Release/v0.2.0

CHANGELOG.md

@@ -3,7 +3,8 @@
 ## [0.2.0] — 2026-10-22
 
 Every deterministic model now gives the same doubles as TerraME at every
-step (checked against `terrame-goldens` v0.1.0). The trajectories of
+step, except Daisyworld, which agrees to about 15 significant digits (checked
+against `terrame-goldens` v0.1.0). The trajectories of
 `Daisyworld` and `MonoLake` change.
 
 - **Fix:** `Daisyworld` now follows TerraME: after updating the white area,
@@ -20,9 +21,13 @@ step (checked against `terrame-goldens` v0.1.0). The trajectories of
 - `PredatorPrey`, `RoomTemperature`, `SIR` and `MonoLake` do their
   arithmetic in TerraME's order, so that the results are the same doubles.
 - `tests/test_goldens.py` (marker `golden`): every model against TerraME
-  2.0.1 runs from `terrame-goldens` v0.1.0. All 16 deterministic cases
-  (13 models, some with two parameter sets) give **the same doubles as
-  TerraME at every step**; RandomWalk is compared by distribution.
+  2.0.1 runs from `terrame-goldens` v0.1.0, generated in the
+  `terrame-docker` image. 15 of the 16 deterministic cases (13 models, some
+  with two parameter sets) give **the same doubles as TerraME at every
+  step**. Daisyworld is compared with a relative tolerance of 10⁻¹³: its
+  temperatures use the C library's `pow`, whose last bit differs between
+  glibc 2.27 (the image) and newer versions. RandomWalk is compared by
+  distribution.
   `tests/goldens/` holds the goldens, their SHA-256 provenance and the
   script that copies them. The strict `xfail`s of `Daisyworld` and
   `MonoLake` in `test_terrame_parity.py` are removed.

README.md

@@ -31,8 +31,11 @@ Every model was run in TerraME 2.0.1 with the `sysdyn` package
 ([`terrame-goldens`](https://github.com/LambdaGeo/terrame-goldens) v0.1.0).
 For the 16 deterministic cases (13 models, some with two parameter sets),
 the Python models give **the same doubles as TerraME at every step**, bit for
-bit, for every variable TerraME records; RandomWalk is compared by
-distribution. `tests/test_goldens.py` runs these comparisons
+bit, for every variable TerraME records, except Daisyworld, whose fourth
+roots go through the C library's `pow`: its last bit depends on the system
+(the goldens come from the `terrame-docker` image, with glibc 2.27), so
+Daisyworld agrees to about 15 significant digits and is compared with a
+relative tolerance of 10⁻¹³. RandomWalk is compared by distribution. `tests/test_goldens.py` runs these comparisons
 (`pytest -m golden`, under a second).
 
 ## Examples

tests/goldens/reference.json

@@ -17,7 +17,7 @@
   "files": {
     "goldens/sysdyn/chaotic_growth.csv.gz": "6328b3248918b88ec4adfc1002a99ec6c4d252fa1e6a3beaa0c1477faa528c75",
     "goldens/sysdyn/coffee.csv.gz": "cf32769b784339a21489a54ef05f1e59c403da6857fa7421dd28e05f1581a4c1",
-    "goldens/sysdyn/daisyworld.csv.gz": "da394371bf4dc84114ae5c5815797d20aaa7fd85b6e7b182446e180f7d747c6c",
+    "goldens/sysdyn/daisyworld.csv.gz": "1b8668e090f51550b12562d5af517d2fe4b2845e1ec68381c4eb21e9bee516b4",
     "goldens/sysdyn/homeostasis.csv.gz": "2cb49a621d7153e98632f7f0e11bfdc47ea04a105a2f659d6976bfa3dfc0fd4f",
     "goldens/sysdyn/homeostasis_100.csv.gz": "6896a8faed843dc0c42f58194b284381fad971e793bed8de531ef26282ac280b",
     "goldens/sysdyn/limited_growth.csv.gz": "9972047016a135b12c2e70e26209d438e1b911dd44ac927c47f4b8eab5f28b87",

tests/test_goldens.py

@@ -8,6 +8,14 @@
 **the same doubles** as TerraME at every step (bit for bit), for every
 variable TerraME records.
 
+The exception is Daisyworld, whose temperatures take a fourth root
+(``x ** 0.25`` in Python, ``x ^ 0.25`` in Lua). Both languages call the C
+library's ``pow``, and its last bit depends on the library: the goldens come
+from the ``terrame-docker`` image (glibc 2.27), and the same TerraME binary
+on glibc 2.39 differs from the 16th significant digit on (see the README of
+``terrame-goldens``). Python on any recent system behaves like the latter, so
+Daisyworld is compared with a relative tolerance of ``POW_RTOL``.
+
 RandomWalk draws random numbers, so it is compared by distribution with the
 50 TerraME runs.
 """
@@ -28,6 +36,9 @@
 
 GOLDENS = Path(__file__).resolve().parent / "goldens"
 
+#: Cases that use the C library's pow, and their relative tolerance (see above).
+POW_RTOL = {"daisyworld": 1e-13}
+
 with open(GOLDENS / "cases.csv", newline="") as _f:
     CASES = [c for c in csv.DictReader(_f) if not c["seeds"]]
 
@@ -63,7 +74,7 @@ def execute(self) -> None:
 
 
 @pytest.mark.parametrize("case", CASES, ids=[c["case"] for c in CASES])
-def test_trajectory_is_bit_identical_to_terrame(case: dict[str, str]) -> None:
+def test_trajectory_matches_terrame(case: dict[str, str]) -> None:
     variables, ref = load(case["case"])
     attrs = [snake(v) for v in variables]
     env = Environment(start_time=1, end_time=int(case["final_time"]))
@@ -75,7 +86,12 @@ def test_trajectory_is_bit_identical_to_terrame(case: dict[str, str]) -> None:
         for var, e, g in zip(variables, expected, got, strict=True):
             if np.isnan(e):  # not defined yet in TerraME (Daisyworld at step 0)
                 continue
-            assert g == e, f"{case['case']}.{var} at step {t}: {g!r} != TerraME {e!r}"
+            rtol = POW_RTOL.get(case["case"])
+            if rtol is None:
+                assert g == e, f"{case['case']}.{var} at step {t}: {g!r} != TerraME {e!r}"
+            else:
+                assert abs(g - e) <= rtol * abs(e), (
+                    f"{case['case']}.{var} at step {t}: {g!r} vs TerraME {e!r}")
 
 
 def test_random_walk_matches_terrame_distribution() -> None: