, models.GradingFailed:
+		return true
+	}
+	return false
 }
 
 // QuestionResult is one question of an attempt with the user's answer.
@@ -638,6 +784,14 @@ func (s *Service) Result(ctx context.Context, userID, attemptID uint) (*AttemptR
 		byQuestion[answers[i].QuestionID] = &answers[i]
 	}
 	res := &AttemptResult{AttemptSummary: AttemptSummary{Attempt: *attempt, AnsweredCount: len(answers)}}
+	for i := range answers {
+		if answers[i].GradingStatus != nil {
+			res.OpenAnswers++
+		}
+		if unreviewed(&answers[i]) {
+			res.Unreviewed++
+		}
+	}
 	for _, q := range questions {
 		res.Questions = append(res.Questions, QuestionResult{Question: q, Answer: byQuestion[q.ID]})
 	}
@@ -668,21 +822,13 @@ func (s *Service) History(ctx context.Context, userID uint, quizSlug string, pag
 	for i, a := range attempts {
 		ids[i] = a.ID
 	}
-	type countRow struct {
-		AttemptID uint
-		N         int
-	}
-	var counts []countRow
-	if err := s.db(ctx).Model(&models.AttemptAnswer{}).Select("attempt_id, COUNT(*) AS n").
-		Where("attempt_id IN ?", ids).Group("attempt_id").Scan(&counts).Error; err != nil {
+	countBy, err := answerCounts(ctx, s.DB, ids)
+	if err != nil {
 		return nil, 0, err
 	}
-	countBy := map[uint]int{}
-	for _, c := range counts {
-		countBy[c.AttemptID] = c.N
-	}
 	for _, a := range attempts {
-		out = append(out, AttemptSummary{Attempt: a, AnsweredCount: countBy[a.ID]})
+		c := countBy[a.ID]
+		out = append(out, AttemptSummary{Attempt: a, AnsweredCount: c.N, OpenAnswers: c.OpenN, Unreviewed: c.UnreviewedN})
 	}
 	return out, total, nil
 }
diff --git a/controllers/api/handlers.go b/controllers/api/handlers.go
index 4ac209f..ce73521 100644
--- a/controllers/api/handlers.go
+++ b/controllers/api/handlers.go
@@ -41,6 +41,7 @@ var (
 		{Field: "answer_ids", Rules: lagoon.ParseRules("nullable|array")},
 		{Field: "answer_ids.*", Rules: lagoon.ParseRules("integer|min:1")},
 		{Field: "known", Rules: lagoon.ParseRules("nullable|boolean")},
+		{Field: "text", Rules: lagoon.ParseRules("nullable|string")},
 	}
 )
 
@@ -116,7 +117,8 @@ func (h *Handlers) StartAttempt(w http.ResponseWriter, r *http.Request) {
 }
 
 // Answer answers POST /attempts/{id}/answers with
-// {"question_id": 1, "answer_ids": [2], "known": true}.
+// {"question_id": 1, "answer_ids": [2], "known": true}; an open question
+// takes {"question_id": 1, "text": "..."}.
 func (h *Handlers) Answer(w http.ResponseWriter, r *http.Request) {
 	userID, ok := h.user(w, r)
 	if !ok {
@@ -147,6 +149,10 @@ func (h *Handlers) Answer(w http.ResponseWriter, r *http.Request) {
 		known := toBool(raw)
 		in.Known = &known
 	}
+	if raw, present := input["text"]; present && raw != nil {
+		text, _ := raw.(string) // the string rule already refused anything else
+		in.Text = &text
+	}
 	fb, err := svc.Answer(r.Context(), userID, attemptID, in)
 	if err != nil {
 		h.writeError(w, r, err)
@@ -331,6 +337,8 @@ var apiMessages = map[string]string{
 	"answer_not_in_question": "The chosen answers do not belong to this question.",
 	"single_one_answer":      "Choose exactly one answer for this question.",
 	"known_required":         "The known field is required for this question.",
+	"text_required":          "Write an answer to this question.",
+	"text_too_long":          "The answer is longer than this question allows.",
 	"unknown_category":       "The quiz has no questions in this category.",
 	"validation_failed":      "The given data was invalid.",
 	"attempt_limit_reached":  "All attempts at this test have been used.",
diff --git a/controllers/api/json.go b/controllers/api/json.go
index 06788bc..346e177 100644
--- a/controllers/api/json.go
+++ b/controllers/api/json.go
@@ -103,7 +103,8 @@ type quizJSON struct {
 
 // questionJSON is a question as the player sees it before answering: no
 // is_correct or explanation, and no answer_text except for self_check
-// flashcards, which have no solution to hide.
+// flashcards, which have no solution to hide. An open question carries its
+// max_length instead (never its rubric or model answer).
 type questionJSON struct {
 	ID        uint    `json:"id"`
 	Text      string  `json:"text"`
@@ -112,8 +113,10 @@ type questionJSON struct {
 	Points    int     `json:"points"`
 	SortOrder int     `json:"sort_order"`
 	// AnswerText is the model answer of a self_check question only.
-	AnswerText *string      `json:"answer_text,omitempty"`
-	Answers    []answerJSON `json:"answers"`
+	AnswerText *string `json:"answer_text,omitempty"`
+	// MaxLength is the longest answer of an open question, in characters.
+	MaxLength *int         `json:"max_length,omitempty"`
+	Answers   []answerJSON `json:"answers"`
 }
 
 type answerJSON struct {
@@ -162,6 +165,10 @@ func question(q models.Question) questionJSON {
 	if q.Type == models.TypeSelfCheck {
 		out.AnswerText = q.AnswerText
 	}
+	if q.Type == models.TypeOpen {
+		n := classes.MaxLength(q)
+		out.MaxLength = &n
+	}
 	for _, a := range q.Answers {
 		out.Answers = append(out.Answers, answerJSON{ID: a.ID, Text: a.Text})
 	}
@@ -191,6 +198,11 @@ type attemptJSON struct {
 	// SecondsLeft is the time left of an open attempt with a limit, counted
 	// by the server.
 	SecondsLeft *int `json:"seconds_left"`
+	// NeedsReview is true while an open answer of the attempt waits for
+	// the owner (pending, graded or failed); OpenAnswers counts its open
+	// answers.
+	NeedsReview bool `json:"needs_review"`
+	OpenAnswers int  `json:"open_answers"`
 }
 
