diff --git a/calibre_toolkit/commands/audit.py b/calibre_toolkit/commands/audit.py
index 34abf7c..5afd6cf 100644
--- a/calibre_toolkit/commands/audit.py
+++ b/calibre_toolkit/commands/audit.py
@@ -86,9 +86,10 @@ def load_audit_records(path: Path) -> list[AuditRecord]:
 
     Builds on the shared `read_audit_entries` reader (one parser for the
     whole toolkit), then filters to records that represent AI-applied writes
-    — those have a non-empty `confidence`, `source`, and `step`. Manual marker
-    writes (mark_mqg_complete and similar) don't carry these and are correctly
-    excluded from the calibration pool. Re-grade writes (carrying a `regrade`
+    — those have a non-empty `confidence`, `source`, and `step`. Flag writes
+    (mark_mqg_complete / clear_mqg_flag, since v1.10) carry source="flag" and
+    step="flag:<command>" but no confidence, so they are correctly excluded
+    from the calibration pool. Re-grade writes (carrying a `regrade`
     marker) are also excluded — they re-run a changed rule and would otherwise
     mix pre- and post-change accuracy in the same tier stats.
     """
diff --git a/calibre_toolkit/commands/clean_titles.py b/calibre_toolkit/commands/clean_titles.py
index b8698c8..d137d85 100644
--- a/calibre_toolkit/commands/clean_titles.py
+++ b/calibre_toolkit/commands/clean_titles.py
@@ -397,7 +397,7 @@ def _mark_complete(
     if not mqg_column or not book_ids:
         return
     with console.status(f"[cyan]Marking {len(book_ids)} {label} books as MQG-01 complete…"):
-        db.mark_mqg_complete(book_ids, mqg_column)
+        db.mark_mqg_complete(book_ids, mqg_column, audit_step="clean-titles")
     console.print(
         f"[dim]Marked {len(book_ids)} books as complete in [bold]{mqg_column}[/bold].[/dim]"
     )
diff --git a/calibre_toolkit/commands/comments_enrich.py b/calibre_toolkit/commands/comments_enrich.py
index 211e440..3e6b5b4 100644
--- a/calibre_toolkit/commands/comments_enrich.py
+++ b/calibre_toolkit/commands/comments_enrich.py
@@ -376,7 +376,7 @@ def run_comments_enrichment(
             f"[bold]{mqg_manual_column}[/bold] for manual review."
         )
         with console.status("Flagging…"):
-            db.mark_mqg_complete(manual_ids, mqg_manual_column)
+            db.mark_mqg_complete(manual_ids, mqg_manual_column, audit_step="comments-enrich")
 
     applied_ids_set = set(applied_ids)
     applied_suggestions = [s for s in suggestions if s.book_id in applied_ids_set]
@@ -453,7 +453,7 @@ def _mark_complete(
     if not mqg_column or not book_ids:
         return
     with console.status(f"[cyan]Marking {len(book_ids)} books as {label} complete…"):
-        db.mark_mqg_complete(book_ids, mqg_column)
+        db.mark_mqg_complete(book_ids, mqg_column, audit_step="comments-enrich")
     console.print(
         f"[dim]Marked {len(book_ids)} books complete in [bold]{mqg_column}[/bold].[/dim]"
     )
diff --git a/calibre_toolkit/commands/enrich_identifiers.py b/calibre_toolkit/commands/enrich_identifiers.py
index f00b7a2..227f3fe 100644
--- a/calibre_toolkit/commands/enrich_identifiers.py
+++ b/calibre_toolkit/commands/enrich_identifiers.py
@@ -553,7 +553,7 @@ def _mark_complete(
     if not mqg_column or not book_ids:
         return
     with console.status(f"[cyan]Marking {len(book_ids)} {label} books as MQG-02 complete…"):
-        db.mark_mqg_complete(book_ids, mqg_column)
+        db.mark_mqg_complete(book_ids, mqg_column, audit_step="enrich-identifiers")
     console.print(
         f"[dim]Marked {len(book_ids)} books as complete in [bold]{mqg_column}[/bold].[/dim]"
     )
@@ -567,4 +567,4 @@ def _mark_manual(
     if not book_ids:
         return
     with console.status(f"[cyan]Flagging {len(book_ids)} books for manual curation…"):
-        db.mark_mqg_complete(book_ids, mqg_manual_column)
+        db.mark_mqg_complete(book_ids, mqg_manual_column, audit_step="enrich-identifiers")
diff --git a/calibre_toolkit/commands/lcc_enrich.py b/calibre_toolkit/commands/lcc_enrich.py
index 3952a09..fc517ff 100644
--- a/calibre_toolkit/commands/lcc_enrich.py
+++ b/calibre_toolkit/commands/lcc_enrich.py
@@ -600,7 +600,7 @@ def run_lcc_enrichment(
             f"[bold]{mqg_manual_column}[/bold] for manual review."
         )
         with console.status("Flagging…"):
-            db.mark_mqg_complete(manual_ids, mqg_manual_column)
+            db.mark_mqg_complete(manual_ids, mqg_manual_column, audit_step="lcc-enrich")
 
