diff --git a/scripts/notebook_tools/check_accent_restoration_invariants.py b/scripts/notebook_tools/check_accent_restoration_invariants.py
new file mode 100644
index 0000000000..bf910e2d3a
--- /dev/null
+++ b/scripts/notebook_tools/check_accent_restoration_invariants.py
@@ -0,0 +1,190 @@
+#!/usr/bin/env python3
+"""Invariant check for accent-restoration tranches (ruling #16638, 2026-10-03).
+
+A restoration tranche must change NOTHING but accents. Reviewing #18814, the
+coordinator found an artefact neither the review nor the dossier had caught:
+33 mid-sentence capitalisations across 17 cells ("la Même convention",
+"l'Inférence bayesienne", "le second Paramètre"). No ACCENT_PAIRS entry maps a
+lowercase form to a capitalised one, so the tooling was not the cause -- the
+tranche *procedure* was. The ruling therefore asks for a per-cell organ:
+
+    strip_accents(base) == strip_accents(head)   for every markdown cell.
+
+strip_accents drops combining marks but NOT case, so the equality is exact on
+everything except accents: capitalisation, rewording, added/removed words,
+renumbering -- all of it breaks the invariant and is reported. Code cells must
+be byte-identical (a prose tranche never touches them), and the cell count and
+each cell type must not move. The comparison base is the tranche's own parent
+commit (--base-sha), never a distant main: renumberings landed between the
+tranche and a later main are NOT tranche changes and must not poison the
+check.
+
+Validated by its false negatives, never by its hits: the #18814 defect
+("la meme convention" -> "la Même convention") MUST be flagged, a clean
+restoration ("theoreme" -> "théorème") MUST pass, and a code-cell edit MUST
+be reported even though it produces no markdown diff.
+
+Usage: python check_accent_restoration_invariants.py <notebook.ipynb>
+                                   --base-sha <sha-or-ref> [--json]
+                                   [--fail-on-findings]
+
+Output: one line per finding `KIND cell #N (excerpt)`, then a total.
+--fail-on-findings exits 2 when any finding exists. Exit 1 is reserved for
+tool failure (git show failed, notebook unreadable) -- a vacuous crash is
+never a clean check.
+"""
+
+import argparse
+import json
+import subprocess
+import sys
+import unicodedata
+from pathlib import Path
+
+# Sibling import (organ-first): the accent-stripping function is the detector's,
+# never a copy. If the two organs ever diverge on what an accent is, they
+# diverge together.
+sys.path.insert(0, str(Path(__file__).resolve().parent))
+from detect_markdown_deaccent import _strip_accents  # noqa: E402
+
+
+def _cell_source(cell: dict) -> str:
+    """Join a cell source given as list-of-lines or plain string."""
+    src = cell.get("source", "")
+    if isinstance(src, list):
+        return "".join(src)
+    return str(src)
+
+
+def _first_diff_context(a: str, b: str, span: int = 30) -> tuple[str, str]:
+    """Excerpt of both strings around their first divergent position."""
+    limit = min(len(a), len(b))
+    i = 0
+    while i < limit and a[i] == b[i]:
+        i += 1
+    lo = max(0, i - span)
+    return a[lo : i + span], b[lo : i + span]
+
+
+def _load_base_notebook(notebook: Path, base_sha: str) -> dict:
+    """Read the notebook as it stands at base_sha via `git show`."""
+    top = subprocess.run(
+        ["git", "rev-parse", "--show-toplevel"],
+        capture_output=True,
+        text=True,
+        encoding="utf-8",
+        errors="replace",
+        check=True,
+    ).stdout.strip()
+    rel = notebook.resolve().relative_to(Path(top)).as_posix()
+    raw = subprocess.run(
+        ["git", "-C", top, "show", f"{base_sha}:{rel}"],
+        capture_output=True,
+        text=True,
+        encoding="utf-8",
+        check=True,
+    ).stdout
+    return json.loads(raw)
+
+
+def _cell_invariant_violations(base_nb: dict, head_nb: dict) -> list[dict]:
+    """Findings for every cell-level change an accent tranche must not make.