 // timeNow is the clock of seconds_left; tests pin it.
@@ -209,6 +221,8 @@ func attempt(s classes.AttemptSummary) attemptJSON {
 		ScorePercent:   a.ScorePercent,
 		QuizMode:       quizMode(a.Quiz),
 		QuizNavigation: quizNavigation(a.Quiz),
+		NeedsReview:    s.Unreviewed > 0,
+		OpenAnswers:    s.OpenAnswers,
 	}
 	if a.Quiz != nil {
 		out.QuizSlug, out.QuizTitle, out.PassingScore = a.Quiz.Slug, a.Quiz.Title, a.Quiz.PassingScore
@@ -240,7 +254,9 @@ func start(st *classes.Started) startJSON {
 	a := *st.Attempt
 	a.Quiz = st.Quiz
 	return startJSON{
-		attemptJSON:  attempt(classes.AttemptSummary{Attempt: a, AnsweredCount: st.AnsweredCount}),
+		attemptJSON: attempt(classes.AttemptSummary{
+			Attempt: a, AnsweredCount: st.AnsweredCount, OpenAnswers: st.OpenAnswers, Unreviewed: st.Unreviewed,
+		}),
 		Resumed:      st.Resumed,
 		AttemptsLeft: classes.AttemptsLeft(st.Quiz, st.AttemptsUsed),
 	}
@@ -261,6 +277,11 @@ type resultJSON struct {
 // withheld (an open test attempt) every solution field, is_correct and
 // points_earned included, is null for every question; answered, answer_ids
 // and known (the student's own input) stay.
+//
+// An open question adds the student's text_answer (always: it is their own
+// input), max_length and, unless withheld, grading_status, grader_feedback
+// and review_comment; its is_correct and points_earned stay null until the
+// answer is graded or reviewed. The rubric is never part of it.
 type questionResultJSON struct {
 	ID               uint               `json:"id"`
 	Text             string             `json:"text"`
@@ -276,6 +297,14 @@ type questionResultJSON struct {
 	Explanation      *string            `json:"explanation"`
 	AnswerText       *string            `json:"answer_text"`
 	Answers          []resultAnswerJSON `json:"answers"`
+	// TextAnswer is the student's text of an open question. It is not
+	// "text" as in the feedback: text is already the question's text here.
+	TextAnswer     *string `json:"text_answer"`
+	GradingStatus  *string `json:"grading_status"`
+	GraderFeedback *string `json:"grader_feedback"`
+	ReviewComment  *string `json:"review_comment"`
+	// MaxLength is the longest answer of an open question.
+	MaxLength *int `json:"max_length,omitempty"`
 }
 
 type resultAnswerJSON struct {
@@ -304,11 +333,24 @@ func result(r *classes.AttemptResult) resultJSON {
 			AnswerIDs: []uint{},
 			Answers:   make([]resultAnswerJSON, 0, len(q.Answers)),
 		}
+		if q.Type == models.TypeOpen {
+			n := classes.MaxLength(q)
+			row.MaxLength = &n
+		}
 		if a := qr.Answer; a != nil {
 			row.Answered = true
 			row.AnswerIDs = wire.Slice(a.AnswerIDs.Data)
 			row.Known = a.SelfCheckKnown
-			if !withheld {
+			if q.Type == models.TypeOpen {
+				row.TextAnswer = a.TextAnswer
+				if !withheld {
+					row.GradingStatus, row.GraderFeedback, row.ReviewComment = a.GradingStatus, a.GraderFeedback, a.ReviewComment
+					if openScored(a) {
+						correct, points := a.IsCorrect, a.Points
+						row.IsCorrect, row.PointsEarned = &correct, &points
+					}
+				}
+			} else if !withheld {
 				correct, points := a.IsCorrect, a.Points
 				row.IsCorrect, row.PointsEarned = &correct, &points
 			}
@@ -340,6 +382,12 @@ func result(r *classes.AttemptResult) resultJSON {
 // feedbackJSON is POST /attempts/{id}/answers. When Withheld (an answer to
 // an open test attempt) every solution field is null and the answers carry
 // only their id and text.
+//
+// The feedback of an open question carries the student's text and, unless
+// withheld, the model answer (answer_text), the explanation, grading_status,
+// grader_feedback and review_comment; is_correct and points stay null until
+// the answer is graded or reviewed. These four keys are null for every other
+// type. The rubric is never part of it.
 type feedbackJSON struct {
 	QuestionID       uint               `json:"question_id"`
 	Withheld         bool               `json:"withheld"`
@@ -349,11 +397,18 @@ type feedbackJSON struct {
 	Explanation      *string            `json:"explanation"`
 	AnswerText       *string            `json:"answer_text"`
 	Answers          []resultAnswerJSON `json:"answers"`
+	Text             *string            `json:"text"`
+	GradingStatus    *string            `json:"grading_status"`
+	GraderFeedback   *string            `json:"grader_feedback"`
+	ReviewComment    *string            `json:"review_comment"`
 }
 
 func feedback(fb *classes.Feedback) feedbackJSON {
 	q := fb.Question
 	out := feedbackJSON{QuestionID: q.ID, Withheld: fb.Withheld, Answers: make([]resultAnswerJSON, 0, len(q.Answers))}
+	if q.Type == models.TypeOpen {
+		return openFeedback(fb, out)
+	}
 	if fb.Withheld {
 		for _, a := range q.Answers {
 			out.Answers = append(out.Answers, resultAnswerJSON{ID: a.ID, Text: a.Text})
@@ -371,4 +426,36 @@ func feedback(fb *classes.Feedback) feedbackJSON {
 	return out
 }
 
