diff --git a/internal/handler/documents.go b/internal/handler/documents.go
index 07d48c8..fa2c661 100644
--- a/internal/handler/documents.go
+++ b/internal/handler/documents.go
@@ -7,9 +7,11 @@ import (
 	"io"
 	"log/slog"
 	"net/http"
+	"regexp"
 	"strconv"
 	"strings"
 	"time"
+	"unicode/utf8"
 
 	"github.com/go-chi/chi/v5"
 
@@ -128,6 +130,34 @@ func (h *DocumentsHandler) HandleListDocuments(w http.ResponseWriter, r *http.Re
 	writeJSON(w, http.StatusOK, resp)
 }
 
+// sanitizeTitle coerces a document title into a value safe for the
+// Postgres text column: valid UTF-8, no C0 control chars (except none —
+// titles are single-line), whitespace-collapsed and trimmed. Invalid
+// byte sequences (a client that mangled a non-ASCII char in the multipart
+// field) are dropped rather than replaced so the title stays clean. An
+// all-garbage title collapses to "" and the caller falls back to the doc id.
+func sanitizeTitle(s string) string {
+	if !utf8.ValidString(s) {
+		s = strings.ToValidUTF8(s, "")
+	}
+	var b strings.Builder
+	b.Grow(len(s))
+	for _, r := range s {
+		if r == utf8.RuneError {
+			continue
+		}
+		if r < 0x20 { // drop control chars incl. tab/newline — titles are one line
+			b.WriteRune(' ')
+			continue
+		}
+		b.WriteRune(r)
+	}
+	return strings.TrimSpace(multiSpaceRe.ReplaceAllString(b.String(), " "))
+}
+
+// multiSpaceRe collapses runs of whitespace in a sanitized title.
+var multiSpaceRe = regexp.MustCompile(`\s{2,}`)
+
 // sourceTypeFromContentType collapses an HTTP Content-Type to the
 // short tag the dashboard's source-type badge expects.
 func sourceTypeFromContentType(ct string) string {
@@ -249,6 +279,12 @@ func (h *DocumentsHandler) HandleIngestDocument(w http.ResponseWriter, r *http.R
 	if title == "" {
 		title = filename
 	}
+	// Sanitize before the DB write: a title with invalid UTF-8 bytes
+	// (e.g. a client that mangled an em-dash) would otherwise make the
+	// Postgres insert fail with "db write failed" and 500 the whole
+	// ingest. Coerce to valid UTF-8 and drop control chars so a bad
+	// title byte can never fail the upload.
+	title = sanitizeTitle(title)
 	if title == "" {
 		title = string(docID)
 	}
diff --git a/internal/handler/sanitize_title_test.go b/internal/handler/sanitize_title_test.go
new file mode 100644
index 0000000..aa66ac2
--- /dev/null
+++ b/internal/handler/sanitize_title_test.go
@@ -0,0 +1,20 @@
+package handler
+
+import "testing"
+
+func TestSanitizeTitle(t *testing.T) {
+	cases := []struct{ in, want string }{
+		{"Attention Is All You Need", "Attention Is All You Need"},
+		{"  spaced   out  title ", "spaced out title"},
+		{"Berkshire — 2023", "Berkshire — 2023"}, // valid UTF-8 em-dash is kept
+		{"bad\xff\xfebyte", "badbyte"},           // invalid UTF-8 bytes dropped
+		{"line\nbreak\ttab", "line break tab"},   // control chars → space, collapsed
+		{"\xff\xfe", ""},                         // all-garbage collapses to empty
+		{"", ""},
+	}
+	for _, c := range cases {
+		if got := sanitizeTitle(c.in); got != c.want {
+			t.Errorf("sanitizeTitle(%q) = %q, want %q", c.in, got, c.want)
+		}
+	}
+}
diff --git a/pkg/ingest/ingest.go b/pkg/ingest/ingest.go
index f4ab4f1..beb92e7 100644
--- a/pkg/ingest/ingest.go
+++ b/pkg/ingest/ingest.go
@@ -714,7 +714,9 @@ func (p *Pipeline) persistTree(ctx context.Context, store docPersister, docID tr
 		// Watermarked PDFs whose overlay text shares a Y coordinate with
 		// the real title produce mojibake like "GGlloobbaall SSttrraatteeggyy"
 		// — we'd rather keep the original filename than show that to a user.
-		if err := store.SetDocumentTitle(ctx, docID, doc.Title); err != nil {
+		// cleanForLLM strips any invalid-UTF-8 bytes so the title write can
+		// never fail the persist.
+		if err := store.SetDocumentTitle(ctx, docID, cleanForLLM(doc.Title)); err != nil {
 			return err
 		}
 	}
diff --git a/pkg/parser/pdf.go b/pkg/parser/pdf.go
index 9cc961f..19962b9 100644
--- a/pkg/parser/pdf.go
+++ b/pkg/parser/pdf.go
@@ -1477,12 +1477,42 @@ var boilerplateFragments = []string{
 	"or scholarly",
 }
 