+
+    Markdown cells are compared through strip_accents (case-preserving), code
+    cells byte-for-byte; cell count and cell types must not move. Findings are
+    ordered by cell index, kind by kind.
+    """
+    findings: list[dict] = []
+    base_cells = base_nb.get("cells", [])
+    head_cells = head_nb.get("cells", [])
+    if len(base_cells) != len(head_cells):
+        findings.append(
+            {
+                "kind": "CELL_COUNT_CHANGED",
+                "cell": None,
+                "detail": f"base {len(base_cells)} cells vs head {len(head_cells)} cells",
+            }
+        )
+    for i, (b, h) in enumerate(zip(base_cells, head_cells)):
+        b_type = b.get("cell_type", "")
+        h_type = h.get("cell_type", "")
+        if b_type != h_type:
+            findings.append(
+                {
+                    "kind": "TYPE_CHANGED",
+                    "cell": i,
+                    "detail": f"base {b_type} vs head {h_type}",
+                }
+            )
+            continue
+        b_src = _cell_source(b)
+        h_src = _cell_source(h)
+        if b_type == "markdown":
+            if _strip_accents(b_src) != _strip_accents(h_src):
+                base_ctx, head_ctx = _first_diff_context(
+                    _strip_accents(b_src), _strip_accents(h_src)
+                )
+                findings.append(
+                    {
+                        "kind": "MARKDOWN_INVARIANT",
+                        "cell": i,
+                        "detail": f"base ...{base_ctx!r} vs head ...{head_ctx!r}",
+                    }
+                )
+        else:
+            if b_src != h_src:
+                base_ctx, head_ctx = _first_diff_context(b_src, h_src)
+                findings.append(
+                    {
+                        "kind": "CODE_MODIFIED",
+                        "cell": i,
+                        "detail": f"base ...{base_ctx!r} vs head ...{head_ctx!r}",
+                    }
+                )
+    return findings
+
+
+def main(argv: list[str] | None = None) -> int:
+    parser = argparse.ArgumentParser(
+        description="Check that an accent-restoration tranche changed nothing but accents."
+    )
+    parser.add_argument("notebook", type=Path, help="notebook on the tranche head")
+    parser.add_argument(
+        "--base-sha",
+        required=True,
+        help="commit (sha or ref) holding the pre-tranche notebook -- typically the tranche parent",
+    )
+    parser.add_argument("--json", action="store_true", help="emit a JSON report")
+    parser.add_argument(
+        "--fail-on-findings", action="store_true", help="exit 2 when any finding exists"
+    )
+    args = parser.parse_args(argv)
+
+    try:
+        head_nb = json.loads(args.notebook.read_text(encoding="utf-8"))
+        base_nb = _load_base_notebook(args.notebook, args.base_sha)
+    except (OSError, json.JSONDecodeError, subprocess.CalledProcessError) as exc:
+        print(f"ERROR: cannot load notebook or base ({exc})", file=sys.stderr)
+        return 1
+
+    findings = _cell_invariant_violations(base_nb, head_nb)
+    report = {
+        "notebook": str(args.notebook),
+        "base_sha": args.base_sha,
+        "total": len(findings),
+        "findings": findings,
+    }
+    if args.json:
+        print(json.dumps(report, ensure_ascii=False, indent=2))
+    else:
+        for f in findings:
+            cell = "cell #" + str(f["cell"]) if f["cell"] is not None else "cells"
+            print(f"{f['kind']} {cell} ({f['detail']})")
+        print(f"TOTAL {report['total']} finding(s) -- {args.notebook}")
+    if findings and args.fail_on_findings:
+        return 2
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())
diff --git a/scripts/tests/test_check_accent_restoration_invariants.py b/scripts/tests/test_check_accent_restoration_invariants.py
new file mode 100644
index 0000000000..edf32a58bd
--- /dev/null
+++ b/scripts/tests/test_check_accent_restoration_invariants.py
@@ -0,0 +1,161 @@
+"""Tests for check_accent_restoration_invariants (ruling #16638 point 1).
