import json import zipfile import pytest import torch from aether.experiments.study import StudyConfig, run_study, replay_checkpoint def small_config(): return StudyConfig(seeds=[3], protocols=['scalar', 'wave'], updates=2, batch_size=2, evaluation_episodes=3, model={'neuron_count': 16, 'scalar_edges': 32, 'region_count': 4, 'region_message_edges': 8, 'state_dimension': 4, 'mode_count': 2, 'complex_dimensions': 2}) def test_study_trains_replays_and_preserves_evidence(tmp_path): config = small_config() result = run_study(config, tmp_path / 'study') assert result['status'] == 'exploratory' assert len(result['runs']) == 2 for run in result['runs']: assert run['initial_digest'] != run['trained_digest'] assert run['query_immutable'] assert run['executed_support_steps'] > 0 assert run['evaluation']['executed_steps'] > 0 replay = replay_checkpoint(tmp_path / 'study' / run['checkpoint']) assert replay['valid'] assert replay['evaluation'] == run['evaluation'] assert replay['retention'] == run['retention'] assert replay['interventions'] == run['interventions'] assert all(s >= 10_000_000 for s in run['evaluation']['episode_seeds']) stored = json.loads((tmp_path / 'study' / 'results.json').read_text()) assert not stored['promotion']['promoted'] with pytest.raises(FileExistsError): run_study(config, tmp_path / 'study') checkpoint = tmp_path / 'study' / result['runs'][-1]['checkpoint'] saved = torch.load(checkpoint, weights_only=True) saved['retention']['episodes'][0]['success'] = not saved['retention']['episodes'][0]['success'] changed = tmp_path / 'changed.pt' torch.save(saved, changed) assert not replay_checkpoint(changed)['valid'] def test_study_source_snapshot_reconstructs_its_recorded_fingerprint(tmp_path): from aether.evidence.receipts import source_fingerprint output = tmp_path / 'study' result = run_study(small_config(), output) restored = tmp_path / 'source' with zipfile.ZipFile(output / 'source-snapshot.zip') as archive: assert 'aether/live/web/app.js' in archive.namelist() assert 'aether/live/web/styles.css' in archive.namelist() assert 'aether/live/web/index.html' in archive.namelist() archive.extractall(restored) assert source_fingerprint(restored) == result['metadata']['source_fingerprint'] @pytest.mark.parametrize('task_family', ['straight', 'distraction', 'waypoint']) def test_study_replays_selected_task_and_records_its_training_inputs(tmp_path, task_family): from aether.environments.stress import stress_expert_batch from aether.kernel.runtime import TissueModel from aether.config import ModelConfig from torch.nn import functional as F config = small_config() config.task_family = task_family output = tmp_path / task_family # A wrong support-task route changes the first loss on the actual expert # observations, even if a short-trained policy happens to choose the same actions. model = TissueModel(ModelConfig(**config.model), 'wave', seed=3) support = stress_expert_batch([300_000, 300_001], task_family=task_family, delay=3, horizon=9) logits = model(torch.from_numpy(support['observations']), disturbance_seed=3000) mask = torch.from_numpy(support['mask']) expected_loss = float(F.cross_entropy(logits[mask], torch.from_numpy(support['actions'])[mask]).detach()) result = run_study(config, output) assert result['config']['task_family'] == task_family assert result['runs'][1]['training_curve'][0]['loss'] == expected_loss for run in result['runs']: path = output / run['checkpoint'] saved = torch.load(path, weights_only=True) assert saved['task_family'] == task_family assert replay_checkpoint(path)['valid'] def test_replay_defaults_legacy_checkpoint_to_straight_task(tmp_path): result = run_study(small_config(), tmp_path / 'study') checkpoint = tmp_path / 'study' / result['runs'][1]['checkpoint'] saved = torch.load(checkpoint, weights_only=True) saved.pop('task_family', None) legacy = tmp_path / 'legacy.pt' torch.save(saved, legacy) assert replay_checkpoint(legacy)['valid'] def test_evaluation_executes_distraction_and_waypoint_contracts(): from aether.experiments.study import evaluate class CueFollower(torch.nn.Module): """Keep the initial direction, but visibly respond to later cue pulses.""" def state_digest(self): return 'fixed-policy' def initial_state(self, batch_size): return torch.zeros(batch_size, 6) def step(self, observation, state, **kwargs): cue = observation[:, :6] state = torch.where(cue.sum(-1, keepdim=True) > 0, cue, state) return torch.cat((state, state.new_zeros(state.shape[0], 1)), -1), state model = CueFollower() straight = evaluate(model, [11], delay=3, horizon=9) distraction = evaluate(model, [11], delay=3, horizon=9, task_family='distraction') waypoint = evaluate(model, [11], delay=3, horizon=9, task_family='waypoint') assert straight['episodes'][0]['success'] is True assert distraction['episodes'][0]['actions'][1] != straight['episodes'][0]['actions'][1] assert waypoint['episodes'][0]['actions'] != straight['episodes'][0]['actions'] @pytest.mark.parametrize('kwargs', [ {'updates': 0}, {'seeds': [1, 1]}, {'seeds': [-1]}, {'seeds': [10001]}, {'protocols': ['fiction']}, {'batch_size': True}, {'learning_rate': float('nan')}, {'protocols': ['wave']}, {'retention_delay': 2}, {'task_family': 'unknown'}, {'task_family': []}, {'task_family': 'waypoint', 'horizon': 5}, ]) def test_invalid_config_rejected(kwargs): with pytest.raises(ValueError): StudyConfig(**kwargs)