Spaces:
Running
Running
File size: 16,775 Bytes
5741b22 ad552a7 5741b22 ad552a7 5741b22 ad552a7 5741b22 ad552a7 5741b22 | 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 | import importlib.util
import json
from pathlib import Path
import shutil
import tempfile
import tracemalloc
import unittest
spec = importlib.util.spec_from_file_location("study_analysis", Path(__file__).parents[1] / "analyze.py")
analysis = importlib.util.module_from_spec(spec)
spec.loader.exec_module(analysis)
class CurveTests(unittest.TestCase):
def setUp(self):
self.temp = tempfile.TemporaryDirectory()
self.root = Path(self.temp.name)
self.config = {"tasks": [{"instance_id": f"task-{i}"} for i in range(5)],
"runs": [{"run_id": f"task-{i}-{condition}", "task": f"task-{i}",
"condition": condition, "category": "benchmark"}
for condition in analysis.CONDITIONS for i in range(5)]}
def tearDown(self):
self.temp.cleanup()
def write_episode(self, task, condition, grades, stop_reason="completed"):
directory = self.root / "episodes" / f"task-{task}-{condition}"
(directory / "runner").mkdir(parents=True)
(directory / "grades.json").write_text(json.dumps(grades))
(directory / "runner" / "result.json").write_text(json.dumps({"stop_reason": stop_reason, "elapsed_seconds": 120}))
(directory / "runner" / "events.jsonl").write_text(json.dumps({"type": "episode_start", "sequence": 1,
"utc": "2026-10-09T00:00:00Z", "elapsed_ms": 0}) + "\n")
def row(self, rows, instant):
return next(row for row in rows if row["axis"] == "wall_seconds" and row["condition"] == "single" and row["time_seconds"] == instant)
def test_regressions_are_preserved_and_terminal_scores_carried(self):
self.write_episode(0, "single", [{"elapsed_seconds": 30, "score": 0.6, "valid": True},
{"elapsed_seconds": 60, "score": 0.2, "valid": True}])
summary, rows = analysis.analyze(self.root, self.config)
self.assertAlmostEqual(self.row(rows, 30)["mean_hidden_test_fraction"], 0.12)
self.assertAlmostEqual(self.row(rows, 60)["mean_hidden_test_fraction"], 0.04)
self.assertAlmostEqual(self.row(rows, 7200)["mean_hidden_test_fraction"], 0.04)
self.assertAlmostEqual(summary["final"]["single"]["mean_hidden_test_fraction"], 0.04)
self.assertEqual(summary["final"]["single"]["tasks_total"], 5)
def test_tasks_are_equal_weighted_not_pooled_by_test_count(self):
self.write_episode(0, "single", [{"elapsed_seconds": 30, "score": 1.0, "valid": True,
"passed": 1, "test_count": 1}])
self.write_episode(1, "single", [{"elapsed_seconds": 30, "score": 0.5, "valid": True,
"passed": 500, "test_count": 1000}])
summary, rows = analysis.analyze(self.root, self.config)
self.assertAlmostEqual(summary["final"]["single"]["mean_hidden_test_fraction"], 0.3)
self.assertEqual(self.row(rows, 30)["tasks_with_valid_grade"], 2)
def test_invalid_grading_retains_last_valid_or_zero(self):
self.write_episode(0, "single", [{"elapsed_seconds": 30, "score": 0.5, "valid": True},
{"elapsed_seconds": 60, "score": 0.9, "valid": False, "error": "grader_unavailable"}], "agent_error")
self.write_episode(1, "single", [{"elapsed_seconds": 30, "score": None, "valid": False}], "setup_error")
summary, _ = analysis.analyze(self.root, self.config)
self.assertAlmostEqual(summary["final"]["single"]["mean_hidden_test_fraction"], 0.1)
episode = next(row for row in summary["episodes"] if row["run_id"] == "task-0-single")
self.assertEqual(episode["status"], "failed")
self.assertEqual(episode["final_score"], 0.5)
self.assertEqual(episode["invalid_grade_count"], 1)
def test_explicit_valid_zero_replaces_previous_success(self):
