| 1 | from __future__ import annotations
|
| 2 |
|
| 3 | import json
|
| 4 | import subprocess
|
| 5 | import sys
|
| 6 | import tempfile
|
| 7 | import unittest
|
| 8 | from pathlib import Path
|
| 9 |
|
| 10 | sys.path.insert(0, str(Path(__file__).resolve().parents[1] / "scripts"))
|
| 11 |
|
| 12 | from oakbench.baseline_book import load_baseline_book, load_targets
|
| 13 | from oakbench.scorecard import evaluate_targets
|
| 14 |
|
| 15 |
|
| 16 | def row(subject: str, elapsed: float, returncode: int = 0, **extra) -> dict:
|
| 17 | base = {
|
| 18 | "bench_id": "b",
|
| 19 | "profile": "smoke",
|
| 20 | "scenario": "tiny_text",
|
| 21 | "operation": "status.clean",
|
| 22 | "run": extra.pop("run", 0),
|
| 23 | "subject": subject,
|
| 24 | "subject_kind": "oak" if subject.startswith("oak") else "git",
|
| 25 | "elapsed_ms": elapsed,
|
| 26 | "returncode": returncode,
|
| 27 | "command": [subject, "status"],
|
| 28 | }
|
| 29 | base.update(extra)
|
| 30 | return base
|
| 31 |
|
| 32 |
|
| 33 | def baseline(value: float = 10.0, noise_floor_pct: float | None = None) -> dict:
|
| 34 | cell = {
|
| 35 | "lane": "core",
|
| 36 | "profile": "smoke",
|
| 37 | "scenario": "tiny_text",
|
| 38 | "operation": "status.clean",
|
| 39 | "subject": "git",
|
| 40 | "metric": "elapsed_ms",
|
| 41 | "statistic": "median",
|
| 42 | "value": value,
|
| 43 | "n": 50,
|
| 44 | }
|
| 45 | if noise_floor_pct is not None:
|
| 46 | cell["noise_floor_pct"] = noise_floor_pct
|
| 47 | return {"schema_version": 1, "cells": [cell]}
|
| 48 |
|
| 49 |
|
| 50 | def baseline_with_ci(value: float, low: float, high: float) -> dict:
|
| 51 | book = baseline(value)
|
| 52 | book["cells"][0]["ci_low"] = low
|
| 53 | book["cells"][0]["ci_high"] = high
|
| 54 | return book
|
| 55 |
|
| 56 |
|
| 57 | def targets(goal: float = 0.8, guards: dict | None = None) -> dict:
|
| 58 | return {
|
| 59 | "schema_version": 1,
|
| 60 | "targets": [
|
| 61 | {
|
| 62 | "id": "t",
|
| 63 | "lane": "core",
|
| 64 | "profile": "smoke",
|
| 65 | "scenario": "tiny_text",
|
| 66 | "operation": "status.clean",
|
| 67 | "subject": "oak_local",
|
| 68 | "comparator": "git",
|
| 69 | "metric": "elapsed_ms",
|
| 70 | "goal": {"type": "ratio_at_most", "value": goal},
|
| 71 | "guards": guards or {"success_rate_at_least": 1.0},
|
| 72 | }
|
| 73 | ],
|
| 74 | }
|
| 75 |
|
| 76 |
|
| 77 | class ScorecardTest(unittest.TestCase):
|
| 78 | def test_failed_trial_identity_is_checked_even_when_failure_guard_allows_it(self):
|
| 79 | result = self.eval([row("oak_local", 5, runner_class="host-a"),
|
| 80 | row("oak_local", 50, returncode=1, runner_class="host-b")],
|
| 81 | targets(0.8, guards={"success_rate_at_least": 0.5}))[0]
|
| 82 | self.assertEqual(result["verdict"], "INVALID")
|
| 83 | self.assertTrue(any("runner_class" in message for message in result["messages"]))
|
| 84 |
|
| 85 | def test_green_when_ratio_ci_clears_goal(self) -> None:
|
| 86 | result = self.eval([row("oak_local", 5.0), row("oak_local", 6.0)], targets(0.8))[0]
|
| 87 | self.assertEqual(result["verdict"], "GREEN")
|
| 88 | self.assertLessEqual(result["ratio_ci_high"], 0.8)
|
| 89 |
|
| 90 | def test_red_when_ratio_ci_misses_goal(self) -> None:
|
| 91 | result = self.eval([row("oak_local", 9.0), row("oak_local", 10.0)], targets(0.8))[0]