+// openFeedback fills the feedback of an open question from its stored row.
+func openFeedback(fb *classes.Feedback, out feedbackJSON) feedbackJSON {
+	q, a := fb.Question, fb.Answer
+	if a != nil {
+		out.Text = a.TextAnswer
+	}
+	if fb.Withheld {
+		return out
+	}
+	out.CorrectAnswerIDs = []uint{}
+	out.Explanation, out.AnswerText = q.Explanation, q.AnswerText
+	if a == nil {
+		return out
+	}
+	out.GradingStatus, out.GraderFeedback, out.ReviewComment = a.GradingStatus, a.GraderFeedback, a.ReviewComment
+	if openScored(a) {
+		correct, points := a.IsCorrect, a.Points
+		out.IsCorrect, out.Points = &correct, &points
+	}
+	return out
+}
+
+// openScored reports whether an open answer has a score to show: graded
+// by a grader or reviewed by the owner. Pending and failed answers have
+// none yet.
+func openScored(a *models.AttemptAnswer) bool {
+	if a == nil || a.GradingStatus == nil {
+		return false
+	}
+	return *a.GradingStatus == models.GradingGraded || *a.GradingStatus == models.GradingReviewed
+}
+
 func wireTime(t time.Time) wire.Time { return wire.Time{Time: t.UTC().Truncate(time.Second)} }
diff --git a/controllers/api/json_test.go b/controllers/api/json_test.go
index afcfaf2..5686a6a 100644
--- a/controllers/api/json_test.go
+++ b/controllers/api/json_test.go
@@ -16,7 +16,11 @@ import (
 func fixtureQuestion(id uint, typ string) models.Question {
 	expl, model, why := "question explanation", "model answer", "answer explanation"
 	q := models.Question{ID: id, Text: "Q", Type: typ, Points: 2, Explanation: &expl, AnswerText: &model}
-	if typ != models.TypeSelfCheck {
+	if typ == models.TypeOpen {
+		rubric := "R-SECRET"
+		q.Rubric = &rubric
+	}
+	if isChoice(typ) {
 		q.Answers = []models.Answer{
 			{ID: id*10 + 1, QuestionID: id, Text: "right", IsCorrect: true, Explanation: &why},
 			{ID: id*10 + 2, QuestionID: id, Text: "wrong", Explanation: &why},
@@ -25,6 +29,28 @@ func fixtureQuestion(id uint, typ string) models.Question {
 	return q
 }
 
+func isChoice(typ string) bool { return typ == models.TypeSingle || typ == models.TypeMultiple }
+
+func strPtr(s string) *string { return &s }
+
+// noRubric fails when a student shape names or contains the rubric or the
+// grading error.
+func noRubric(t *testing.T, where, raw string) {
+	t.Helper()
+	if strings.Contains(raw, `"rubric"`) || strings.Contains(raw, "R-SECRET") || strings.Contains(raw, "grading_error") || strings.Contains(raw, "provider exploded") {
+		t.Errorf("%s leaks the rubric or the grading error: %s", where, raw)
+	}
+}
+
+// openRow is a stored open answer in the given grading status.
+func openRow(questionID uint, status string) *models.AttemptAnswer {
+	return &models.AttemptAnswer{
+		AttemptID: 5, QuestionID: questionID,
+		TextAnswer: strPtr("moja odpowiedź"), GradingStatus: strPtr(status), GradingSeq: 1,
+		GradingError: strPtr("provider exploded"),
+	}
+}
+
 func decode(t *testing.T, v any) (map[string]any, string) {
 	t.Helper()
 	b, err := json.Marshal(v)
@@ -83,21 +109,43 @@ func TestWithheldSerializers(t *testing.T) {
 				CorrectAnswerIDs: classes.CorrectAnswerIDs(answered),
 				Withheld:         true,
 			}
+			// gradedOpen is an open answer with every grading field set, so
+			// a serializer that forgets to withhold one shows it.
+			gradedOpen := func() *models.AttemptAnswer {
+				r := openRow(1, models.GradingGraded)
+				two := 2.0
+				r.IsCorrect, r.Points, r.GraderPoints = true, 2, &two
+				r.GraderFeedback, r.ReviewComment = strPtr("feedback"), strPtr("comment")
+				return r
+			}
+			if typ == models.TypeOpen {
+				fb.Answer = gradedOpen()
+			}
 			m, raw := decode(t, feedback(fb))
 			if m["withheld"] != true {
 				t.Errorf("feedback withheld = %v, want true", m["withheld"])
 			}
-			noSolution(t, "feedback", m, "is_correct", "points", "correct_answer_ids", "explanation", "answer_text")
+			noSolution(t, "feedback", m, "is_correct", "points", "correct_answer_ids", "explanation", "answer_text",
+				"grading_status", "grader_feedback", "review_comment")
 			noCorrectness(t, "feedback", raw)
-			if typ != models.TypeSelfCheck && len(m["answers"].([]any)) != 2 {
+			noRubric(t, "feedback", raw)
+			if isChoice(typ) && len(m["answers"].([]any)) != 2 {
 				t.Errorf("feedback answers = %v, want both answers", m["answers"])
 			}
+			if typ == models.TypeOpen && m["text"] != "moja odpowiedź" {
+				t.Errorf("withheld open feedback lost the student's text: %v", m["text"])
+			} else if typ != models.TypeOpen && m["text"] != nil {
+				t.Errorf("%s feedback text = %v, want null", typ, m["text"])
+			}
 
 			known := true
 			row := &models.AttemptAnswer{AttemptID: 5, QuestionID: 1, IsCorrect: true, Points: 2}
-			if typ == models.TypeSelfCheck {
+			switch typ {
+			case models.TypeSelfCheck:
 				row.SelfCheckKnown = &known
-			} else {
+			case models.TypeOpen:
+				row = gradedOpen()
+			default:
 				row.AnswerIDs = lagoon.Jsonable[[]uint]{Data: []uint{11}, Valid: true}
 			}
 			res := &classes.AttemptResult{
@@ -109,24 +157,33 @@ func TestWithheldSerializers(t *testing.T) {
 				t.Errorf("result solutions_withheld = %v quiz_mode = %v", m["solutions_withheld"], m["quiz_mode"])
 			}
 			noCorrectness(t, "result", raw)
+			noRubric(t, "result", raw)
 			questions := m["questions"].([]any)
 			if len(questions) != 2 {
 				t.Fatalf("result questions = %d, want 2", len(questions))
 			}
 			for i, raw := range questions {
 				q := raw.(map[string]any)
-				noSolution(t, "result question "+string(rune('0'+i)), q, "is_correct", "points_earned", "correct_answer_ids", "explanation", "answer_text")
+				noSolution(t, "result question "+string(rune('0'+i)), q, "is_correct", "points_earned", "correct_answer_ids", "explanation", "answer_text",
+					"grading_status", "grader_feedback", "review_comment")
 			}
 			first := questions[0].(map[string]any)
 			if first["answered"] != true {
 				t.Errorf("answered question lost answered: %v", first)
 			}
-			if typ == models.TypeSelfCheck {
+			switch typ {
+			case models.TypeSelfCheck:
 				if first["known"] != true {
 					t.Errorf("self_check lost the student's known: %v", first)
 				}
-			} else if ids, _ := first["answer_ids"].([]any); len(ids) != 1 || ids[0] != float64(11) {
-				t.Errorf("answer_ids = %v, want the student's [11]", first["answer_ids"])
+			case models.TypeOpen:
+				if first["text_answer"] != "moja odpowiedź" {
+					t.Errorf("open lost the student's text: %v", first)
+				}
+			default:
+				if ids, _ := first["answer_ids"].([]any); len(ids) != 1 || ids[0] != float64(11) {
+					t.Errorf("answer_ids = %v, want the student's [11]", first["answer_ids"])
+				}
 			}
 