+
+Validated by false negatives: the #18814 capitalisation defect MUST be
+flagged, a clean restoration MUST pass, a code edit MUST be reported.
+"""
+
+import json
+import subprocess as subprocess_mod
+import sys
+from pathlib import Path
+
+import pytest
+
+sys.path.insert(0, str(Path(__file__).resolve().parents[1] / "notebook_tools"))
+import check_accent_restoration_invariants as cri  # noqa: E402
+
+
+def _nb(cells: list[dict]) -> dict:
+    return {"cells": cells}
+
+
+def _md(src: str) -> dict:
+    return {"cell_type": "markdown", "source": [src]}
+
+
+def _code(src: str) -> dict:
+    return {"cell_type": "code", "source": [src]}
+
+
+class TestCellInvariantViolations:
+    def test_clean_restoration_passes(self):
+        """theoreme -> théorème: accents only, no finding."""
+        base = _nb([_md("Le theoreme de Bayes s'applique ici."), _code("x = 1")])
+        head = _nb([_md("Le théorème de Bayes s'applique ici."), _code("x = 1")])
+        assert cri._cell_invariant_violations(base, head) == []
+
+    def test_18814_capitalisation_defect_is_flagged(self):
+        """The founding defect: 'la meme convention' -> 'la Même convention'."""
+        base = _nb([_md("On suit la meme convention que precedemment.")])
+        head = _nb([_md("On suit la Même convention que precedemment.")])
+        findings = cri._cell_invariant_violations(base, head)
+        assert len(findings) == 1
+        assert findings[0]["kind"] == "MARKDOWN_INVARIANT"
+        assert findings[0]["cell"] == 0
+
+    def test_case_preserved_by_strip_means_case_change_breaks_invariant(self):
+        """Lowercase->capital mid-word, no accent involved at all."""
+        base = _nb([_md("le second parametre vaut 2")])
+        head = _nb([_md("le second Parametre vaut 2")])
+        findings = cri._cell_invariant_violations(base, head)
+        assert [f["kind"] for f in findings] == ["MARKDOWN_INVARIANT"]
+
+    def test_rewording_is_flagged(self):
+        """A real word change (not accents) breaks the invariant."""
+        base = _nb([_md("Le modele est simple")])
+        head = _nb([_md("Le modèle est très simple")])
+        assert [f["kind"] for f in cri._cell_invariant_violations(base, head)] == [
+            "MARKDOWN_INVARIANT"
+        ]
+
+    def test_code_cell_edit_is_flagged(self):
+        """A code edit produces no markdown diff but must be reported."""
+        base = _nb([_md("texte"), _code("resultat = 1")])
+        head = _nb([_md("texte"), _code("resultat = 2")])
+        findings = cri._cell_invariant_violations(base, head)
+        assert [f["kind"] for f in findings] == ["CODE_MODIFIED"]
+        assert findings[0]["cell"] == 1
+
+    def test_cell_count_change_is_flagged(self):
+        base = _nb([_md("a")])
+        head = _nb([_md("a"), _md("b")])
+        findings = cri._cell_invariant_violations(base, head)
+        assert [f["kind"] for f in findings] == ["CELL_COUNT_CHANGED"]
+
+    def test_cell_type_change_is_flagged(self):
+        base = _nb([_md("theoreme")])
+        head = _nb([_code("theoreme")])
+        assert [f["kind"] for f in cri._cell_invariant_violations(base, head)] == [
+            "TYPE_CHANGED"
+        ]
+
+    def test_list_and_string_sources_both_handled(self):
+        """nbformat source is list-of-lines; some tools emit a plain string."""