self.write_episode(0, "single", [{"elapsed_seconds": 30, "score": 0.8, "valid": True},
{"elapsed_seconds": 60, "score": 0.0, "valid": True}])
summary, rows = analysis.analyze(self.root, self.config)
self.assertEqual(summary["final"]["single"]["mean_hidden_test_fraction"], 0)
self.assertEqual(self.row(rows, 60)["tasks_with_valid_grade"], 1)
def test_grade_summary_fallback_uses_official_aggregate_only(self):
self.write_episode(0, "single", [{"elapsed_seconds": 30, "sha256": "abc", "valid": True}])
directory = self.root / "episodes" / "task-0-single" / "evaluations" / "abc" / "task-0"
directory.mkdir(parents=True)
(directory / "summary.json").write_text(json.dumps({"complete": True, "score": 0.75, "test_count": 100}))
summary, _ = analysis.analyze(self.root, self.config)
self.assertEqual(summary["final"]["single"]["mean_hidden_test_fraction"], 0.15)
self.assertTrue((self.root / "curves.csv").exists())
self.assertTrue((self.root / "summary.json").exists())
def test_summary_fallback_respects_submission_and_infrastructure_classification(self):
fixtures = [{"complete": False, "valid": True, "analysis_score": 0.0,
"score": 0.0, "scoring_status": "submission_failed"},
{"complete": False, "valid": False, "analysis_score": None,
"score": 0.9, "scoring_status": "infrastructure_failed"}]
for task, summary in enumerate(fixtures):
self.write_episode(task, "single", [{"elapsed_seconds": 30, "sha256": "abc", "valid": True}])
directory = self.root / "episodes" / f"task-{task}-single" / "evaluations" / "abc" / f"task-{task}"
directory.mkdir(parents=True)
(directory / "summary.json").write_text(json.dumps(summary))
summary, _ = analysis.analyze(self.root, self.config)
episodes = {row["run_id"]: row for row in summary["episodes"]}
self.assertTrue(episodes["task-0-single"]["has_valid_grade"])
self.assertEqual(episodes["task-0-single"]["final_score"], 0.0)
self.assertFalse(episodes["task-1-single"]["has_valid_grade"])
self.assertEqual(episodes["task-1-single"]["invalid_grade_count"], 1)
def test_controller_grading_failure_overrides_completed_runner(self):
self.write_episode(0, "single", [{"elapsed_seconds": 30, "score": None, "valid": False}])
directory = self.root / "episodes" / "task-0-single"
(directory / "status.json").write_text(json.dumps({"status": "failed", "error": "grader failed"}))
summary, _ = analysis.analyze(self.root, self.config)
episode = next(row for row in summary["episodes"] if row["run_id"] == "task-0-single")
self.assertEqual(episode["status"], "failed")
self.assertEqual(episode["invalid_grade_count"], 1)
self.assertEqual(summary["final"]["single"]["episodes_completed"], 0)
def test_normal_wall_time_cutoff_is_a_completed_attempt(self):
self.write_episode(0, "single", [{"elapsed_seconds": 120, "score": 0.2, "valid": True}], "wall_time_limit")
summary, _ = analysis.analyze(self.root, self.config)
episode = next(row for row in summary["episodes"] if row["run_id"] == "task-0-single")
self.assertEqual(episode["status"], "completed")
self.assertEqual(episode["stop_reason"], "wall_time_limit")
directory = self.root / "episodes" / "task-0-single"
(directory / "status.json").write_text(json.dumps({"status": "interrupted", "stop_reason": "controller_restart"}))
summary, _ = analysis.analyze(self.root, self.config)
episode = next(row for row in summary["episodes"] if row["run_id"] == "task-0-single")
self.assertEqual(episode["status"], "interrupted")
self.assertEqual(episode["stop_reason"], "controller_restart")
def test_analysis_memory_does_not_retain_repeated_prompt_payloads(self):
self.write_episode(0, "single", [{"elapsed_seconds": 4, "score": 0.2, "valid": True}])