|
| 92 | self.assertEqual(result["verdict"], "RED")
|
| 93 |
|
| 94 | def test_unmeasured_when_rows_absent_or_skipped(self) -> None:
|
| 95 | result = self.eval([], targets(0.8))[0]
|
| 96 | self.assertEqual(result["verdict"], "UNMEASURED")
|
| 97 | skipped = row("oak_local", 0.0, returncode=77, skipped=True, skip_reason="missing remote")
|
| 98 | result = self.eval([skipped], targets(0.8))[0]
|
| 99 | self.assertEqual(result["verdict"], "UNMEASURED")
|
| 100 | self.assertIn("missing remote", result["messages"][0])
|
| 101 |
|
| 102 | def test_invalid_when_guard_fails(self) -> None:
|
| 103 | failed = row("oak_local", 5.0, returncode=1)
|
| 104 | result = self.eval([failed], targets(0.8))[0]
|
| 105 | self.assertEqual(result["verdict"], "INVALID")
|
| 106 |
|
| 107 | def test_undecidable_when_goal_tighter_than_noise_floor(self) -> None:
|
| 108 | result = self.eval(
|
| 109 | [row("oak_local", 9.8), row("oak_local", 10.0)],
|
| 110 | targets(0.99),
|
| 111 | baseline_book=baseline(noise_floor_pct=5.0),
|
| 112 | )[0]
|
| 113 | self.assertEqual(result["verdict"], "UNDECIDABLE")
|
| 114 |
|
| 115 | def test_obvious_green_survives_positive_noise_floor(self) -> None:
|
| 116 | result = self.eval(
|
| 117 | [row("oak_local", 5.0), row("oak_local", 5.0)],
|
| 118 | targets(1.0),
|
| 119 | baseline_book=baseline(noise_floor_pct=5.0),
|
| 120 | )[0]
|
| 121 | self.assertEqual(result["verdict"], "GREEN")
|
| 122 |
|
| 123 | def test_lucky_narrow_ci_inside_noise_floor_is_not_green(self) -> None:
|
| 124 | # Ratio 0.97 with a zero-width bootstrap CI clears a 0.99 goal by less
|
| 125 | # than the 5% noise floor: instrument luck, not a verdict (ADR-0008).
|
| 126 | result = self.eval(
|
| 127 | [row("oak_local", 9.7), row("oak_local", 9.7)],
|
| 128 | targets(0.99),
|
| 129 | baseline_book=baseline(noise_floor_pct=5.0),
|
| 130 | )[0]
|
| 131 | self.assertEqual(result["verdict"], "UNDECIDABLE")
|
| 132 |
|
| 133 | def test_goal_above_parity_measures_margin_from_parity(self) -> None:
|
| 134 | # A 1.5x ceiling has a 50% margin; a 5% noise floor must not make a
|
| 135 | # straddling result UNDECIDABLE — it is AMBER with a required n.
|
| 136 | result = self.eval(
|
| 137 | [row("oak_local", 14.0), row("oak_local", 16.0)],
|
| 138 | targets(1.5),
|
| 139 | baseline_book=baseline(noise_floor_pct=5.0),
|
| 140 | )[0]
|
| 141 | self.assertEqual(result["verdict"], "AMBER")
|
| 142 | self.assertIsNotNone(result.get("required_n"))
|
| 143 |
|
| 144 | def test_null_returncode_rows_are_failures_not_crashes(self) -> None:
|
| 145 | broken = row("oak_local", 5.0)
|
| 146 | broken["returncode"] = None
|
| 147 | result = self.eval([broken], targets(0.8))[0]
|
| 148 | self.assertEqual(result["verdict"], "INVALID")
|
| 149 |
|
| 150 | def test_rows_must_match_lane_and_profile_exactly(self) -> None:
|
| 151 | workflow_row = row("oak_local", 5.0, phase="workflow")
|
| 152 | result = self.eval([workflow_row], targets(0.8))[0]
|
| 153 | self.assertEqual(result["verdict"], "UNMEASURED")
|
| 154 |
|
| 155 | null_profile_target = targets(0.8)
|
| 156 | null_profile_target["targets"][0]["profile"] = None
|
| 157 | result = self.eval([row("oak_local", 5.0)], null_profile_target)[0]
|
| 158 | self.assertEqual(result["verdict"], "UNMEASURED")