     applied_ids_set = set(applied_ids)
     flagged = len(manual_ids) if mqg_manual_column else 0
@@ -716,5 +716,5 @@ def _mark_complete(db: CalibreDB, mqg_column: str, book_ids: list[int], label: s
     if not mqg_column or not book_ids:
         return
     with console.status(f"[cyan]Marking {len(book_ids)} books as {label} complete…"):
-        db.mark_mqg_complete(book_ids, mqg_column)
+        db.mark_mqg_complete(book_ids, mqg_column, audit_step="lcc-enrich")
     console.print(f"[dim]Marked {len(book_ids)} books complete in [bold]{mqg_column}[/bold].[/dim]")
diff --git a/calibre_toolkit/commands/tags_enrich.py b/calibre_toolkit/commands/tags_enrich.py
index 4994430..25c9497 100644
--- a/calibre_toolkit/commands/tags_enrich.py
+++ b/calibre_toolkit/commands/tags_enrich.py
@@ -310,7 +310,7 @@ def run_tags_enrichment(
             f"[bold]{mqg_manual_column}[/bold] for manual review."
         )
         with console.status("Flagging…"):
-            db.mark_mqg_complete(manual_ids, mqg_manual_column)
+            db.mark_mqg_complete(manual_ids, mqg_manual_column, audit_step="tags-enrich")
 
     applied_ids_set = set(applied_ids)
     flagged = len(manual_ids) if mqg_manual_column else 0
@@ -397,7 +397,7 @@ def _mark_complete(
     if not mqg_column or not book_ids:
         return
     with console.status(f"[cyan]Marking {len(book_ids)} books as {label} complete…"):
-        db.mark_mqg_complete(book_ids, mqg_column)
+        db.mark_mqg_complete(book_ids, mqg_column, audit_step="tags-enrich")
     console.print(
         f"[dim]Marked {len(book_ids)} books complete in [bold]{mqg_column}[/bold].[/dim]"
     )
diff --git a/calibre_toolkit/commands/tags_review.py b/calibre_toolkit/commands/tags_review.py
index e8b1661..7b8d7c6 100644
--- a/calibre_toolkit/commands/tags_review.py
+++ b/calibre_toolkit/commands/tags_review.py
@@ -185,9 +185,9 @@ def _prompt_action(has_ai: bool) -> str:
 
 def _lock(db: CalibreDB, book_id: int, reviewed_column: str, mqg_column: str | None) -> None:
     """Mark a book as reviewed and (if configured) as MQG-05 complete."""
-    db.mark_mqg_complete([book_id], reviewed_column)
+    db.mark_mqg_complete([book_id], reviewed_column, audit_step="tags-review")
     if mqg_column:
-        db.mark_mqg_complete([book_id], mqg_column)
+        db.mark_mqg_complete([book_id], mqg_column, audit_step="tags-review")
 