 			done := time.Now()
@@ -207,15 +264,15 @@ func TestSummaryAndAttemptFields(t *testing.T) {
 		t.Errorf("test summary = %v", m)
 	}
 	types, _ := m["question_types"].(map[string]any)
-	if len(types) != 3 || types["single"] != float64(3) || types["multiple"] != float64(0) || types["self_check"] != float64(0) {
-		t.Errorf("question_types = %v, want all three keys", m["question_types"])
+	if len(types) != 4 || types["single"] != float64(3) || types["multiple"] != float64(0) || types["self_check"] != float64(0) || types["open"] != float64(0) {
+		t.Errorf("question_types = %v, want all four keys", m["question_types"])
 	}
 	m, _ = decode(t, quizSummary(classes.QuizSummary{Quiz: models.Quiz{ID: 4, Slug: "l"}, AttemptsUsed: 9}))
 	if v, ok := m["attempts_left"]; !ok || v != nil {
 		t.Errorf("learning attempts_left = %v (present %v), want null", v, ok)
 	}
-	if types, _ := m["question_types"].(map[string]any); len(types) != 3 {
-		t.Errorf("learning question_types = %v, want all three keys", m["question_types"])
+	if types, _ := m["question_types"].(map[string]any); len(types) != 4 {
+		t.Errorf("learning question_types = %v, want all four keys", m["question_types"])
 	}
 	if m["mode"] != models.ModeLearning || m["navigation"] != models.NavigationFree {
 		t.Errorf("zero-value quiz mode / navigation = %v / %v", m["mode"], m["navigation"])
@@ -241,3 +298,119 @@ func TestSummaryAndAttemptFields(t *testing.T) {
 		t.Errorf("start json = %v", m)
 	}
 }
+
+func TestOpenSerializers(t *testing.T) {
+	q := fixtureQuestion(1, models.TypeOpen)
+	q.AnswerText = strPtr("MODEL")
+	learning := &models.Quiz{Slug: "l", PassingScore: 50}
+
+	// Feedback of a pending learning answer: the model answer and the
+	// student's text, no score yet, never the rubric.
+	m, raw := decode(t, feedback(&classes.Feedback{Question: q, Answer: openRow(1, models.GradingPending)}))
+	noRubric(t, "pending feedback", raw)
+	if m["answer_text"] != "MODEL" || m["text"] != "moja odpowiedź" || m["grading_status"] != models.GradingPending {
+		t.Errorf("pending feedback = %s", raw)
+	}
+	if m["is_correct"] != nil || m["points"] != nil {
+		t.Errorf("pending feedback has a score: %s", raw)
+	}
+	if a, ok := m["answers"].([]any); !ok || len(a) != 0 {
+		t.Errorf("open feedback answers = %v, want []", m["answers"])
+	}
+	if ids, ok := m["correct_answer_ids"].([]any); !ok || len(ids) != 0 {
+		t.Errorf("open feedback correct_answer_ids = %v, want []", m["correct_answer_ids"])
+	}
+
+	m, raw = decode(t, feedback(&classes.Feedback{Question: q, Answer: openRow(1, models.GradingFailed)}))
+	noRubric(t, "failed feedback", raw)
+	if m["grading_status"] != models.GradingFailed || m["is_correct"] != nil || m["points"] != nil {
+		t.Errorf("failed feedback = %s", raw)
+	}
+
+	graded := openRow(1, models.GradingGraded)
+	graded.Points, graded.IsCorrect, graded.GraderFeedback = 1.5, false, strPtr("Brakuje przykładu.")
+	m, raw = decode(t, feedback(&classes.Feedback{Question: q, Answer: graded}))
+	noRubric(t, "graded feedback", raw)
+	if m["points"] != 1.5 || m["is_correct"] != false || m["grader_feedback"] != "Brakuje przykładu." || m["review_comment"] != nil {
+		t.Errorf("graded feedback = %s", raw)
+	}
+
+	reviewed := openRow(1, models.GradingReviewed)
+	reviewed.Points, reviewed.IsCorrect, reviewed.ReviewComment = 2, true, strPtr("Dobrze.")
+	m, raw = decode(t, feedback(&classes.Feedback{Question: q, Answer: reviewed}))
+	if m["points"] != float64(2) || m["is_correct"] != true || m["review_comment"] != "Dobrze." || m["grading_status"] != models.GradingReviewed {
+		t.Errorf("reviewed feedback = %s", raw)
+	}
+
+	// The question before answering: max_length, no model answer, no rubric.
+	m, raw = decode(t, question(q))
+	noRubric(t, "open question", raw)
+	if m["max_length"] != float64(models.DefaultMaxLength) {
+		t.Errorf("open question max_length = %v, want %d", m["max_length"], models.DefaultMaxLength)
+	}
+	if _, ok := m["answer_text"]; ok {
+		t.Errorf("open question leaks its model answer: %s", raw)
+	}
+	five := 5
+	q5 := q
+	q5.MaxLength = &five
+	if m, _ = decode(t, question(q5)); m["max_length"] != float64(5) {
+		t.Errorf("open question max_length = %v, want 5", m["max_length"])
+	}
+	if _, raw = decode(t, question(fixtureQuestion(2, models.TypeSingle))); strings.Contains(raw, "max_length") {
+		t.Errorf("single question names max_length: %s", raw)
+	}
+	_, raw = decode(t, quizDetail(&classes.QuizDetail{Quiz: *learning, Questions: []models.Question{q}}))
+	noRubric(t, "quiz detail", raw)
+
+	// The result of a learning attempt with a pending open answer.
+	res := &classes.AttemptResult{
+		AttemptSummary: classes.AttemptSummary{
+			Attempt:       models.Attempt{ID: 5, StartedAt: time.Now(), Quiz: learning},
+			AnsweredCount: 1, OpenAnswers: 1, Unreviewed: 1,
+		},
+		Questions: []classes.QuestionResult{{Question: q, Answer: openRow(1, models.GradingPending)}, {Question: fixtureQuestion(2, models.TypeSingle)}},
+	}
+	m, raw = decode(t, result(res))
+	noRubric(t, "result", raw)
+	if m["needs_review"] != true || m["open_answers"] != float64(1) {
+		t.Errorf("result needs_review / open_answers = %v / %v", m["needs_review"], m["open_answers"])
+	}
+	row := m["questions"].([]any)[0].(map[string]any)
+	if row["points_earned"] != nil || row["is_correct"] != nil {
+		t.Errorf("pending open result has a score: %v", row)
+	}
+	if row["answer_text"] != "MODEL" || row["text_answer"] != "moja odpowiedź" || row["text"] != "Q" || row["grading_status"] != models.GradingPending {
+		t.Errorf("pending open result = %v", row)
+	}
+	if row["max_length"] != float64(models.DefaultMaxLength) {
+		t.Errorf("open result max_length = %v", row["max_length"])
+	}
+	choice := m["questions"].([]any)[1].(map[string]any)
+	for _, k := range []string{"text_answer", "grading_status", "grader_feedback", "review_comment"} {
+		if v, ok := choice[k]; !ok || v != nil {
+			t.Errorf("single result %s = %v (present %v), want null", k, v, ok)
+		}
+	}
+	if _, ok := choice["max_length"]; ok {
+		t.Errorf("single result names max_length: %v", choice)
+	}
+
+	res.Questions[0].Answer = reviewed
+	res.Unreviewed = 0
+	m, _ = decode(t, result(res))
+	row = m["questions"].([]any)[0].(map[string]any)
+	if row["points_earned"] != float64(2) || row["is_correct"] != true || row["review_comment"] != "Dobrze." || m["needs_review"] != false {
+		t.Errorf("reviewed open result = %v (needs_review %v)", row, m["needs_review"])
+	}
+
+	// Every attempt row carries the two review keys.
+	m, _ = decode(t, attempt(classes.AttemptSummary{Attempt: models.Attempt{ID: 1, StartedAt: time.Now(), Quiz: learning}}))
+	if m["needs_review"] != false || m["open_answers"] != float64(0) {
+		t.Errorf("attempt without open answers = %v / %v", m["needs_review"], m["open_answers"])
+	}
+	m, _ = decode(t, start(&classes.Started{Attempt: &models.Attempt{ID: 2, StartedAt: time.Now()}, Quiz: learning, OpenAnswers: 2, Unreviewed: 1}))
+	if m["needs_review"] != true || m["open_answers"] != float64(2) {
+		t.Errorf("start needs_review / open_answers = %v / %v", m["needs_review"], m["open_answers"])
+	}
+}
diff --git a/graders/graders.go b/graders/graders.go
new file mode 100644
index 0000000..12a0b42
--- /dev/null
+++ b/graders/graders.go
@@ -0,0 +1,62 @@
+// Package graders is the grader registry of golem15.quizzes. A plugin that
+// grades open answers implements HasGraders, and golem15.quizzes collects
+// its graders at boot. Manual is built in and is never called: an answer of
+// a manual quiz waits for the owner. The package imports only the standard
+// library, so a grading plugin can depend on it without pulling in the rest
+// of golem15.quizzes.
+package graders
+
+import "context"
+
+// Manual is the built-in grader code: the owner grades every open answer.
+const Manual = "manual"
+
+// ManualLabel is the lang key of Manual in the admin grader dropdown.
+const ManualLabel = "golem15.quizzes::lang.graders.manual"
+
+// Input is one open answer to grade.
+type Input struct {
+	QuizID     uint
+	QuestionID uint
+	AttemptID  uint
+	AnswerID   uint
+	UserID     uint
+	// Question is the owner-authored question text.
+	Question string
+	// ModelAnswer is the owner-authored model answer; it may be empty.
+	ModelAnswer string
+	// Rubric is the owner-authored grading rubric; it may be empty and is
+	// owner-only: it must never reach the student.
+	Rubric string
+	// Answer is the student's text. It is untrusted input.
+	Answer string
+	// Points is the full value of the question.
+	Points int
+}
+
+// Result is a grader's verdict. golem15.quizzes clamps Points to
+// 0..Input.Points in steps of 0.5 and refuses a value that is not a finite
+// number.
+type Result struct {
+	Points   float64
+	Feedback string
+}
+
+// Grader grades open answers. Grade runs in a background job with no
+// transaction open, so it may take seconds; an error is retried a few times
+// and then leaves the answer failed for the owner to grade.
+type Grader interface {
+	// Code identifies the grader in a quiz's grader field, such as "ai".
+	// It must match ^[a-z][a-z0-9_]{0,63}$ and must not be Manual.
+	Code() string
+	// Label is the lang key or text the admin grader dropdown shows.
+	Label() string
+	// Grade scores one answer.
+	Grade(ctx context.Context, in Input) (Result, error)
+}
+
+// HasGraders is implemented by a plugin that provides graders.
+// golem15.quizzes calls it once at boot.
+type HasGraders interface {
+	QuizGraders() []Grader
+}
diff --git a/lang/en/lang.yaml b/lang/en/lang.yaml
index 0dcd08c..584fe1c 100644
--- a/lang/en/lang.yaml
+++ b/lang/en/lang.yaml
@@ -15,6 +15,10 @@ api:
     attempt_limit_reached: All attempts at this test have been used.
     answer_locked: An answer in this test cannot be changed.
     category_not_allowed: A test covers the whole quiz; a category cannot be chosen.
+    text_required: Write an answer to this question.
+    text_too_long: The answer is longer than this question allows.
+graders:
+    manual: Manual (the owner grades)
 permissions:
     access_quizzes: Manage quizzes, questions and answers
     access_attempts: View quiz attempts
diff --git a/lang/pl/lang.yaml b/lang/pl/lang.yaml
index 5cef763..9f09635 100644
--- a/lang/pl/lang.yaml
+++ b/lang/pl/lang.yaml
@@ -15,6 +15,10 @@ api:
     attempt_limit_reached: Wykorzystano już wszystkie podejścia do tego testu.
     answer_locked: W tym teście nie można zmienić udzielonej odpowiedzi.
     category_not_allowed: Test obejmuje cały quiz; nie można wybrać kategorii.
+    text_required: Wpisz odpowiedź na to pytanie.
+    text_too_long: Odpowiedź jest dłuższa, niż pozwala to pytanie.
+graders:
+    manual: Ręcznie (ocenia właściciel)
 permissions:
     access_quizzes: Zarządzanie quizami, pytaniami i odpowiedziami
     access_attempts: Podgląd podejść do quizów
diff --git a/models/attempt_answer.go b/models/attempt_answer.go
index b04f6ed..7a5a6fb 100644
--- a/models/attempt_answer.go
+++ b/models/attempt_answer.go
@@ -8,8 +8,31 @@ import (
 	"git.golem15.com/golem15/summercms/modules/lagoon"
 )
 
