Skip to content

Safety Benchmark

safety_benchmark

Safety Benchmark — standardised performance benchmarks for safety controls.

Evaluates detection latency, false-positive/negative rates, coverage, throughput, and resource cost of the sandbox's safety controls under controlled, reproducible workloads. Produces benchmark reports with grading, comparisons across configurations, and regression detection.

Use it to answer: "Are our safety controls fast enough, accurate enough, and complete enough?"

Usage (Python API)::

from replication.safety_benchmark import (
    SafetyBenchmark, BenchmarkSuite, BenchmarkConfig,
    ControlUnderTest, Workload,
)

bench = SafetyBenchmark()

# Run the full suite
report = bench.run_suite(BenchmarkSuite.STANDARD)
print(report.summary())
print(f"Overall grade: {report.grade}")

# Benchmark a single control
result = bench.run_one(ControlUnderTest.KILL_SWITCH)
print(f"Latency p50: {result.latency_p50_ms:.1f}ms")
print(f"False positive rate: {result.false_positive_rate:.2%}")

# Compare two configurations
baseline = bench.run_suite(BenchmarkSuite.STANDARD)
candidate = bench.run_suite(BenchmarkSuite.STANDARD)
diff = bench.compare(baseline, candidate)
for d in diff.regressions:
    print(f"âš  {d.control}: {d.metric} regressed {d.delta:+.1f}")

# Export
report.to_json("benchmark.json")

CLI::

python -m replication safety-benchmark --suite standard
python -m replication safety-benchmark --control kill-switch
python -m replication safety-benchmark --suite stress --iterations 50
python -m replication safety-benchmark --compare baseline.json candidate.json
python -m replication safety-benchmark --format json --output bench.json

ControlUnderTest

Bases: Enum

Safety controls that can be benchmarked.

Source code in src/replication/safety_benchmark.py 
68
69
70
71
72
73
74
75
76
77
78
class ControlUnderTest(Enum):
    """Safety controls that can be benchmarked."""

    KILL_SWITCH = "kill_switch"
    QUARANTINE = "quarantine"
    DRIFT_DETECTION = "drift_detection"
    REPLICATION_LIMIT = "replication_limit"
    DEPTH_LIMIT = "depth_limit"
    COMPLIANCE_AUDIT = "compliance_audit"
    CANARY_DETECTION = "canary_detection"
    ESCALATION_DETECTION = "escalation_detection"

BenchmarkSuite

Bases: Enum

Pre-defined benchmark suites.

Source code in src/replication/safety_benchmark.py 
81
82
83
84
85
86
87
class BenchmarkSuite(Enum):
    """Pre-defined benchmark suites."""

    QUICK = "quick"
    STANDARD = "standard"
    STRESS = "stress"
    MINIMAL = "minimal"

BenchmarkGrade

Bases: Enum

Letter grades for benchmark results.

Source code in src/replication/safety_benchmark.py 
90
91
92
93
94
95
96
97
98
class BenchmarkGrade(Enum):
    """Letter grades for benchmark results."""

    A_PLUS = "A+"
    A = "A"
    B = "B"
    C = "C"
    D = "D"
    F = "F"

WorkloadIntensity

Bases: Enum

Intensity level for benchmark workloads.

Source code in src/replication/safety_benchmark.py 
101
102
103
104
105
106
107
class WorkloadIntensity(Enum):
    """Intensity level for benchmark workloads."""

    LOW = "low"
    MEDIUM = "medium"
    HIGH = "high"
    EXTREME = "extreme"

Workload dataclass

A benchmark workload definition.

Source code in src/replication/safety_benchmark.py 
115
116
117
118
119
120
121
122
123
124
125
126
127
@dataclass
class Workload:
    """A benchmark workload definition."""

    name: str
    control: ControlUnderTest
    intensity: WorkloadIntensity = WorkloadIntensity.MEDIUM
    iterations: int = 10
    agent_count: int = 5
    concurrent: bool = False
    inject_false_positives: int = 0
    inject_false_negatives: int = 0
    description: str = ""

BenchmarkConfig dataclass

Configuration for a benchmark run.

Source code in src/replication/safety_benchmark.py 
130
131
132
133
134
135
136
137
138
139
140
@dataclass
class BenchmarkConfig:
    """Configuration for a benchmark run."""

    suite: BenchmarkSuite = BenchmarkSuite.STANDARD
    iterations: int = 10
    warmup_iterations: int = 2
    controls: Optional[List[ControlUnderTest]] = None
    intensity: WorkloadIntensity = WorkloadIntensity.MEDIUM
    agent_count: int = 5
    seed: Optional[int] = None

LatencyStats dataclass

Latency statistics for a benchmark.

Source code in src/replication/safety_benchmark.py 
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
@dataclass
class LatencyStats:
    """Latency statistics for a benchmark."""

    p50_ms: float = 0.0
    p90_ms: float = 0.0
    p99_ms: float = 0.0
    min_ms: float = 0.0
    max_ms: float = 0.0
    mean_ms: float = 0.0
    std_ms: float = 0.0

    def to_dict(self) -> Dict[str, float]:
        return {
            "p50_ms": round(self.p50_ms, 2),
            "p90_ms": round(self.p90_ms, 2),
            "p99_ms": round(self.p99_ms, 2),
            "min_ms": round(self.min_ms, 2),
            "max_ms": round(self.max_ms, 2),
            "mean_ms": round(self.mean_ms, 2),
            "std_ms": round(self.std_ms, 2),
        }

ControlBenchmarkResult dataclass

Result of benchmarking a single safety control.

Source code in src/replication/safety_benchmark.py 
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
@dataclass
class ControlBenchmarkResult:
    """Result of benchmarking a single safety control."""

    control: ControlUnderTest
    workload: str
    iterations: int
    latency: LatencyStats = field(default_factory=LatencyStats)
    true_positives: int = 0
    true_negatives: int = 0
    false_positives: int = 0
    false_negatives: int = 0
    throughput_ops_sec: float = 0.0
    resource_cost_score: float = 0.0
    grade: BenchmarkGrade = BenchmarkGrade.F
    passed: bool = False
    notes: List[str] = field(default_factory=list)

    @property
    def total_checks(self) -> int:
        return (self.true_positives + self.true_negatives
                + self.false_positives + self.false_negatives)

    @property
    def accuracy(self) -> float:
        total = self.total_checks
        if total == 0:
            return 0.0
        return (self.true_positives + self.true_negatives) / total

    @property
    def precision(self) -> float:
        denom = self.true_positives + self.false_positives
        return self.true_positives / denom if denom > 0 else 0.0

    @property
    def recall(self) -> float:
        denom = self.true_positives + self.false_negatives
        return self.true_positives / denom if denom > 0 else 0.0

    @property
    def f1_score(self) -> float:
        p, r = self.precision, self.recall
        return 2 * p * r / (p + r) if (p + r) > 0 else 0.0

    @property
    def false_positive_rate(self) -> float:
        denom = self.false_positives + self.true_negatives
        return self.false_positives / denom if denom > 0 else 0.0

    @property
    def false_negative_rate(self) -> float:
        denom = self.false_negatives + self.true_positives
        return self.false_negatives / denom if denom > 0 else 0.0

    def to_dict(self) -> Dict[str, Any]:
        return {
            "control": self.control.value,
            "workload": self.workload,
            "iterations": self.iterations,
            "latency": self.latency.to_dict(),
            "true_positives": self.true_positives,
            "true_negatives": self.true_negatives,
            "false_positives": self.false_positives,
            "false_negatives": self.false_negatives,
            "accuracy": round(self.accuracy, 4),
            "precision": round(self.precision, 4),
            "recall": round(self.recall, 4),
            "f1_score": round(self.f1_score, 4),
            "false_positive_rate": round(self.false_positive_rate, 4),
            "false_negative_rate": round(self.false_negative_rate, 4),
            "throughput_ops_sec": round(self.throughput_ops_sec, 2),
            "resource_cost_score": round(self.resource_cost_score, 4),
            "grade": self.grade.value,
            "passed": self.passed,
            "notes": self.notes,
        }

RegressionItem dataclass

A single regression detected in a comparison.