 
 def _inline_edit(current: list[str], suggested: list[str] | None) -> list[str]:
diff --git a/calibre_toolkit/commands/unflag_manual.py b/calibre_toolkit/commands/unflag_manual.py
index be977b8..2871357 100644
--- a/calibre_toolkit/commands/unflag_manual.py
+++ b/calibre_toolkit/commands/unflag_manual.py
@@ -123,7 +123,7 @@ def run_unflag_manual(
     with console.status(f"Clearing flags for {len(books)} book(s)…"):
         for book in books:
             try:
-                db.clear_mqg_flag(book.id, mqg_manual_column)
+                db.clear_mqg_flag(book.id, mqg_manual_column, audit_step="unflag-manual")
                 cleared += 1
             except RuntimeError as e:
                 console.print(f"[red]Error on book {book.id}: {e}[/red]")
diff --git a/calibre_toolkit/db.py b/calibre_toolkit/db.py
index 40f6bb0..302e539 100644
--- a/calibre_toolkit/db.py
+++ b/calibre_toolkit/db.py
@@ -10,7 +10,7 @@
 from pathlib import Path
 from dataclasses import dataclass, field
 
-from .logging_config import get_logger
+from .logging_config import audit_log, get_logger
 
 _log = get_logger(__name__)
 
@@ -291,12 +291,21 @@ def _one(item):
                     progress_callback(done, len(updates), len(failures))
         return applied, failures
 
-    def mark_mqg_complete(self, book_ids: list[int], column: str) -> None:
+    def mark_mqg_complete(
+        self, book_ids: list[int], column: str, audit_step: str = "",
+    ) -> None:
         """Mark a list of books as complete for a given MQG column.
 
         Writes directly to SQLite in a single transaction — avoids spawning
         one calibredb process per book, which is prohibitively slow for
         large batches.
+
+        `audit_step` names the command performing the flag write; each book
+        gets one audit entry with source="flag" and step="flag:<audit_step>".
+        The "flag:" namespace is load-bearing: the regrade staleness selector
+        filters on exact AI step names ("lcc-enrich" etc.), so a bare command
+        name here would make a flagged book look freshly enriched. Entries
+        carry no confidence, which keeps them out of the calibration pool.
         """
         label = column.lstrip("#")
         with self._connect() as ro:
@@ -317,7 +326,13 @@ def mark_mqg_complete(self, book_ids: list[int], column: str) -> None:
             )
             conn.commit()
 
-    def clear_mqg_flag(self, book_id: int, column: str) -> None:
+        step = f"flag:{audit_step}" if audit_step else "flag"
+        for bid in book_ids:
+            audit_log(bid, column, True, source="flag", step=step)
+
+    def clear_mqg_flag(
+        self, book_id: int, column: str, audit_step: str = "",
+    ) -> None:
         """Clear a custom boolean MQG column for a single book.
 
         Deletes the row rather than writing value=0: Calibre's bool-column
@@ -326,6 +341,10 @@ def clear_mqg_flag(self, book_id: int, column: str) -> None:
         `not #<manual>:true` exclusion and the book would never be
         re-queued. Removing the row restores the undefined state that
         `not #col:true` matches.
+
+        Audited the same way as mark_mqg_complete (source="flag",
+        step="flag:<audit_step>"), with new_value=None recording the
+        cleared state.
         """
         label = column.lstrip("#")
         with self._connect() as ro:
@@ -340,6 +359,9 @@ def clear_mqg_flag(self, book_id: int, column: str) -> None:
             conn.execute(f"DELETE FROM {table} WHERE book = ?", (book_id,))
             conn.commit()
 
+        step = f"flag:{audit_step}" if audit_step else "flag"
+        audit_log(book_id, column, None, source="flag", step=step)
+
     def get_identifiers(self, book_id: int) -> dict[str, str]:
         """Return {type: value} for all identifiers currently on a book."""
         query = "SELECT type, val FROM identifiers WHERE book = ?"
diff --git a/tests/conftest.py b/tests/conftest.py
index 1219456..c99cf00 100644
--- a/tests/conftest.py
+++ b/tests/conftest.py
@@ -16,6 +16,18 @@
 FIXTURES_DIR = Path(__file__).parent / "fixtures"
 
