File size: 9,963 Bytes
80d8c84
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
from __future__ import annotations

import json

from replicalab.agents.lab_manager_agent import LabManagerAgent
from replicalab.env import ReplicaLabEnv
from replicalab.models import ScientistAction
from replicalab.oracle import Oracle
from replicalab.oracle_models import (
    AdjudicatorRoundScore,
    EnvironmentEvent,
    OracleLabManagerObservation,
    PostMortem,
    Scenario,
)


def _scenario_payload() -> dict:
    return {
        "paper": {
            "title": "Reproducing a Small Vision Benchmark",
            "domain": "ml_benchmark",
            "claim": "A compact model can recover >90% of reference accuracy under budget.",
            "method_summary": "Train a compact CNN with fixed augmentations and evaluate on a held-out split.",
            "original_sample_size": 1200,
            "original_duration_days": 3,
            "original_technique": "compact_cnn",
            "required_controls": ["seed_control", "baseline_model"],
            "required_equipment": ["GPU cluster", "validation server"],
            "required_reagents": ["dataset snapshot"],
            "statistical_test": "accuracy_gap",
        },
        "lab_constraints": {
            "budget_total": 2400.0,
            "budget_remaining": 2400.0,
            "equipment": [
                {
                    "name": "GPU cluster",
                    "available": True,
                    "condition": "shared_booking",
                    "booking_conflicts": ["Monday"],
                    "cost_per_use": 250.0,
                },
                {
                    "name": "Validation server",
                    "available": True,
                    "condition": "operational",
                    "booking_conflicts": [],
                    "cost_per_use": 20.0,
                },
            ],
            "reagents": [
                {
                    "name": "dataset snapshot",
                    "in_stock": True,
                    "quantity_available": 1.0,
                    "unit": "copy",
                    "lead_time_days": 0,
                    "cost": 0.0,
                }
            ],
            "staff": [
                {
                    "name": "Alex",
                    "role": "engineer",
                    "available_days": ["Monday", "Tuesday"],
                    "skills": ["training", "evaluation"],
                }
            ],
            "max_duration_days": 5,
            "safety_rules": ["No external internet during training."],
            "valid_substitutions": [
                {
                    "original": "GPU cluster",
                    "substitute": "single high-memory GPU",
                    "validity": "acceptable_with_caveats",
                    "caveats": "Lower throughput is acceptable if evaluation fidelity is preserved.",
                }
            ],
        },
        "minimum_viable_spec": {
            "min_sample_size": 800,
            "must_keep_controls": ["seed_control", "baseline_model"],
            "acceptable_techniques": ["compact_cnn", "distilled_cnn"],
            "min_duration_days": 2,
            "critical_equipment": ["Validation server"],
            "flexible_equipment": ["GPU cluster"],
            "critical_reagents": ["dataset snapshot"],
            "flexible_reagents": [],
            "power_threshold": 0.8,
        },
        "difficulty": "medium",
        "narrative_hook": "The compute team just reduced your preferred GPU window.",
    }


def _round_score_payload() -> dict:
    return {
        "rigor_flags": ["kept baseline_model"],
        "feasibility_flags": ["GPU window narrowed"],
        "info_gain": 0.6,
        "protocol_delta": 0.4,
        "momentum": 0.7,
        "contradiction_detected": False,
        "stalling_detected": False,
        "step_reward": 0.55,
        "notes": "Scientist asked a useful scheduling question and preserved controls.",
    }


def _post_mortem_payload() -> dict:
    return {
        "overall_summary": "The Scientist converged on a feasible compact CNN plan.",
        "rigor_explanation": "Controls and the validation server were preserved.",
        "feasibility_explanation": "The final plan fit the available compute and duration window.",
        "fidelity_explanation": "The protocol stayed close to the benchmark setup.",
        "key_decisions": ["Kept seed control", "Accepted lower-throughput compute"],
        "missed_opportunities": ["Could have asked about booking conflicts earlier"],
        "comparison_note": "An optimal Scientist would have requested the alternate GPU window one round sooner.",
    }


class _FakeMessagesAPI:
    def __init__(self, payloads: list[dict]) -> None:
        self._payloads = payloads
        self.calls = 0

    def create(self, **_: object):
        payload = self._payloads[self.calls]
        self.calls += 1

        class _Chunk:
            def __init__(self, text: str) -> None:
                self.text = text

        class _Response:
            def __init__(self, text: str) -> None:
                self.content = [_Chunk(text)]

        return _Response(json.dumps(payload))


class _FakeClient:
    def __init__(self, payloads: list[dict]) -> None:
        self.messages = _FakeMessagesAPI(payloads)


