diff --git a/tests/test_experts_confirmed_gaps.py b/tests/test_experts_confirmed_gaps.py
new file mode 100644
index 0000000..30eabcd
--- /dev/null
+++ b/tests/test_experts_confirmed_gaps.py
@@ -0,0 +1,330 @@
+"""Behaviour of ``semantic/experts.py`` that no test pinned.
+
+Each test here closes a mutant that survived the *full* suite (not just a
+targeted subset) on ``28ce1c3``. Every one was also checked to be observable:
+the mutated code produces different output on the input used below, so these
+are real gaps, not equivalent mutants.
+"""
+from __future__ import annotations
+
+import itertools
+import math
+
+import pandas as pd
+import pytest
+
+import freshdata as fd
+from freshdata.semantic import experts
+from freshdata.semantic.experts import (
+    _resolve_date,
+    looks_like_identifier_value,
+    parse_currency_parts,
+    parse_number_words,
+)
+
+# ── accounting negatives need BOTH parentheses (bool#10) ────────────────────
+
+
+@pytest.mark.parametrize("value", ["(1,200", "(1.50", "1,200)", "($5", "(€1.200,50"])
+def test_an_unbalanced_parenthesis_is_not_an_accounting_negative(value):
+    """``(1,200)`` is -1200; ``(1,200`` is malformed and must not parse.
+
+    Treating a lone opening parenthesis as the accounting-negative marker
+    strips the first *and last* characters, so ``"(1,200"`` became ``"1,20"``
+    and was read as -1.2, unambiguous -- a malformed value silently turned
+    into a confident wrong number.
+    """
+    assert parse_currency_parts(value) == (None, False)
+
+
+def test_a_balanced_accounting_negative_still_parses():
+    """Control for the test above."""
+    assert parse_currency_parts("(1,200.50)") == (-1200.5, False)
+
+
+# ── no ISO churn when the column cannot convert (cmp#33) ────────────────────
+
+
+def test_valid_iso_dates_are_left_alone_beside_an_unresolvable_value():
+    """Canonical ISO strings are re-encoded only alongside a *low-risk* repair.
+
+    Here the only non-ISO value is day/month-ambiguous (high risk), so the
+    column cannot convert as a whole. Rewriting the valid ISO dates to
+    timestamps anyway would be pure representation churn, and would leave an
+    object column mixing ``Timestamp``s with the one string still under
+    review. ``fix_dtypes=False`` keeps the values as strings until the
+    semantic stage; with it on, the dtype step converts the column first and
+    this path is never reached.
+    """
+    values = ["2021-03-15", "2021-04-20", "2021-06-30", "05/01/2021"] * 5
+    result = fd.clean(pd.DataFrame({"d": values}), semantic_mode="auto",
+                      fix_dtypes=False, verbose=False)
+
+    assert result.data["d"].tolist() == values
+    semantic = [a for a in result.report().actions if a.step == "semantic"]
+    assert [a.status for a in semantic] == ["suggested"]
+    assert "05/01/2021" in semantic[0].description
+
+
+# ── a bare "hundred" means one hundred (const_num#34) ──────────────────────
+
+
+@pytest.mark.parametrize("text,expected", [
+    ("hundred", 100), ("hundred and five", 105), ("hundred thousand", 100_000),
+])
+def test_a_bare_hundred_counts_as_one_hundred(text, expected):
+    """``hundred`` with no count before it is one hundred, like ``thousand``.
+
+    Existing tests only used ``"one hundred"``, where the explicit count hides
+    the default: ``(current or 1) * 100`` could default to any number and no
+    test noticed.
+    """
+    assert parse_number_words(text) == expected
+
+
+# ── lakh grouping checks every inner group (const_num#46) ──────────────────
+
+
+@pytest.mark.parametrize("value", ["₹12,3,456", "$1,2,345", "₹1,2,34,567"])
+def test_a_malformed_inner_group_is_rejected(value):
+    """Indian grouping is 2-digit groups then a final 3-digit group.