event_path = self.root / "episodes" / "task-0-single" / "runner" / "events.jsonl"
payload = "x" * (256 * 1024)
with event_path.open("a") as stream:
for index in range(64):
stream.write(json.dumps({"type": "model_request_start", "elapsed_ms": index,
"agent_id": "a", "context": {"messages": [payload]}}) + "\n")
stream.write(json.dumps({"type": "tool_start", "elapsed_ms": 1000,
"agent_id": "a", "sequence": 65, "command": payload}) + "\n")
stream.write(json.dumps({"type": "tool_end", "elapsed_ms": 3000,
"agent_id": "a", "causal_event_id": 65,
"duration_ms": 2000, "output": payload}) + "\n")
tracemalloc.start()
try:
episode = analysis.episode_data(self.root, self.config["runs"][0], [], 10_000, 265)
_, peak = tracemalloc.get_traced_memory()
finally:
tracemalloc.stop()
self.assertEqual(episode["points"][0]["reference_seconds"], 2)
self.assertLess(peak, 4 * 1024 * 1024, f"Retained prompt payloads: {peak / 1024**2:.1f} MiB")
def test_unicode_separators_in_transcript_are_not_jsonl_boundaries(self):
self.write_episode(0, "single", [{"elapsed_seconds": 4, "score": 0.2, "valid": True}])
event_path = self.root / "episodes" / "task-0-single" / "runner" / "events.jsonl"
events = [{"type": "tool_start", "elapsed_ms": 1000, "agent_id": "a", "sequence": 2},
{"type": "tool_end", "elapsed_ms": 3000, "agent_id": "a", "causal_event_id": 2,
"duration_ms": 2000, "output": "before\u0085middle\u2028after\u2029end"}]
with event_path.open("a") as stream:
for event in events:
stream.write(json.dumps(event, ensure_ascii=False) + "\n")
episode = analysis.episode_data(self.root, self.config["runs"][0], [], 10_000, 265)
self.assertEqual(episode["points"][0]["reference_seconds"], 2)
self.assertTrue(episode["clock_valid"])
def test_recovery_attempts_keep_spent_time_and_tokens_without_selecting_best_score(self):
self.write_episode(0, "single", [{"elapsed_seconds": 10, "score": 0.8, "valid": True},
{"elapsed_seconds": 120, "score": None, "valid": False},
{"elapsed_seconds": 7200, "score": 1.0, "valid": True}], "agent_error")
active = self.root / "episodes" / "task-0-single"
archive = self.root / "recovery-attempts" / "task-0-single" / "attempt-1"
archive.parent.mkdir(parents=True)
shutil.move(active, archive)
self.write_episode(0, "single", [{"elapsed_seconds": 120, "score": 0.2, "valid": True}])
(active / "runner" / "events.jsonl").write_text(json.dumps({"type": "episode_start", "sequence": 1,
"utc": "2026-10-09T01:00:00Z", "elapsed_ms": 0}) + "\n")
requests = [{"id": ident, "client_request_id": "reused-worker-id", "agent_id": "main",
"status": "complete", "prompt_tokens": 0, "completion_tokens": output,
"context_charged": output, "charged": output,
"started_at": f"2026-10-09T{hour:02d}:00:01Z",
"ended_at": f"2026-10-09T{hour:02d}:00:02Z"}
for ident, hour, output in [(1, 0, 265), (2, 1, 530)]]
episode = analysis.episode_data(self.root, self.config["runs"][0], requests, 10_000, 265)
self.assertEqual(episode["final_score"], 0.2)
self.assertEqual(episode["elapsed_seconds"], 240)
self.assertEqual(episode["api_tokens_charged"], 795)
self.assertEqual(episode["attempt_count"], 2)
self.assertTrue(episode["has_valid_final_grade"])
self.assertEqual(episode["invalid_grade_count"], 1)
self.assertEqual(episode["current_attempt_invalid_grade_count"], 0)
self.assertEqual([(point["wall_seconds"], point["reference_seconds"], point["score"])
for point in episode["points"]], [(10, 1, 0.8), (240, 3, 0.2)])
def test_invalid_terminal_grade_is_distinct_from_earlier_valid_zero(self):
self.write_episode(0, "single", [{"elapsed_seconds": 30, "score": 0, "valid": True},