 
+@pytest.fixture(autouse=True)
+def _isolated_audit_log(tmp_path: Path, monkeypatch):
+    """Keep every test's audit trail out of ~/.calibre-toolkit/audit.log.
+
+    Flag writes audit through db.mark_mqg_complete / clear_mqg_flag since
+    v1.10, so any test touching those methods would otherwise append to the
+    user's real audit log. Tests that need a specific path still win — their
+    own monkeypatch.setenv overwrites this one.
+    """
+    monkeypatch.setenv("CALIBRE_TOOLKIT_AUDIT_LOG", str(tmp_path / "audit.log"))
+
+
 @pytest.fixture
 def fixtures_dir() -> Path:
     """Absolute path to the per-suite fixtures directory."""
diff --git a/tests/test_flag_audit.py b/tests/test_flag_audit.py
new file mode 100644
index 0000000..20614ed
--- /dev/null
+++ b/tests/test_flag_audit.py
@@ -0,0 +1,170 @@
+"""Contract tests for flag-write audit logging (v1.10 item 3).
+
+mark_mqg_complete and clear_mqg_flag write one audit entry per book with
+source="flag" and step="flag:<command>". Three reader contracts are
+load-bearing:
+
+  1. The calibration loader (load_audit_records) excludes flag entries —
+     they carry no confidence.
+  2. The regrade staleness selector (find_stale_books) filters on exact AI
+     step names, so the "flag:" namespace keeps a flag write from making a
+     book look freshly enriched.
+  3. A flag write inside a regrade context picks up the regrade marker at
+     the audit_log choke point — harmless, since contracts 1 and 2 exclude
+     the entry on other fields.
+"""
+
+from __future__ import annotations
+
+import json
+import sqlite3
+from datetime import datetime, timezone
+from pathlib import Path
+
+import pytest
+
+from calibre_toolkit.db import CalibreDB
+from calibre_toolkit.logging_config import audit_log, regrade_audit
+from calibre_toolkit.commands.audit import load_audit_records
+from calibre_toolkit.commands.regrade import find_stale_books
+
+
+def _build_minimal_calibre_db(library_path: Path) -> None:
+    library_path.mkdir(parents=True, exist_ok=True)
+    conn = sqlite3.connect(str(library_path / "metadata.db"))
+    conn.executescript(
+        """
+        CREATE TABLE books (
+            id INTEGER PRIMARY KEY,
+            title TEXT,
+            sort TEXT,
+            author_sort TEXT,
+            pubdate TEXT
+        );
+        CREATE TABLE custom_columns (
+            id INTEGER PRIMARY KEY,
+            label TEXT,
+            datatype TEXT
+        );
+        CREATE TABLE custom_column_7 (
+            book INTEGER,
+            value INTEGER,
+            UNIQUE(book)
+        );
+        """
+    )
+    conn.executemany("INSERT INTO books (id, title) VALUES (?, ?)", [
+        (1, "Book One"),
+        (2, "Book Two"),
+    ])
+    conn.execute(
+        "INSERT INTO custom_columns (id, label, datatype) VALUES (7, 'mqg_lcc_manual', 'bool')"
+    )
+    conn.commit()
+    conn.close()
+
+
+@pytest.fixture
+def db(tmp_path: Path) -> CalibreDB:
+    library = tmp_path / "lib"
+    _build_minimal_calibre_db(library)
+    return CalibreDB(library_path=str(library), calibredb_path="calibredb")
+
+
+@pytest.fixture
+def audit_path(tmp_path: Path) -> Path:
+    # The conftest autouse fixture points CALIBRE_TOOLKIT_AUDIT_LOG here.
+    return tmp_path / "audit.log"
+
+
+def _entries(path: Path) -> list[dict]:
+    if not path.exists():
+        return []
+    return [json.loads(l) for l in path.read_text(encoding="utf-8").splitlines() if l.strip()]