+// Grading states of an open answer (AttemptAnswer.GradingStatus).
+const (
+	// GradingPending waits for a grader or for the owner.
+	GradingPending = "pending"
+	// GradingGraded holds a grader's score the owner has not confirmed.
+	GradingGraded = "graded"
+	// GradingFailed is an answer the grader could not score; the owner
+	// grades it.
+	GradingFailed = "failed"
+	// GradingReviewed is final: the owner confirmed or set the score.
+	GradingReviewed = "reviewed"
+)
+
+// GradingStatuses lists the valid values of AttemptAnswer.GradingStatus.
+var GradingStatuses = []string{GradingPending, GradingGraded, GradingFailed, GradingReviewed}
+
 // AttemptAnswer is the user's answer to one question of an attempt. There is
 // one row per attempt and question; answering again overwrites it.
+//
+// TextAnswer and the grading fields belong to open questions only:
+// GradingStatus is NULL for every other row. GradingSeq grows with every new
+// text, so a grading job for an older text is ignored; GradingAttempts counts
+// the grader runs for the current text. The review fields record the owner's
+// decision: ReviewedByUserID for a frontend user (MCP), ReviewedByAdminID for
+// a backend administrator.
 type AttemptAnswer struct {
 	ID             uint                    `gorm:"column:id;primaryKey"`
 	AttemptID      uint                    `gorm:"column:attempt_id"`
@@ -20,6 +43,20 @@ type AttemptAnswer struct {
 	Points         float64                 `gorm:"column:points"`
 	CreatedAt      time.Time               `gorm:"column:created_at"`
 	UpdatedAt      time.Time               `gorm:"column:updated_at"`
+
+	TextAnswer        *string    `gorm:"column:text_answer"`
+	GradingStatus     *string    `gorm:"column:grading_status"`
+	GradingSeq        int        `gorm:"column:grading_seq"`
+	GradingAttempts   int        `gorm:"column:grading_attempts"`
+	Grader            *string    `gorm:"column:grader"`
+	GraderPoints      *float64   `gorm:"column:grader_points"`
+	GraderFeedback    *string    `gorm:"column:grader_feedback"`
+	GradingError      *string    `gorm:"column:grading_error"`
+	GradedAt          *time.Time `gorm:"column:graded_at"`
+	ReviewComment     *string    `gorm:"column:review_comment"`
+	ReviewedAt        *time.Time `gorm:"column:reviewed_at"`
+	ReviewedByUserID  *uint      `gorm:"column:reviewed_by_user_id"`
+	ReviewedByAdminID *uint      `gorm:"column:reviewed_by_admin_id"`
 }
 
