Add Control v2 incremental experiment results (evidence, visuals, 42 fixed samples)
a64600b verified Download method_v2/evidence/code/suite/tests/test_null_commitment.py from CharlesZhang-USTC/snf_experiments: direct link, hf CLI and curl.
- Browser
- Download file 4.33 kB
-
https://huggingface.co/CharlesZhang-USTC/snf_experiments/resolve/main/method_v2/evidence/code/suite/tests/test_null_commitment.py
- Command line
-
hf download hf://CharlesZhang-USTC/snf_experiments/method_v2/evidence/code/suite/tests/test_null_commitment.py
-
curl -L -o test_null_commitment.py https://huggingface.co/CharlesZhang-USTC/snf_experiments/resolve/main/method_v2/evidence/code/suite/tests/test_null_commitment.py
4.33 kB
| """Analytical and independent-seed checks; no neural-model dependencies.""" | |
| import unittest | |
| import numpy as np | |
| from suite.null_commitment import ( | |
| ORDERS, commitment_steps, exact_curve, fixed_order, monte_carlo, | |
| oracle_remap, response_statistics, target_support, | |
| ) | |
| class IndependentTokenNullTests(unittest.TestCase): | |
| def setUp(self): | |
| self.mask = target_support(256) | |
| self.alpha = np.linspace(0, 1, 65) | |
| self.steps = 18 | |
| self.b = .15 | |
| def commit(self, order): | |
| return commitment_steps(fixed_order(self.mask, order), self.steps) | |
| def test_orders_preserve_positions_and_reveal_budget(self): | |
| counts = [] | |
| for order in ORDERS: | |
| permutation = fixed_order(self.mask, order) | |
| np.testing.assert_array_equal(np.sort(permutation), np.arange(256)) | |
| commit = commitment_steps(permutation, self.steps) | |
| self.assertTrue(np.all(commit >= 0)) | |
| counts.append(np.bincount(commit, minlength=self.steps)) | |
| for count in counts[1:]: | |
| np.testing.assert_array_equal(count, counts[0]) | |
| balanced = fixed_order(self.mask, "balanced") | |
| np.testing.assert_array_equal(np.cumsum(self.mask[balanced]), np.arange(1, 257) * 64 // 256) | |
| def test_zero_coupling_can_have_narrow_duration_response(self): | |
| first = exact_curve(self.alpha, self.commit("target_first"), self.mask, self.steps, self.b) | |
| random = exact_curve(self.alpha, self.commit("random"), self.mask, self.steps, self.b) | |
| fs = response_statistics(self.alpha, first["target_match"]) | |
| rs = response_statistics(self.alpha, random["target_match"]) | |
| self.assertLess(fs["width80"], .3) | |
| self.assertGreater(rs["width80"], .55) | |
| self.assertGreater(rs["width80"] / fs["width80"], 2) | |
| self.assertAlmostEqual(fs["endpoint_gain"], 1 - self.b) | |
| self.assertAlmostEqual(rs["endpoint_gain"], 1 - self.b) | |
| def test_amplitude_schedule_is_linear_for_every_order(self): | |
| for order in ORDERS: | |
| curve = exact_curve(self.alpha, self.commit(order), self.mask, self.steps, | |
| self.b, "amplitude") | |
| np.testing.assert_allclose(curve["target_match"], self.b + (1 - self.b) * self.alpha, atol=1e-14) | |
| self.assertAlmostEqual(response_statistics(self.alpha, curve["target_match"])["width80"], .8) | |
| def test_context_only_is_dormant_and_outside_support_is_untouched(self): | |
| commit = self.commit("target_first") | |
| for copy in (True, False): | |
| curve = exact_curve(self.alpha, commit, self.mask, self.steps, self.b, | |
| reference_copy=copy) | |
| np.testing.assert_allclose(curve["outside_match"], self.b, atol=1e-14) | |
| np.testing.assert_array_equal(curve["direct_copy_count_outside"], 0) | |
| if not copy: | |
| np.testing.assert_allclose(curve["target_match"], self.b, atol=1e-14) | |
| self.assertIsNone(response_statistics(self.alpha, curve["target_match"])["width80"]) | |
| mc = monte_carlo(self.alpha, commit, self.mask, self.steps, self.b, "duration", 32) | |
| self.assertEqual(mc["outside_changed_token_observations"], 0) | |
| def test_oracle_inverse_recovers_continuous_intermediate_expectations(self): | |
| for order in ORDERS: | |
| oracle = oracle_remap(self.alpha, self.commit(order), self.mask, self.steps, self.b) | |
| self.assertLess(oracle["max_absolute_error"], 1e-12) | |
| self.assertIn("not held-out", oracle["caveat"]) | |
| np.testing.assert_allclose(oracle["achieved_level"], self.alpha, atol=1e-12) | |
| def test_monte_carlo_agrees_with_exact_without_changing_outside(self): | |
| commit = self.commit("random") | |
| for schedule in ("duration", "amplitude"): | |
| exact = exact_curve(self.alpha, commit, self.mask, self.steps, self.b, schedule) | |
| mc = monte_carlo(self.alpha, commit, self.mask, self.steps, self.b, schedule, 1024) | |
| error = np.abs(mc["target_match_mean"] - exact["target_match"]) | |
| self.assertTrue(np.all(error < 5 * mc["target_match_se"] + 1e-12)) | |
| self.assertLess(float(error.max()), .01) | |
| self.assertEqual(mc["outside_changed_token_observations"], 0) | |
| if __name__ == "__main__": | |
| unittest.main() | |