+        base = _nb([{"cell_type": "markdown", "source": "theoreme ok"}])
+        head = _nb([{"cell_type": "markdown", "source": ["théorème ok"]}])
+        assert cri._cell_invariant_violations(base, head) == []
+
+
+class TestLoadBaseNotebook:
+    def test_git_show_roundtrip(self, tmp_path, monkeypatch):
+        """base notebook is read via git show sha:relpath from the repo root."""
+        nb = {"cells": [{"cell_type": "markdown", "source": ["theoreme"]}]}
+
+        def fake_run(cmd, **kwargs):
+            if cmd[:3] == ["git", "rev-parse", "--show-toplevel"]:
+                return subprocess_mod.CompletedProcess(cmd, 0, stdout=str(tmp_path))
+            assert cmd[:4] == ["git", "-C", str(tmp_path), "show"]
+            assert cmd[4] == "abc123:dir/nb.ipynb"
+            return subprocess_mod.CompletedProcess(
+                cmd, 0, stdout=json.dumps(nb, ensure_ascii=False)
+            )
+
+        monkeypatch.setattr(subprocess_mod, "run", fake_run)
+        nb_path = tmp_path / "dir" / "nb.ipynb"
+        loaded = cri._load_base_notebook(nb_path, "abc123")
+        assert loaded == nb
+
+    def test_git_show_failure_raises(self, tmp_path, monkeypatch):
+        """A missing sha at the base path surfaces as CalledProcessError."""
+
+        def fake_run(cmd, **kwargs):
+            if cmd[:3] == ["git", "rev-parse", "--show-toplevel"]:
+                return subprocess_mod.CompletedProcess(cmd, 0, stdout=str(tmp_path))
+            raise subprocess_mod.CalledProcessError(128, cmd)
+
+        monkeypatch.setattr(subprocess_mod, "run", fake_run)
+        with pytest.raises(subprocess_mod.CalledProcessError):
+            cri._load_base_notebook(tmp_path / "nb.ipynb", "deadbeef")
+
+
+class TestMain:
+    def _write_nb(self, tmp_path, cells: list[dict]) -> Path:
+        path = tmp_path / "nb.ipynb"
+        path.write_text(json.dumps(_nb(cells), ensure_ascii=False), encoding="utf-8")
+        return path
+
+    def test_clean_trancho_exits_zero(self, tmp_path, monkeypatch, capsys):
+        path = self._write_nb(tmp_path, [_md("Le théorème tient.")])
+        monkeypatch.setattr(
+            cri, "_load_base_notebook", lambda p, sha: _nb([_md("Le theoreme tient.")])
+        )
+        rc = cri.main([str(path), "--base-sha", "abc", "--fail-on-findings"])
+        assert rc == 0
+        assert "TOTAL 0" in capsys.readouterr().out
+
+    def test_finding_exits_two(self, tmp_path, monkeypatch, capsys):
+        """The #18814 defect, end to end through main."""
+        path = self._write_nb(tmp_path, [_md("la Même convention")])
+        monkeypatch.setattr(
+            cri, "_load_base_notebook", lambda p, sha: _nb([_md("la meme convention")])
+        )
+        rc = cri.main([str(path), "--base-sha", "abc", "--fail-on-findings"])
+        assert rc == 2
+        out = capsys.readouterr().out
+        assert "MARKDOWN_INVARIANT cell #0" in out
+
+    def test_json_report_shape(self, tmp_path, monkeypatch, capsys):
+        path = self._write_nb(tmp_path, [_md("le Parametre")])
+        monkeypatch.setattr(
+            cri, "_load_base_notebook", lambda p, sha: _nb([_md("le parametre")])
+        )
+        rc = cri.main([str(path), "--base-sha", "abc", "--json"])
+        assert rc == 0
+        report = json.loads(capsys.readouterr().out)
+        assert report["total"] == 1
+        assert report["findings"][0]["kind"] == "MARKDOWN_INVARIANT"
+        assert report["base_sha"] == "abc"
+
+    def test_missing_file_exits_one(self, tmp_path):
+        rc = cri.main([str(tmp_path / "nope.ipynb"), "--base-sha", "abc"])
+        assert rc == 1