 // TableName is the plugin-prefixed table.
diff --git a/models/question.go b/models/question.go
index c8a3272..a93824a 100644
--- a/models/question.go
+++ b/models/question.go
@@ -9,13 +9,25 @@ const (
 	TypeSingle    = "single"
 	TypeMultiple  = "multiple"
 	TypeSelfCheck = "self_check"
+	TypeOpen      = "open"
 )
 
 // QuestionTypes lists the valid values of Question.Type.
-var QuestionTypes = []string{TypeSingle, TypeMultiple, TypeSelfCheck}
+var QuestionTypes = []string{TypeSingle, TypeMultiple, TypeSelfCheck, TypeOpen}
+
+// Limits of the answer length of an open question, in characters.
+const (
+	// DefaultMaxLength applies when MaxLength is nil.
+	DefaultMaxLength = 2000
+	// MaxMaxLength is the largest MaxLength the column accepts.
+	MaxMaxLength = 20000
+)
 
 // Question is one question of a quiz. A self_check question is a flashcard:
-// the user reveals AnswerText and marks whether they knew it.
+// the user reveals AnswerText and marks whether they knew it. An open
+// question is answered in free text and graded against AnswerText (the
+// model answer) and the owner-only Rubric; MaxLength caps the answer
+// (DefaultMaxLength when nil).
 type Question struct {
 	ID          uint      `gorm:"column:id;primaryKey"`
 	QuizID      uint      `gorm:"column:quiz_id"`
@@ -24,6 +36,8 @@ type Question struct {
 	Category    *string   `gorm:"column:category"`
 	AnswerText  *string   `gorm:"column:answer_text"`
 	Explanation *string   `gorm:"column:explanation"`
+	Rubric      *string   `gorm:"column:rubric"`
+	MaxLength   *int      `gorm:"column:max_length"`
 	SortOrder   int       `gorm:"column:sort_order"`
 	Points      int       `gorm:"column:points"`
 	CreatedAt   time.Time `gorm:"column:created_at"`
@@ -37,7 +51,7 @@ func (Question) TableName() string { return "golem15_quizzes_questions" }
 
 // Fillable lists the columns mass assignment may set.
 func (Question) Fillable() []string {
-	return []string{"quiz_id", "text", "type", "category", "answer_text", "explanation", "sort_order", "points"}
+	return []string{"quiz_id", "text", "type", "category", "answer_text", "explanation", "sort_order", "points", "rubric", "max_length"}
 }
 
 // Rules are checked by the admin API on every save.
@@ -45,9 +59,10 @@ func (Question) Rules() map[string]string {
 	return map[string]string{
 		"quiz_id":    "required|integer|min:1",
 		"text":       "required",
-		"type":       "required|in:single,multiple,self_check",
+		"type":       "required|in:single,multiple,self_check,open",
 		"category":   "nullable|max:255",
 		"sort_order": "nullable|integer",
 		"points":     "required|integer|min:0",
+		"max_length": "nullable|integer|between:1,20000",
 	}
 }
diff --git a/models/quiz.go b/models/quiz.go
index ca2a60b..d5a76fc 100644
--- a/models/quiz.go
+++ b/models/quiz.go
@@ -7,6 +7,7 @@ import (
 	"time"
 	"unicode"
 
+	"git.golem15.com/golem15/sm-quizzes-plugin/graders"
 	"gorm.io/gorm"
 )
 
@@ -32,6 +33,8 @@ var QuizNavigations = []string{NavigationFree, NavigationOneWay}
 // free-form link to a record of another plugin, such as "acme.shop:course:12".
 // Mode test withholds solutions until the attempt is finished; Navigation
 // (free or one_way) and MaxAttempts (nil is unlimited) only apply to tests.
+// Grader is the code of the grader of its open answers (graders.Manual: the
+// owner grades them); an unknown code behaves as manual.
 type Quiz struct {
 	ID               uint       `gorm:"column:id;primaryKey"`
 	Title            string     `gorm:"column:title"`
@@ -46,6 +49,7 @@ type Quiz struct {
 	Mode             string     `gorm:"column:mode"`
 	Navigation       string     `gorm:"column:navigation"`
 	MaxAttempts      *int       `gorm:"column:max_attempts"`
+	Grader           string     `gorm:"column:grader"`
 	CreatedAt        time.Time  `gorm:"column:created_at"`
 	UpdatedAt        time.Time  `gorm:"column:updated_at"`
 	Questions        []Question `gorm:"foreignKey:QuizID"`
@@ -59,7 +63,7 @@ func (Quiz) Fillable() []string {
 	return []string{
 		"title", "slug", "description", "context", "is_published",
 		"time_limit_seconds", "passing_score", "shuffle_questions", "round_size",
-		"mode", "navigation", "max_attempts",
+		"mode", "navigation", "max_attempts", "grader",
 	}
 }
 