|
| 159 |
|
| 160 | def test_missing_guard_metric_is_invalid(self) -> None:
|
| 161 | result = self.eval(
|
| 162 | [row("oak_local", 5.0, information_recall=1.0), row("oak_local", 5.0)],
|
| 163 | targets(0.8, guards={"success_rate_at_least": 1.0, "information_recall_at_least": 1.0}),
|
| 164 | )[0]
|
| 165 | self.assertEqual(result["verdict"], "INVALID")
|
| 166 | self.assertIn("information_recall guard is unmeasured", "; ".join(result["messages"]))
|
| 167 |
|
| 168 | result = self.eval(
|
| 169 | [row("oak_local", 5.0)],
|
| 170 | targets(0.8, guards={"success_rate_at_least": 1.0, "pipe_compatible_unified": True}),
|
| 171 | )[0]
|
| 172 | self.assertEqual(result["verdict"], "INVALID")
|
| 173 |
|
| 174 | def test_weak_content_integrity_source_is_measured_without_public_trust_guard(self) -> None:
|
| 175 | measured_only = row(
|
| 176 | "oak_local",
|
| 177 | 5.0,
|
| 178 | content_integrity_check_passed=True,
|
| 179 | content_integrity_source="worktree_payload_sha256",
|
| 180 | content_integrity_sources=["worktree_payload_sha256"],
|
| 181 | content_integrity_public_trust_passed=False,
|
| 182 | )
|
| 183 | result = self.eval([measured_only], targets(0.8))[0]
|
| 184 | self.assertEqual(result["verdict"], "GREEN")
|
| 185 |
|
| 186 | def test_public_content_trust_guard_rejects_weak_sources(self) -> None:
|
| 187 | weak_sources = [
|
| 188 | ("mixed_payload_sources", ["git_head_payload_sha256", "git_lfs_worktree_payload_sha256"]),
|
| 189 | ("worktree_payload_sha256", ["worktree_payload_sha256"]),
|
| 190 | ("git_lfs_worktree_payload_sha256", ["git_lfs_worktree_payload_sha256"]),
|
| 191 | ]
|
| 192 | for source, sources in weak_sources:
|
| 193 | with self.subTest(source=source):
|
| 194 | measured_only = row(
|
| 195 | "oak_local",
|
| 196 | 5.0,
|
| 197 | content_integrity_check_passed=True,
|
| 198 | content_integrity_source=source,
|
| 199 | content_integrity_sources=sources,
|
| 200 | content_integrity_public_trust_passed=False,
|
| 201 | )
|
| 202 | result = self.eval(
|
| 203 | [measured_only],
|
| 204 | targets(0.8, guards={"success_rate_at_least": 1.0, "content_integrity_public_trust": True}),
|
| 205 | )[0]
|
| 206 | self.assertEqual(result["verdict"], "INVALID")
|
| 207 | self.assertIn("weak content_integrity_source", "; ".join(result["messages"]))
|
| 208 |
|
| 209 | def test_public_content_trust_guard_accepts_strong_source(self) -> None:
|
| 210 | strong = row(
|
| 211 | "oak_local",
|
| 212 | 5.0,
|
| 213 | content_integrity_check_passed=True,
|
| 214 | content_integrity_source="git_head_payload_sha256",
|
| 215 | content_integrity_sources=["git_head_payload_sha256"],
|
| 216 | content_integrity_public_trust_passed=True,
|
| 217 | )
|
| 218 | result = self.eval(
|
| 219 | [strong],
|
| 220 | targets(0.8, guards={"success_rate_at_least": 1.0, "content_integrity_public_trust": True}),
|
| 221 | )[0]
|
| 222 | self.assertEqual(result["verdict"], "GREEN")
|
| 223 |
|
| 224 | def test_partial_metric_coverage_caps_green_to_amber(self) -> None:
|
| 225 | # 5 successful rows but the metric is measured on only 1; a decisive
|
| 226 | # under-goal value from that lone row must not be GREEN (ADR-0008
|
| 227 | # partial-coverage rule), and n must report contributing rows only.