+
+    A group of one digit anywhere before the last is malformed and must not
+    parse. Only the first inner group was being checked by any test, so a
+    grouping check that skipped the one before last went unnoticed and
+    ``"$1,2,345"`` would have read as 12345 -- unambiguous.
+    """
+    assert parse_currency_parts(value) == (None, False)
+
+
+@pytest.mark.parametrize("value,expected", [
+    ("₹1,23,456", 123456.0), ("₹12,34,567.50", 1234567.5),
+])
+def test_well_formed_lakh_grouping_still_parses(value, expected):
+    """Control for the test above."""
+    assert parse_currency_parts(value) == (expected, False)
+
+
+# ── a grouping separator cannot lead (const_num#38) ────────────────────────
+
+
+@pytest.mark.parametrize("value", ["$,123", "€,123.45", "£,000"])
+def test_a_leading_grouping_separator_is_rejected(value):
+    """``,123`` has a thousands separator with no thousands before it.
+
+    The first group is required to be non-empty; checking the second instead
+    would have accepted these as 123 and 123.45.
+    """
+    assert parse_currency_parts(value) == (None, False)
+
+
+def test_a_leading_decimal_point_is_still_a_decimal():
+    """Control: ``.500`` is a decimal, not a grouping, and parses."""
+    assert parse_currency_parts("$.500") == (0.5, False)
+
+
+# ── an accounting negative of exactly one, and of zero (const_num#55) ──────
+
+
+@pytest.mark.parametrize("value", ["($1.00)", "(1.00)", "(€1,00)"])
+def test_an_accounting_negative_of_one_is_minus_one(value):
+    """The zero special case must not swallow a debit of exactly one.
+
+    ``-abs(value) if value != 0 else 0.0`` exists only to keep zero from
+    becoming ``-0.0``. Keyed on the wrong value it would turn a one-unit
+    debit into 0.0 -- a dollar silently erased -- and no test had a debit of
+    exactly one.
+    """
+    assert parse_currency_parts(value)[0] == -1.0
+
+
+def test_an_accounting_negative_zero_is_positive_zero():
+    """``($0.00)`` is zero, not negative zero (which prints as ``-0.0``)."""
+    value, _ = parse_currency_parts("($0.00)")
+    assert value == 0.0
+    assert math.copysign(1.0, value) == 1.0
+
+
+# ── date resolutions: a valid confidence, and the partial-date guess ───────
+
+
+def _resolve(raw: str, dayfirst: bool | None = None, reference_date: str | None = None):
+    return _resolve_date(raw, dayfirst=dayfirst, reference_date=reference_date)
+
+
+@pytest.mark.parametrize("raw,kwargs", [
+    ("2021-03-15", {}), ("2021/03/15", {}), ("March 5, 2021", {}), ("5 March 2021", {}),
+    ("13/01/2021", {}), ("05/01/2021", {}), ("05/01/2021", {"dayfirst": True}),
+    ("2025-03", {}), ("yesterday", {}), ("yesterday", {"reference_date": "2021-03-15"}),
+    ("yesterday", {"reference_date": "not a date"}),
+])
+def test_every_date_resolution_has_a_valid_confidence(raw, kwargs):
+    """A confidence is a score in [0, 1] at the point it is produced (const_num#91).
+
+    The report clamps downstream, which hid an ISO-date confidence of 1.03
+    from every existing test; the resolver's own output must be valid.
+    """
+    res = _resolve(raw, **kwargs)
+    assert res is not None
+    assert 0.0 <= res.confidence <= 1.0
+
+
+def test_a_partial_date_offers_the_first_of_that_month_for_review():
+    """``2025-03`` has no day: never applied, but the reviewer sees 1 March 2025.
+
+    Reading the month group as the year (const_num#92) turned the guess into
+    year 0003 and no test looked at the guess itself.