@@ -76,6 +80,7 @@ func (Quiz) Rules() map[string]string {
 		"mode":               "nullable|in:learning,test",
 		"navigation":         "nullable|in:free,one_way",
 		"max_attempts":       "nullable|integer|min:1",
+		"grader":             "nullable|max:64",
 	}
 }
 
@@ -94,8 +99,8 @@ func (q *Quiz) BeforeSave(*gorm.DB) error {
 	return nil
 }
 
-// fillDefaults stores an empty mode or navigation as learning / free and a
-// non-positive attempt limit as unlimited. GORM writes Go zero values, so
+// fillDefaults stores an empty mode or navigation as learning / free, an
+// empty grader as manual and a non-positive attempt limit as unlimited. GORM writes Go zero values, so
 // without it a models.Quiz{} would hit the columns' CHECK constraints.
 func (q *Quiz) fillDefaults() {
 	if q.Mode == "" {
@@ -104,6 +109,9 @@ func (q *Quiz) fillDefaults() {
 	if q.Navigation == "" {
 		q.Navigation = NavigationFree
 	}
+	if q.Grader == "" {
+		q.Grader = graders.Manual
+	}
 	if q.MaxAttempts != nil && *q.MaxAttempts <= 0 {
 		q.MaxAttempts = nil
 	}
diff --git a/models/quiz_test.go b/models/quiz_test.go
index 8fd6178..2799211 100644
--- a/models/quiz_test.go
+++ b/models/quiz_test.go
@@ -1,6 +1,12 @@
 package models
 
-import "testing"
+import (
+	"slices"
+	"strings"
+	"testing"
+
+	"git.golem15.com/golem15/sm-quizzes-plugin/graders"
+)
 
 func TestQuizDefaults(t *testing.T) {
 	zero, negative := 0, -1
@@ -16,14 +22,52 @@ func TestQuizDefaults(t *testing.T) {
 			if q.Mode != ModeLearning || q.Navigation != NavigationFree || q.MaxAttempts != nil {
 				t.Errorf("%s with max_attempts %d = %q %q %v, want learning free nil", name, *limit, q.Mode, q.Navigation, q.MaxAttempts)
 			}
+			if q.Grader != graders.Manual {
+				t.Errorf("%s left grader %q, want %q", name, q.Grader, graders.Manual)
+			}
 		}
 		two := 2
-		q := &Quiz{Title: "Q", Mode: ModeTest, Navigation: NavigationOneWay, MaxAttempts: &two}
+		q := &Quiz{Title: "Q", Mode: ModeTest, Navigation: NavigationOneWay, MaxAttempts: &two, Grader: "ai"}
 		if err := hook(q); err != nil {
 			t.Fatalf("%s: %v", name, err)
 		}
 		if q.Mode != ModeTest || q.Navigation != NavigationOneWay || q.MaxAttempts == nil || *q.MaxAttempts != 2 {
 			t.Errorf("%s changed a test quiz: %q %q %v", name, q.Mode, q.Navigation, q.MaxAttempts)
 		}
+		if q.Grader != "ai" {
+			t.Errorf("%s changed an explicit grader to %q", name, q.Grader)
+		}
+	}
+}
+
+func TestOpenQuestionModel(t *testing.T) {
+	if !slices.Contains(QuestionTypes, TypeOpen) || len(QuestionTypes) != 4 {
+		t.Errorf("QuestionTypes = %v, want the open type appended", QuestionTypes)
+	}
+	if QuestionTypes[0] != TypeSingle || QuestionTypes[1] != TypeMultiple || QuestionTypes[2] != TypeSelfCheck {
+		t.Errorf("QuestionTypes reordered the existing types: %v", QuestionTypes)
+	}
+	rules := Question{}.Rules()
+	if !strings.Contains(rules["type"], "open") || !strings.Contains(rules["type"], "self_check") {
+		t.Errorf("type rule = %q, want every type including open", rules["type"])
+	}
+	if rules["max_length"] != "nullable|integer|between:1,20000" {
+		t.Errorf("max_length rule = %q", rules["max_length"])
+	}
+	fill := Question{}.Fillable()
+	if !slices.Contains(fill, "rubric") || !slices.Contains(fill, "max_length") {
+		t.Errorf("Fillable = %v, want rubric and max_length", fill)
+	}
+	if DefaultMaxLength != 2000 || MaxMaxLength != 20000 {
+		t.Errorf("max length limits = %d / %d", DefaultMaxLength, MaxMaxLength)
+	}
+	if len(GradingStatuses) != 4 || GradingStatuses[0] != GradingPending {
+		t.Errorf("GradingStatuses = %v", GradingStatuses)
+	}
+	if !slices.Contains(Quiz{}.Fillable(), "grader") {
+		t.Error("Quiz.Fillable lacks grader")
+	}
+	if len(AttemptAnswer{}.Fillable()) != 0 {
+		t.Error("AttemptAnswer.Fillable must stay empty")
 	}
 }
diff --git a/plugin_test.go b/plugin_test.go
index 7d2a50b..4a3b9dd 100644
--- a/plugin_test.go
+++ b/plugin_test.go
@@ -43,8 +43,8 @@ func TestPluginActivates(t *testing.T) {
 	if len(migrations) == 0 {
 		t.Fatal("no migrations")
 	}
-	if got := migrations[len(migrations)-1].ID; got != "20261109120000_add_test_mode_to_quizzes" {
-		t.Errorf("last migration = %s, want 20261109120000_add_test_mode_to_quizzes", got)
+	if got := migrations[len(migrations)-1].ID; got != "20261109180000_add_open_questions_and_grading" {
+		t.Errorf("last migration = %s, want 20261109180000_add_open_questions_and_grading", got)
 	}
 	if app.Config.Bool("golem15.quizzes.scoring.multiple_partial") {
 		t.Error("scoring.multiple_partial defaults to true, want false")
diff --git a/registry.gen.go b/registry.gen.go
index 6117492..6416719 100644
--- a/registry.gen.go
+++ b/registry.gen.go
@@ -27,6 +27,7 @@ func generatedMigrations() []*gormigrate.Migration {
 		updates.AddCategoryToAttempts(),
 		updates.AddCompletedAttemptsIndex(),
 		updates.AddTestModeToQuizzes(),
+		updates.AddOpenQuestionsAndGrading(),
 	}
 }
 