+
+
+# ── Writer contract ──────────────────────────────────────────────────────────
+
+
+def test_mark_writes_one_entry_per_book(db: CalibreDB, audit_path: Path):
+    db.mark_mqg_complete([1, 2], "#mqg_lcc_manual", audit_step="lcc-enrich")
+    entries = _entries(audit_path)
+    assert len(entries) == 2
+    assert {e["book_id"] for e in entries} == {1, 2}
+    for e in entries:
+        assert e["field"] == "#mqg_lcc_manual"
+        assert e["new_value"] is True
+        assert e["source"] == "flag"
+        assert e["step"] == "flag:lcc-enrich"
+        assert "confidence" not in e
+
+
+def test_clear_writes_entry_with_null_value(db: CalibreDB, audit_path: Path):
+    db.mark_mqg_complete([1], "#mqg_lcc_manual", audit_step="lcc-enrich")
+    db.clear_mqg_flag(1, "#mqg_lcc_manual", audit_step="unflag-manual")
+    entries = _entries(audit_path)
+    assert len(entries) == 2
+    clear = entries[-1]
+    assert clear["book_id"] == 1
+    assert clear["new_value"] is None
+    assert clear["source"] == "flag"
+    assert clear["step"] == "flag:unflag-manual"
+
+
+def test_default_step_is_bare_flag(db: CalibreDB, audit_path: Path):
+    db.mark_mqg_complete([1], "#mqg_lcc_manual")
+    assert _entries(audit_path)[0]["step"] == "flag"
+
+
+def test_mark_unknown_column_noop_writes_no_entry(db: CalibreDB, audit_path: Path):
+    db.mark_mqg_complete([1], "#no_such_column", audit_step="lcc-enrich")
+    assert _entries(audit_path) == []
+
+
+def test_clear_unknown_column_raises_and_writes_no_entry(db: CalibreDB, audit_path: Path):
+    with pytest.raises(RuntimeError, match="not found"):
+        db.clear_mqg_flag(1, "#no_such_column", audit_step="unflag-manual")
+    assert _entries(audit_path) == []
+
+
+# ── Reader contract 1: calibration excludes flag entries ─────────────────────
+
+
+def test_calibration_loader_excludes_flag_entries(db: CalibreDB, audit_path: Path):
+    audit_log(1, "#lcc", "PR6059", confidence="high", source="ai", step="lcc-enrich")
+    db.mark_mqg_complete([1], "#mqg_lcc_manual", audit_step="lcc-enrich")
+
+    records = load_audit_records(audit_path)
+    assert len(records) == 1
+    assert records[0].field == "#lcc"
+    assert records[0].source == "ai"
+
+
+# ── Reader contract 2: regrade staleness ignores flag entries ────────────────
+
+
+def test_flag_entry_does_not_refresh_regrade_staleness():
+    entries = [
+        {"timestamp": "2026-05-01T00:00:00+00:00", "book_id": 1,
+         "field": "#lcc", "new_value": "PR6059",
+         "confidence": "high", "source": "ai", "step": "lcc-enrich"},
+        # Later flag write for the same book — must NOT make it look fresh.
+        {"timestamp": "2026-06-09T00:00:00+00:00", "book_id": 1,
+         "field": "#mqg_lcc", "new_value": True,
+         "source": "flag", "step": "flag:lcc-enrich"},
+    ]
+    before = datetime(2026, 6, 1, tzinfo=timezone.utc)
+    assert find_stale_books(entries, "lcc-enrich", before) == [1]
+
+
+# ── Reader contract 3: regrade marker rides along harmlessly ─────────────────
+
+
+def test_flag_inside_regrade_context_carries_marker_and_stays_excluded(
+    db: CalibreDB, audit_path: Path,
+):
+    with regrade_audit("2026-06-01"):
+        db.mark_mqg_complete([1], "#mqg_lcc_manual", audit_step="lcc-enrich")
+    entry = _entries(audit_path)[0]
+    assert entry["regrade"] == "2026-06-01"
+    assert load_audit_records(audit_path) == []