+    """
+    res = _resolve("2025-03")
+    assert res.value == pd.Timestamp(2025, 3, 1)
+    assert res.risk == "high"
+
+
+# ── the leading-zero identifier veto (const_num#113) ───────────────────────
+
+
+@pytest.mark.parametrize("value,expected", [
+    ("05", True), ("0123", True), ("007", True), ("00", True),
+    ("10", False), ("100", False), ("1", False), ("0", False),
+])
+def test_a_leading_zero_marks_a_value_as_an_identifier(value, expected):
+    """``05`` and ``0123`` are codes; ``10`` and ``100`` are numbers.
+
+    This veto keeps the spelled-number and category experts off zero-padded
+    codes. Every existing test used ``"007"`` -- which has a zero in its first
+    *and* second place -- so checking the wrong character position went
+    unnoticed: ``05`` and ``0123`` lost their protection and ``10`` gained it.
+    """
+    assert looks_like_identifier_value(value) is expected
+
+
+# ── every expert proposes with a valid base confidence (const_num#120, #129) ──
+
+_EXPERT_CORPUS = {
+    "amount": ["$15.00", "$1,200.50", "(1,200)", "$20.00", "€1.234,56", "$1.200"],
+    "price": ["$15.00", "£20.00", "€30.00", "$40.00", "¥500", "$60.00"],
+    "weight_kg": ["5 kg", "12 kg", "7 lb", "3 kg", "8 kg", "10 g"],
+    "active": ["yes", "no", "Y", "N", "true", "0"],
+    "gender": ["M", "F", "male", "Female", "m", "f"],
+    "status": ["Active", "active", "ACTIVE", "Inactive", "inactive", "Pending"],
+    "qty": ["three", "twelve", "forty two", "7", "one hundred", "9"],
+    "when": ["yesterday", "2021-03-15", "today", "March 5, 2021", "2021-04-01", "05/01/2021"],
+}
+
+
+def test_every_expert_proposes_with_a_base_confidence_in_range(monkeypatch):
+    """A base confidence is a score in [0, 1] where it is produced.
+
+    ``confidence_from_evidence`` clamps to 0.999 downstream, so a literal of
+    1.01 in an expert (the currency and unit experts' 0.96, mutated) reached
+    no report and failed no test. This checks the contract at the source, for
+    every proposal every expert makes on a corpus touching each of them.
+    """
+    seen: list[float] = []
+    real = experts.make_proposal
+
+    def spy(**kwargs):
+        seen.append(kwargs["base_confidence"])
+        return real(**kwargs)
+
+    monkeypatch.setattr(experts, "make_proposal", spy)
+    for column, values in _EXPERT_CORPUS.items():
+        fd.clean(pd.DataFrame({column: values * 6}), semantic_mode="auto",
+                 fix_dtypes=False, verbose=False)
+    assert len(seen) >= 30, f"corpus reached too few proposals ({len(seen)})"
+    out_of_range = sorted({c for c in seen if not 0.0 <= c <= 1.0})
+    assert not out_of_range, out_of_range
+
+
+# ── an empty body is not ambiguous (const_bool#12) ──────────────────────────
+
+
+@pytest.mark.parametrize("value", ["", "   ", "()", "( )"])
+def test_nothing_between_the_parentheses_is_neither_money_nor_ambiguous(value):
+    """``ambiguous`` describes a value; with no value there is nothing to be unsure of."""
+    assert parse_currency_parts(value) == (None, False)
+
+
+# ── equivalence proofs: survivors that cannot be killed ─────────────────────
+#
+# These are deliberately not source-text tripwires. The mutation harness
+# writes each mutant through ``ast.unparse``, which reformats the whole file,
+# so a test that greps ``inspect.getsource`` fails on *every* mutant and would
+# report nearly all of them killed. Each proof quotes its premise instead.