def test_oracle_generate_scenario_parses_json() -> None:
    oracle = Oracle(_FakeClient([_scenario_payload()]))

    scenario = oracle.generate_scenario(seed=7, difficulty="medium", domain="ml_benchmark")

    assert isinstance(scenario, Scenario)
    assert scenario.paper.domain == "ml_benchmark"
    assert scenario.lab_constraints.equipment[0].name == "GPU cluster"


def test_oracle_score_round_parses_structured_payload() -> None:
    oracle = Oracle(_FakeClient([_round_score_payload()]))
    scenario = Scenario.model_validate(_scenario_payload())
    action = ScientistAction(
        action_type="request_info",
        sample_size=0,
        controls=[],
        technique="",
        duration_days=0,
        required_equipment=[],
        required_reagents=[],
        questions=["When is the GPU cluster available?"],
        rationale="",
    )
    lab_manager = LabManagerAgent(_FakeClient([{
        "response_type": "feasibility_report",
        "feasible": False,
        "issues": ["GPU cluster is shared-booked on Monday"],
        "suggestions": ["Use the single high-memory GPU instead"],
        "cost_estimate": 250.0,
        "time_estimate_days": 3,
        "message": "The GPU cluster is shared-booked Monday; the single high-memory GPU is acceptable with caveats.",
    }]))
    response = lab_manager.respond(
        OracleLabManagerObservation(
            lab_constraints=scenario.lab_constraints,
            current_protocol=None,
            scientist_action=action,
            round_number=1,
        )
    )

    score = oracle.score_round(
        scenario=scenario,
        round_number=1,
        scientist_action=action,
        lab_manager_response=response,
        conversation_history=[],
        current_protocol=None,
        previous_scores=[],
    )

    assert isinstance(score, AdjudicatorRoundScore)
    assert score.step_reward == 0.55


def test_oracle_maybe_inject_event_returns_optional_event() -> None:
    oracle = Oracle(_FakeClient([{"inject": True, "event": {
        "event_type": "budget_cut",
        "description": "Finance reduced the remaining budget.",
        "state_changes": {"lab_constraints.budget_remaining": 1800.0},
        "severity": "moderate",
    }}]))

    event = oracle.maybe_inject_event(
        scenario=Scenario.model_validate(_scenario_payload()),
        round_number=3,
        current_protocol=None,
        conversation_history=[],
        inject_enabled=True,
    )

    assert isinstance(event, EnvironmentEvent)
    assert event.event_type == "budget_cut"


def test_oracle_generate_post_mortem_parses_json() -> None:
    oracle = Oracle(_FakeClient([_post_mortem_payload()]))
    from replicalab.oracle_models import AdjudicatorTerminalScore

    post_mortem = oracle.generate_post_mortem(
        scenario=Scenario.model_validate(_scenario_payload()),
        final_protocol={"technique": "compact_cnn"},
        conversation_history=[],
        terminal_score=AdjudicatorTerminalScore(
            rigor=0.9,
            feasibility=0.8,
            fidelity=0.85,
            parsimony=0.9,
            robustness=0.8,
            power_preservation=0.8,
            efficiency_bonus=0.2,
            communication_bonus=0.1,
            penalties={},
            terminal_reward=5.0,
            total_reward=5.6,
        ),
    )

    assert isinstance(post_mortem, PostMortem)
    assert "feasible compact CNN plan" in post_mortem.overall_summary


def test_env_can_reset_from_oracle_scenario_without_changing_outer_contract() -> None:
    class _FakeOracle:
        def __init__(self) -> None:
            self.scenario = Scenario.model_validate(_scenario_payload())

        def generate_scenario(self, seed: int, difficulty: str, domain: str) -> Scenario:
            assert seed == 11
            assert difficulty == "medium"
            assert domain == "ml_benchmark"
            return self.scenario

        def score_round(self, **_: object):
            return AdjudicatorRoundScore.model_validate(_round_score_payload())

        def maybe_inject_event(self, **_: object):
            return None

        def generate_post_mortem(self, **_: object):
            return PostMortem.model_validate(_post_mortem_payload())

    env = ReplicaLabEnv(
        oracle=_FakeOracle(),
        enable_oracle_post_mortem=True,
    )
    observation = env.reset(seed=11, scenario="ml_benchmark", difficulty="medium")

    assert observation.scientist is not None
    assert observation.scientist.paper_title == "Reproducing a Small Vision Benchmark"
    assert observation.lab_manager is not None
    assert "Validation server" in observation.lab_manager.equipment_available