|
| 228 | rows = [row("oak_local", 5.0, run=0)] + [row("oak_local", None, run=i) for i in range(1, 5)]
|
| 229 | result = self.eval(rows, targets(0.8))[0]
|
| 230 | self.assertEqual(result["verdict"], "AMBER")
|
| 231 | self.assertEqual(result["n"], 1)
|
| 232 | self.assertEqual(result["n_successful"], 5)
|
| 233 | self.assertIn("metric unmeasured on 4 of 5 successful rows", result["messages"])
|
| 234 |
|
| 235 | def test_full_metric_coverage_reports_matching_counts_and_green(self) -> None:
|
| 236 | result = self.eval([row("oak_local", 5.0), row("oak_local", 6.0)], targets(0.8))[0]
|
| 237 | self.assertEqual(result["verdict"], "GREEN")
|
| 238 | self.assertEqual(result["n"], 2)
|
| 239 | self.assertEqual(result["n_successful"], 2)
|
| 240 |
|
| 241 | def test_partial_metric_coverage_caps_red_to_amber(self) -> None:
|
| 242 | # Symmetric cap: missing data cannot fail a target either.
|
| 243 | rows = [row("oak_local", 15.0, run=0)] + [row("oak_local", None, run=i) for i in range(1, 5)]
|
| 244 | result = self.eval(rows, targets(0.8))[0]
|
| 245 | self.assertEqual(result["verdict"], "AMBER")
|
| 246 | self.assertEqual(result["n"], 1)
|
| 247 | self.assertEqual(result["n_successful"], 5)
|
| 248 | self.assertIn("metric unmeasured on 4 of 5 successful rows", result["messages"])
|
| 249 |
|
| 250 | def test_baseline_uncertainty_makes_ratio_ci_conservative(self) -> None:
|
| 251 | result = self.eval(
|
| 252 | [row("oak_local", 8.0), row("oak_local", 8.0)],
|
| 253 | targets(0.8),
|
| 254 | baseline_book=baseline_with_ci(10.0, 9.0, 11.0),
|
| 255 | )[0]
|
| 256 | self.assertEqual(result["verdict"], "AMBER")
|
| 257 | self.assertGreater(result["ratio_ci_high"], 0.8)
|
| 258 |
|
| 259 | def test_cli_json_output(self) -> None:
|
| 260 | with tempfile.TemporaryDirectory() as tmp:
|
| 261 | root = Path(tmp)
|
| 262 | rows = root / "rows.jsonl"
|
| 263 | rows.write_text(json.dumps(row("oak_local", 5.0)) + "\n")
|
| 264 | book = root / "book.json"
|
| 265 | book.write_text(json.dumps(baseline()))
|
| 266 | target_file = root / "targets.json"
|
| 267 | target_file.write_text(json.dumps(targets(0.8)))
|
| 268 | proc = subprocess.run(
|
| 269 | [
|
| 270 | sys.executable,
|
| 271 | "scripts/scorecard.py",
|
| 272 | "--mode", "diagnostic",
|
| 273 | "--baseline-book",
|
| 274 | str(book),
|
| 275 | "--targets",
|
| 276 | str(target_file),
|
| 277 | "--bootstrap-samples",
|
| 278 | "20",
|
| 279 | "--json",
|
| 280 | str(rows),
|
| 281 | ],
|
| 282 | cwd=Path(__file__).resolve().parents[1],
|
| 283 | text=True,
|
| 284 | stdout=subprocess.PIPE,
|
| 285 | stderr=subprocess.PIPE,
|
| 286 | check=False,
|
| 287 | )
|
| 288 | self.assertEqual(proc.returncode, 0, proc.stderr)
|
| 289 | parsed = json.loads(proc.stdout)
|
| 290 | self.assertEqual(parsed[0]["verdict"], "GREEN")
|
| 291 |
|
| 292 | def eval(self, rows: list[dict], target_config: dict, baseline_book: dict | None = None) -> list[dict]:
|
| 293 | with tempfile.TemporaryDirectory() as tmp:
|
| 294 | root = Path(tmp)
|
| 295 | target_path = root / "targets.json"
|
| 296 | target_path.write_text(json.dumps(target_config))
|
| 297 | book_path = root / "book.json"
|
| 298 | book_path.write_text(json.dumps(baseline_book or baseline()))
|
| 299 | return evaluate_targets(
|
| 300 | rows,
|
| 301 | load_targets(target_path),
|
| 302 | load_baseline_book(book_path),
|
| 303 | bootstrap_samples=50,
|
| 304 | )
|
| 305 |
|
| 306 |
|
| 307 | if __name__ == "__main__":
|
| 308 | unittest.main()
|