+
+
+def test_the_decimal_separator_tiebreak_cannot_tie():
+    """``cmp#7``: ``rfind(".") > rfind(",")`` -> ``>=`` is equivalent.
+
+    The comparison sits under ``if dots and commas:``, so both characters are
+    present and each ``rfind`` is a real index. Two different characters never
+    share an index, so the two sides are never equal and ``>`` and ``>=``
+    agree. Checked exhaustively over every body of length <= 7 on the
+    alphabet that matters.
+    """
+    for n in range(2, 8):
+        for body in map("".join, itertools.product("1.,", repeat=n)):
+            if "." in body and "," in body:
+                assert body.rfind(".") != body.rfind(",")
+
+
+def test_the_single_separator_tail_ignores_maxsplit():
+    """``const_num#47``: ``body.rsplit(sep, 1)[1]`` -> ``rsplit(sep, 2)`` is equivalent.
+
+    ``tail`` is read only in the branch where the separator occurs exactly
+    once (``occurrences > 1`` short-circuits to grouping first). With one
+    occurrence, ``rsplit`` yields the same two parts for any ``maxsplit >= 1``.
+    Checked exhaustively over every body of length <= 7.
+    """
+    for sep in ".,":
+        for n in range(1, 8):
+            for body in map("".join, itertools.product(f"1{sep}", repeat=n)):
+                if body.count(sep) == 1:
+                    assert body.rsplit(sep, 1)[1] == body.rsplit(sep, 2)[1]
+
+
+def test_the_integer_part_ignores_maxsplit():
+    """``const_num#51``: ``split(decimal_sep, maxsplit=1)[0]`` -> ``maxsplit=2`` is equivalent.
+
+    Element ``[0]`` is the text before the *first* separator, whatever the
+    ``maxsplit`` (for any ``maxsplit >= 1``). Checked exhaustively over every
+    body of length <= 7.
+    """
+    for sep in ".,":
+        for n in range(1, 8):
+            for body in map("".join, itertools.product("1.,", repeat=n)):
+                assert body.split(sep, maxsplit=1)[0] == body.split(sep, maxsplit=2)[0]
+
+
+def test_a_thirteenth_month_is_rejected_whichever_branch_reads_it():
+    """``const_num#105``/``#107``: ``b <= 12`` -> ``b <= 13`` (and ``a``) are equivalent.
+
+    The day/month decision is ``a > 12 and b <= 12`` (day-first), then
+    ``b > 12 and a <= 12`` (month-first), then ``a > 12 and b > 12`` -> not a
+    date. Loosening 12 to 13 changes only inputs where the would-be month is
+    exactly 13. The original sends those to "not a date"; the mutant assigns
+    month 13, and ``_safe_timestamp`` rejects month 13, so the result is
+    ``None`` either way. Checked for every day and a range of years.
+    """
+    for year in (1970, 2000, 2024, 2099):
+        for day in range(0, 100):
+            assert experts._safe_timestamp(year, 13, day) is None
+    for other in range(13, 100):
+        assert _resolve(f"{other}/13/2024") is None
+        assert _resolve(f"13/{other}/2024") is None
+
+
+def test_a_grouping_check_never_sees_an_empty_remainder():
+    """``const_bool#5``: ``if not rest: return True`` -> ``False`` is equivalent.
+
+    ``_valid_grouping`` returns early unless ``sep in part``. Its only caller
+    passes ``group_sep``, which is always ``"."`` or ``","``, and
+    ``lstrip("+-")`` cannot remove either, so ``split(sep)`` yields at least
+    two groups and ``rest`` is never empty. Checked exhaustively over every
+    part of length <= 7 on the alphabet that matters.
+    """
+    for sep in ".,":
+        for n in range(1, 8):
+            for part in map("".join, itertools.product("1.,+-", repeat=n)):
+                if sep in part:
+                    assert len(part.lstrip("+-").split(sep)[1:]) >= 1