diff --git a/evals/dimensional/__init__.py b/evals/dimensional/__init__.py
new file mode 100644
index 00000000..e8985f53
--- /dev/null
+++ b/evals/dimensional/__init__.py
@@ -0,0 +1,7 @@
+"""Dimensional-reasoning lane — the third diversity-panel domain.
+
+The interlingua's unit algebra (``generate.binding_graph.units``) is the system
+under test; an independent dimensional oracle (``evals.dimensional.oracle``) is the
+gold. Structurally distinct from logic/grounding (dimensional exponent arithmetic),
+it exercises the binding-graph interlingua as a load-bearing reasoner beyond GSM8K.
+"""
diff --git a/evals/dimensional/oracle.py b/evals/dimensional/oracle.py
new file mode 100644
index 00000000..df04f216
--- /dev/null
+++ b/evals/dimensional/oracle.py
@@ -0,0 +1,99 @@
+"""Independent dimensional oracle — the gold for the dimensional-reasoning lane.
+
+This is **deliberately a second, independent decision procedure** for "what is the
+dimension of ``left <op> right``?". It carries its **own** unit→base-exponent table
+and its **own** exponent arithmetic and canonical-string renderer, and shares
+**no code** with :mod:`generate.binding_graph.units` (the system under test). Two
+independent dimensional reasoners agreeing on the same cases is real evidence the
+interlingua's unit algebra is correct; a shared-code "oracle" would only prove the
+algebra agrees with itself (INV-25).
+
+Base-dimension order matches the SUT's so the rendered strings are directly
+comparable (the comparison is about the *decision*, not the spelling).
+"""
+
+from __future__ import annotations
+
+from typing import Final
+
+# Same axis order as the SUT's BASE_DIMENSIONS so canonical strings line up.
+_BASE: Final[tuple[str, ...]] = ("length", "time", "mass", "money", "count", "temperature")
+
+# An INDEPENDENT unit→base-exponent table (hand-authored; NOT the en_units_v1 pack
+# the SUT reads). Only concrete base units; composites are resolved structurally.
+_UNIT_DIMS: Final[dict[str, dict[str, int]]] = {
+    "meter": {"length": 1}, "mile": {"length": 1}, "foot": {"length": 1},
+    "second": {"time": 1}, "hour": {"time": 1}, "minute": {"time": 1}, "day": {"time": 1},
+    "kilogram": {"mass": 1}, "pound": {"mass": 1},
+    "dollar": {"money": 1},
+    "degree": {"temperature": 1},
+}
+
+
+class OracleError(ValueError):
+    """Unit outside the oracle's independent vocabulary — refuse, never guess."""
+
+
+def _depluralize(unit: str) -> str:
+    """Independent conservative plural strip (mirrors the SUT's contract, own code)."""
+    if unit in _UNIT_DIMS:
+        return unit
+    for cand in (
+        unit[:-3] + "y" if unit.endswith("ies") and len(unit) > 3 else "",
+        unit[:-2] if unit.endswith("es") and len(unit) > 2 else "",
+        unit[:-1] if unit.endswith("s") and len(unit) > 1 else "",
+    ):
+        if cand and cand in _UNIT_DIMS:
+            return cand
+    return unit
+
+
+def _vector(unit: str) -> tuple[int, ...]:
+    """Resolve a unit id (incl. plural and ``X_per_Y`` composites) to an exponent
+    tuple over ``_BASE``. Raises :class:`OracleError` on anything else."""
+    if not isinstance(unit, str) or not unit:
+        raise OracleError(f"empty unit: {unit!r}")
+    canon = _depluralize(unit)
+    if canon in _UNIT_DIMS:
+        dims = _UNIT_DIMS[canon]
+        return tuple(dims.get(b, 0) for b in _BASE)
+    if "_per_" in unit:
+        num, _, denom = unit.partition("_per_")
+        nv = _vector(num)
+        dv = _vector(denom)
+        return tuple(a - b for a, b in zip(nv, dv))
+    raise OracleError(f"unknown_unit: {unit!r}")
+
+
+def _render(exps: tuple[int, ...]) -> str:
+    """Canonical dimension string, format-compatible with the SUT's renderer."""
+    nums: list[str] = []
+    dens: list[str] = []
+    for dim, e in zip(_BASE, exps):
+        if e > 0:
+            nums.append(dim if e == 1 else f"{dim}^{e}")
+        elif e < 0:
+            dens.append(dim if e == -1 else f"{dim}^{-e}")
+    if not nums and not dens:
+        return "dimensionless"
+    num_part = "*".join(nums) if nums else "1"
+    return num_part if not dens else f"{num_part}/{'*'.join(dens)}"
+
+
+def dimensional_result(op: str, left: str, right: str) -> str:
+    """The oracle verdict: the canonical dimension of ``left <op> right``.
+
+    ``op`` is ``"product"`` or ``"quotient"``. Returns ``"refused"`` if either
+    unit is outside the oracle's vocabulary or ``op`` is unknown — the same
+    refusal posture as the SUT.