diff --git a/updates/20261109180000_add_open_questions_and_grading.go b/updates/20261109180000_add_open_questions_and_grading.go
new file mode 100644
index 0000000..6418b5d
--- /dev/null
+++ b/updates/20261109180000_add_open_questions_and_grading.go
@@ -0,0 +1,108 @@
+// Code generated by summer make. DO NOT EDIT.
+
+package updates
+
+import (
+	"fmt"
+
+	"github.com/go-gormigrate/gormigrate/v2"
+	"gorm.io/gorm"
+)
+
+// AddOpenQuestionsAndGrading returns the 20261109180000_add_open_questions_and_grading gormigrate entry.
+// It adds the open question type (answered in free text, graded against the
+// model answer and an owner-only rubric), the quiz grader and the grading
+// and review state of each attempt answer. Every existing question, quiz and
+// answer keeps today's behaviour: the grader is manual, no question is open
+// and every existing answer row has no grading state.
+func AddOpenQuestionsAndGrading() *gormigrate.Migration {
+	return &gormigrate.Migration{
+		ID: "20261109180000_add_open_questions_and_grading",
+		Migrate: func(tx *gorm.DB) error {
+			// The original type CHECK is an unnamed inline constraint, so
+			// it is found by the column it checks rather than by an assumed
+			// name. Its definition cannot be matched on "IN": Postgres
+			// stores it as type = ANY (ARRAY[...]).
+			for _, stmt := range []string{
+				`DO $$
+DECLARE c RECORD;
+BEGIN
+    FOR c IN
+        SELECT con.conname FROM pg_constraint con
+        JOIN pg_attribute att ON att.attrelid = con.conrelid AND att.attname = 'type'
+        WHERE con.conrelid = 'golem15_quizzes_questions'::regclass
+          AND con.contype = 'c'
+          AND con.conkey = ARRAY[att.attnum]
+    LOOP
+        EXECUTE format('ALTER TABLE golem15_quizzes_questions DROP CONSTRAINT %I', c.conname);
+    END LOOP;
+END $$`,
+				`ALTER TABLE golem15_quizzes_questions
+    ADD CONSTRAINT golem15_quizzes_questions_type_check CHECK (type IN ('single', 'multiple', 'self_check', 'open')),
+    ADD COLUMN rubric TEXT NULL,
+    ADD COLUMN max_length INTEGER NULL CHECK (max_length IS NULL OR max_length BETWEEN 1 AND 20000)`,
+				`ALTER TABLE golem15_quizzes_quizzes
+    ADD COLUMN grader VARCHAR(64) NOT NULL DEFAULT 'manual'`,
+				`ALTER TABLE golem15_quizzes_attempt_answers
+    ADD COLUMN text_answer TEXT NULL,
+    ADD COLUMN grading_status VARCHAR(20) NULL CHECK (grading_status IS NULL OR grading_status IN ('pending', 'graded', 'failed', 'reviewed')),
+    ADD COLUMN grading_seq INTEGER NOT NULL DEFAULT 0,
+    ADD COLUMN grading_attempts INTEGER NOT NULL DEFAULT 0,
+    ADD COLUMN grader VARCHAR(64) NULL,
+    ADD COLUMN grader_points DOUBLE PRECISION NULL,
+    ADD COLUMN grader_feedback TEXT NULL,
+    ADD COLUMN grading_error TEXT NULL,
+    ADD COLUMN graded_at TIMESTAMPTZ NULL,
+    ADD COLUMN review_comment TEXT NULL,
+    ADD COLUMN reviewed_at TIMESTAMPTZ NULL,
+    ADD COLUMN reviewed_by_user_id INTEGER NULL,
+    ADD COLUMN reviewed_by_admin_id INTEGER NULL`,
+				`CREATE INDEX golem15_quizzes_attempt_answers_grading_index ON golem15_quizzes_attempt_answers (grading_status)
+    WHERE grading_status IN ('pending', 'graded', 'failed')`,
+			} {
+				if err := tx.Exec(stmt).Error; err != nil {
+					return err
+				}
+			}
+			return nil
+		},
+		Rollback: func(tx *gorm.DB) error {
+			var n int64
+			if err := tx.Raw(`SELECT COUNT(*) FROM golem15_quizzes_questions WHERE type = 'open'`).Scan(&n).Error; err != nil {
+				return err
+			}
+			if n > 0 {
+				return fmt.Errorf("quizzes: rollback refused: %d open questions exist; delete them first", n)
+			}
+			for _, stmt := range []string{
+				`DROP INDEX IF EXISTS golem15_quizzes_attempt_answers_grading_index`,
+				`ALTER TABLE golem15_quizzes_attempt_answers
+    DROP COLUMN IF EXISTS reviewed_by_admin_id,
+    DROP COLUMN IF EXISTS reviewed_by_user_id,
+    DROP COLUMN IF EXISTS reviewed_at,
+    DROP COLUMN IF EXISTS review_comment,
+    DROP COLUMN IF EXISTS graded_at,
+    DROP COLUMN IF EXISTS grading_error,
+    DROP COLUMN IF EXISTS grader_feedback,
+    DROP COLUMN IF EXISTS grader_points,
+    DROP COLUMN IF EXISTS grader,
+    DROP COLUMN IF EXISTS grading_attempts,
+    DROP COLUMN IF EXISTS grading_seq,
+    DROP COLUMN IF EXISTS grading_status,
+    DROP COLUMN IF EXISTS text_answer`,
+				`ALTER TABLE golem15_quizzes_quizzes DROP COLUMN IF EXISTS grader`,
+				`ALTER TABLE golem15_quizzes_questions
+    DROP COLUMN IF EXISTS max_length,
+    DROP COLUMN IF EXISTS rubric,
+    DROP CONSTRAINT IF EXISTS golem15_quizzes_questions_type_check`,
+				`ALTER TABLE golem15_quizzes_questions
+    ADD CONSTRAINT golem15_quizzes_questions_type_check CHECK (type IN ('single', 'multiple', 'self_check'))`,
+			} {
+				if err := tx.Exec(stmt).Error; err != nil {
+					return err
+				}
+			}
+			return nil
+		},
+	}
+}