+// isRepeatedMarkerLine reports whether a row is a run of the same short
+// token repeated — the artifact of an icon glyph rendered as text and
+// stamped once per item, e.g. the ORCID "iD" logo repeated once per author
+// on a paper's byline ("iD iD iD iD"). Real prose never repeats an
+// identical ≤2-character token as the majority of a ≥3-token line, so
+// dropping these is safe and removes the junk author-block "headings"
+// (e.g. PRISMA's contributor page) the font heuristic would otherwise
+// surface as sections.
+func isRepeatedMarkerLine(s string) bool {
+	fields := strings.Fields(s)
+	if len(fields) < 3 {
+		return false
+	}
+	counts := map[string]int{}
+	for _, f := range fields {
+		if len([]rune(f)) <= 2 {
+			counts[f]++
+		}
+	}
+	for _, c := range counts {
+		if c*100 >= len(fields)*60 { // ≥60% of the line is one short token
+			return true
+		}
+	}
+	return false
+}
+
 // isBoilerplateLine reports whether a row is publisher/license noise.
 // Matches the curated signature list, the bare arXiv id stamp
-// ("arXiv:2401.01234v2 [cs.CL] 5 Jan 2024"), and short license-tail
-// fragments.
+// ("arXiv:2401.01234v2 [cs.CL] 5 Jan 2024"), short license-tail
+// fragments, and repeated icon-glyph marker rows (ORCID "iD iD iD").
 func isBoilerplateLine(s string) bool {
 	low := strings.ToLower(strings.TrimSpace(s))
+	if isRepeatedMarkerLine(low) {
+		return true
+	}
 	for _, sig := range boilerplateSignatures {
 		if strings.Contains(low, sig) {
 			return true
diff --git a/pkg/parser/repeated_marker_test.go b/pkg/parser/repeated_marker_test.go
new file mode 100644
index 0000000..e3db729
--- /dev/null
+++ b/pkg/parser/repeated_marker_test.go
@@ -0,0 +1,32 @@
+package parser
+
+import "testing"
+
+// TestIsRepeatedMarkerLine locks the filter that drops icon-glyph marker
+// rows (e.g. an ORCID "iD" repeated once per author) while never touching
+// real prose or short headings.
+func TestIsRepeatedMarkerLine(t *testing.T) {
+	drop := []string{
+		"id id id id",        // 4/4 = 100%
+		"iD iD iD iD",        // case-insensitive caller lower-cases first; test raw too
+		"id id id",           // 3/3
+		"id id id id author", // 4/5 = 80% ≥ 60%
+	}
+	keep := []string{
+		"the right to erasure is established in article 17",
+		"multi-head attention",                     // 2 tokens, below the 3-token floor
+		"introduction",                             // single token
+		"id id id penny whiting17, david moher 22", // 3 of 9 ≈ 33% → real author line, kept
+		"id id elie a. akl 8, sue e. brennan",      // 2 of 8 → kept
+	}
+	for _, s := range drop {
+		if !isRepeatedMarkerLine(s) {
+			t.Errorf("expected DROP for %q", s)
+		}
+	}
+	for _, s := range keep {
+		if isRepeatedMarkerLine(s) {
+			t.Errorf("expected KEEP for %q", s)
+		}
+	}
+}