+    """
+    try:
+        lv = _vector(left)
+        rv = _vector(right)
+    except OracleError:
+        return "refused"
+    if op == "product":
+        return _render(tuple(a + b for a, b in zip(lv, rv)))
+    if op == "quotient":
+        return _render(tuple(a - b for a, b in zip(lv, rv)))
+    return "refused"
diff --git a/evals/dimensional/runner.py b/evals/dimensional/runner.py
new file mode 100644
index 00000000..c62ff3ad
--- /dev/null
+++ b/evals/dimensional/runner.py
@@ -0,0 +1,80 @@
+"""Dimensional-reasoning lane runner — the interlingua's unit algebra vs. the
+independent dimensional oracle.
+
+The system under test is the binding-graph interlingua (``generate.binding_graph.
+units``): given two unit ids and an operation, what is the dimension of the result?
+Each committed case is scored as ``correct`` (SUT == gold), ``wrong`` (SUT
+committed a different dimension than gold — MUST stay 0), or ``refused`` (SUT
+declined a unit outside the closed vocabulary). The gold is the independent oracle
+verdict, frozen into ``cases.jsonl`` at generation (INV-25).
+"""
+
+from __future__ import annotations
+
+import json
+from collections import Counter
+from pathlib import Path
+
+from generate.binding_graph.units import (
+    UnitAlgebraError,
+    parse_unit,
+    unit_product,
+    unit_quotient,
+)
+
+_ROOT = Path(__file__).resolve().parent
+_REFUSED = "refused"
+
+
+def decide(op: str, left: str, right: str) -> str:
+    """The SUT verdict: the canonical dimension of ``left <op> right`` via the
+    interlingua's unit algebra, or ``"refused"`` for an unknown unit / op."""
+    try:
+        lv = parse_unit(left)
+        rv = parse_unit(right)
+    except UnitAlgebraError:
+        return _REFUSED
+    if op == "product":
+        return unit_product(lv, rv).to_canonical_string()
+    if op == "quotient":
+        return unit_quotient(lv, rv).to_canonical_string()
+    return _REFUSED
+
+
+def _load(path: Path) -> list[dict]:
+    return [json.loads(line) for line in path.read_text(encoding="utf-8").splitlines() if line.strip()]
+
+
+def build_report(cases: list[dict]) -> dict:
+    counts: Counter[str] = Counter({"correct": 0, "wrong": 0, "refused": 0})
+    wrong_examples: list[dict] = []
+    for case in cases:
+        gold = case["gold"]
+        got = decide(case["op"], case["left"], case["right"])
+        if got == gold:
+            counts["refused" if gold == _REFUSED else "correct"] += 1
+        else:
+            counts["wrong"] += 1
+            if len(wrong_examples) < 10:
+                wrong_examples.append({"id": case["id"], "gold": gold, "got": got})
+    return {
+        "n": len(cases),
+        "counts": dict(counts),
+        "all_correct": counts["wrong"] == 0,
+        "wrong_examples": wrong_examples,
+    }
+
+
+def main() -> int:
+    report = build_report(_load(_ROOT / "v1" / "cases.jsonl"))
+    c = report["counts"]
+    print(f"[dimensional] n={report['n']} correct={c['correct']} wrong={c['wrong']} refused={c['refused']}")
+    for w in report["wrong_examples"]:
+        print(f"  WRONG {w['id']}: gold={w['gold']} got={w['got']}")
+    return 0 if report["all_correct"] else 1
+
+
+if __name__ == "__main__":
+    import sys
+
+    sys.exit(main())
diff --git a/evals/dimensional/v1/cases.jsonl b/evals/dimensional/v1/cases.jsonl
new file mode 100644
index 00000000..538bb77f
--- /dev/null
+++ b/evals/dimensional/v1/cases.jsonl
@@ -0,0 +1,12 @@
+{"gold": "length^2", "id": "dim-v1-0001", "left": "meter", "op": "product", "right": "meter"}
+{"gold": "length/time", "id": "dim-v1-0002", "left": "meter", "op": "quotient", "right": "second"}
+{"gold": "length", "id": "dim-v1-0003", "left": "meter_per_second", "op": "product", "right": "second"}
+{"gold": "money/time", "id": "dim-v1-0004", "left": "dollar", "op": "quotient", "right": "hour"}
+{"gold": "time*money", "id": "dim-v1-0005", "left": "dollar", "op": "product", "right": "hour"}
+{"gold": "mass/length", "id": "dim-v1-0006", "left": "kilogram", "op": "quotient", "right": "meter"}
+{"gold": "length*time", "id": "dim-v1-0007", "left": "meter", "op": "product", "right": "second"}
+{"gold": "length/time", "id": "dim-v1-0008", "left": "mile", "op": "quotient", "right": "hour"}
+{"gold": "dimensionless", "id": "dim-v1-0009", "left": "foot", "op": "quotient", "right": "foot"}
+{"gold": "length^2", "id": "dim-v1-0010", "left": "foot", "op": "product", "right": "foot"}
+{"gold": "refused", "id": "dim-v1-0011", "left": "dollar", "op": "quotient", "right": "apple"}
+{"gold": "refused", "id": "dim-v1-0012", "left": "gravitons", "op": "product", "right": "second"}
diff --git a/tests/test_architectural_invariants.py b/tests/test_architectural_invariants.py
index 5ccfad00..26cd7bdc 100644
--- a/tests/test_architectural_invariants.py
+++ b/tests/test_architectural_invariants.py
@@ -1095,6 +1095,14 @@ INDEPENDENT_GOLD_LANES: tuple[IndependentGoldLane, ...] = (
             "generate.logic_equivalence",
         ),
     ),
+    # The dimensional-reasoning lane: the interlingua's own unit algebra
+    # (generate.binding_graph.units) is the SUT, so its gold oracle must share no
+    # code with the binding graph.