{"elapsed_seconds": 120, "score": None, "valid": False}])
episode = analysis.episode_data(self.root, self.config["runs"][0], [], 10_000, 265)
self.assertTrue(episode["has_valid_grade"])
self.assertFalse(episode["has_valid_final_grade"])
class ClockTests(unittest.TestCase):
def event(self, type_, seconds, sequence, **kwargs):
return {"type": type_, "elapsed_ms": seconds * 1000, "sequence": sequence, **kwargs}
def request(self, ident, agent, prompt, context, output, started=0, ended=1):
return {"id": ident, "agent_id": agent, "client_request_id": ident, "status": "complete",
"prompt_tokens": prompt, "completion_tokens": output, "context_charged": context,
"started_at": f"2026-10-09T00:00:{started:02d}Z", "ended_at": f"2026-10-09T00:00:{ended:02d}Z"}
def test_parallel_clocks_message_handoff_and_child_inheritance(self):
events = [self.event("episode_start", 0, 1, utc="2026-10-09T00:00:00Z"),
self.event("agent_created", 0, 2, agent_id="a", parent_id=None),
self.event("agent_created", 0, 3, agent_id="b", parent_id=None),
self.event("tool_start", 1.1, 4, agent_id="a"),
self.event("tool_end", 4.1, 5, agent_id="a", causal_event_id=4, duration_ms=3000),
self.event("message_sent", 4.2, 6, agent_id="a"),
self.event("message_delivered", 4.3, 7, agent_id="b", causal_event_id=6),
self.event("tool_start", 4.4, 8, agent_id="b"),
self.event("tool_end", 6.4, 9, agent_id="b", causal_event_id=8, duration_ms=2000),
self.event("agent_created", 6.5, 10, agent_id="child", parent_id="b"),
self.event("tool_start", 6.6, 11, agent_id="child"),
self.event("tool_end", 7.6, 12, agent_id="child", causal_event_id=11, duration_ms=1000),
self.event("snapshot", 7.7, 13, path="snapshot.tar.gz")]
requests = [self.request("req-a", "a", 20_000, 10_265, 265),
self.request("req-b", "b", 0, 265, 265)]
clocks = analysis.clock_trace(events, requests)
self.assertTrue(clocks["valid"])
self.assertEqual(analysis.clocks_at(clocks, 1), (2, 3)) # max, not sum
self.assertEqual(analysis.clocks_at(clocks, 100, "snapshot.tar.gz"), (8, 9))
with tempfile.TemporaryDirectory() as directory:
path = Path(directory) / "events.jsonl"
with path.open("w") as stream:
for event in events:
stream.write(json.dumps({**event, "content": "omitted transcript\u0085text"},
ensure_ascii=False) + "\n")
projected = analysis.clock_trace(analysis.read_clock_events(path), requests)
self.assertEqual(projected, clocks)
def test_gateway_only_compaction_is_counted(self):
events = [self.event("episode_start", 0, 1, utc="2026-10-09T00:00:00Z")]
requests = [self.request("ordinary", "a", 0, 265, 265),
self.request("compaction", "a", 0, 530, 530, started=2, ended=3)]
clocks = analysis.clock_trace(events, requests)
self.assertTrue(clocks["valid"])
self.assertEqual(analysis.clocks_at(clocks, 4), (3, 3))
def test_unknown_usage_invalidates_normalized_curve(self):
events = [self.event("episode_start", 0, 1, utc="2026-10-09T00:00:00Z")]
request = self.request("unknown", "a", 0, 0, 0)
request.update(completion_tokens=None, prompt_tokens=None, status="unknown_usage_conservatively_charged")
clocks = analysis.clock_trace(events, [request])
self.assertFalse(clocks["valid"])
episode = {"condition": "single", "task": "a", "clock_valid": False,
"points": [{"wall_seconds": 30, "reference_seconds": 2, "score": 0.5, "order": 0}]}
row = analysis.aggregate_at([episode], ["a"], "single", 5, "reference_seconds")
self.assertIsNone(row["mean_hidden_test_fraction"])
self.assertEqual(analysis.aggregate_at([episode], ["a"], "single", 30, "wall_seconds")["mean_hidden_test_fraction"], 0.5)
if __name__ == "__main__":
unittest.main()
|