+    IndependentGoldLane(
+        name="dimensional",
+        oracle_module="evals/dimensional/oracle.py",
+        sut_import_prefixes=("generate.binding_graph",),
+    ),
 )
 
 _DEDUCTIVE_CASE_FILES: tuple[str, ...] = (
diff --git a/tests/test_dimensional_reasoning_lane.py b/tests/test_dimensional_reasoning_lane.py
new file mode 100644
index 00000000..458c8d12
--- /dev/null
+++ b/tests/test_dimensional_reasoning_lane.py
@@ -0,0 +1,73 @@
+"""Dimensional-reasoning lane — the third diversity-panel domain.
+
+Proves the binding-graph interlingua's unit algebra decides the dimension of a
+unit operation identically to an independent dimensional oracle (sharing no code
+with it), on every committed case — the same independent-gold discipline as the
+deductive and finite-entity lanes, applied to a structurally distinct domain.
+"""
+
+from __future__ import annotations
+
+import json
+from pathlib import Path
+
+import pytest
+
+from evals.dimensional.oracle import dimensional_result
+from evals.dimensional.runner import decide
+
+_FIXTURE = Path(__file__).resolve().parents[1] / "evals" / "dimensional" / "v1" / "cases.jsonl"
+
+
+def _load() -> list[dict]:
+    return [json.loads(line) for line in _FIXTURE.read_text(encoding="utf-8").splitlines() if line.strip()]
+
+
+@pytest.mark.parametrize("case", _load(), ids=lambda c: c["id"])
+def test_interlingua_agrees_with_independent_oracle_and_gold(case: dict) -> None:
+    """The lane gate: the interlingua's unit algebra (SUT) == the independent
+    oracle == the committed gold, with wrong == 0 by construction."""
+    oracle = dimensional_result(case["op"], case["left"], case["right"])
+    sut = decide(case["op"], case["left"], case["right"])
+    assert oracle == case["gold"], f"{case['id']}: independent oracle != committed gold"
+    assert sut == case["gold"], f"{case['id']}: interlingua committed a wrong dimension (wrong=0 breach)"
+
+
+def test_fixture_has_nontrivial_and_refusal_signal() -> None:
+    """Guard against a vacuous fixture: it must carry composite dimensions AND a
+    refusal, or the lane proves nothing."""
+    golds = [c["gold"] for c in _load()]
+    assert any("/" in g or "*" in g or "^" in g for g in golds), "no composite dimensions"
+    assert "refused" in golds, "no refusal-boundary case"
+
+
+def test_oracle_is_independent_of_the_sut() -> None:
+    """The oracle must not IMPORT the interlingua it checks (AST, not substring —
+    the docstring may name the SUT). Also enforced structurally by INV-25's
+    registry; this is a fast local sanity check."""
+    import ast
+    import evals.dimensional.oracle as oracle_mod
+
+    path = oracle_mod.__file__
+    assert path is not None
+    tree = ast.parse(Path(path).read_text(encoding="utf-8"))
+    imported: set[str] = set()
+    for node in ast.walk(tree):
+        if isinstance(node, ast.Import):
+            imported |= {a.name for a in node.names}
+        elif isinstance(node, ast.ImportFrom) and node.module:
+            imported.add(node.module)
+    assert not any(m.startswith("generate.binding_graph") for m in imported), (
+        "oracle must share no code with the SUT (generate.binding_graph)"
+    )
+
+
+def test_dimensionless_and_distinct_units_same_dimension() -> None:
+    # mile/hour and meter/second are different units, same dimension.
+    assert decide("quotient", "mile", "hour") == decide("quotient", "meter", "second")
+    assert decide("quotient", "foot", "foot") == "dimensionless"
+
+
+def test_unknown_unit_refuses() -> None:
+    assert decide("product", "gravitons", "second") == "refused"
+    assert dimensional_result("product", "gravitons", "second") == "refused"