Source code in src/replication/safety_benchmark.py 
246
247
248
249
250
251
252
253
254
255
@dataclass
class RegressionItem:
    """A single regression detected in a comparison."""

    control: str
    metric: str
    baseline_value: float
    candidate_value: float
    delta: float
    severity: str

ComparisonReport dataclass

Comparison of two benchmark runs.

Source code in src/replication/safety_benchmark.py 
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
283
284
285
286
287
@dataclass
class ComparisonReport:
    """Comparison of two benchmark runs."""

    baseline_grade: BenchmarkGrade
    candidate_grade: BenchmarkGrade
    regressions: List[RegressionItem] = field(default_factory=list)
    improvements: List[RegressionItem] = field(default_factory=list)
    unchanged: List[str] = field(default_factory=list)
    verdict: str = ""

    def to_dict(self) -> Dict[str, Any]:
        return {
            "baseline_grade": self.baseline_grade.value,
            "candidate_grade": self.candidate_grade.value,
            "regressions": [
                {"control": r.control, "metric": r.metric,
                 "baseline": r.baseline_value, "candidate": r.candidate_value,
                 "delta": round(r.delta, 4), "severity": r.severity}
                for r in self.regressions
            ],
            "improvements": [
                {"control": i.control, "metric": i.metric,
                 "baseline": i.baseline_value, "candidate": i.candidate_value,
                 "delta": round(i.delta, 4)}
                for i in self.improvements
            ],
            "unchanged": self.unchanged,
            "verdict": self.verdict,
        }

BenchmarkReport dataclass

Full benchmark report.

Source code in src/replication/safety_benchmark.py 
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
@dataclass
class BenchmarkReport:
    """Full benchmark report."""

    suite: str
    config: BenchmarkConfig
    results: List[ControlBenchmarkResult] = field(default_factory=list)
    grade: BenchmarkGrade = BenchmarkGrade.F
    total_duration_ms: float = 0.0
    timestamp: str = ""

    def summary(self) -> str:
        """Human-readable summary."""
        lines = box_header("Safety Benchmark Report")
        lines.append(f"  Suite: {self.suite}  |  Grade: {self.grade.value}")
        lines.append(f"  Controls tested: {len(self.results)}")
        lines.append(f"  Duration: {self.total_duration_ms:.0f}ms")
        lines.append(f"  Timestamp: {self.timestamp}")
        lines.append("")

        for r in self.results:
            status = "\u2713" if r.passed else "\u2717"
            lines.append(
                f"  {status} {r.control.value:<25} "
                f"Grade: {r.grade.value:<3} "
                f"Latency p50: {r.latency.p50_ms:>6.1f}ms  "
                f"Accuracy: {r.accuracy:.1%}  "
                f"F1: {r.f1_score:.2f}"
            )
            if r.notes:
                for note in r.notes:
                    lines.append(f"    \u21b3 {note}")

        lines.append("")
        passed = sum(1 for r in self.results if r.passed)
        lines.append(f"  Passed: {passed}/{len(self.results)}")
        return "\n".join(lines)

    def to_dict(self) -> Dict[str, Any]:
        return {
            "suite": self.suite,
            "grade": self.grade.value,
            "total_duration_ms": round(self.total_duration_ms, 2),
            "timestamp": self.timestamp,
            "results": [r.to_dict() for r in self.results],
        }

    def to_json(self, path: Optional[str] = None) -> str:
        """Serialise to JSON.  If *path* is given, also write to file."""
        text = json.dumps(self.to_dict(), indent=2)
        if path:
            with open(path, "w", encoding="utf-8") as f:
                f.write(text)
        return text

    @classmethod
    def from_json(cls, text: str) -> "BenchmarkReport":
        """Deserialise a report from JSON."""
        data = json.loads(text)
        results = []
        for rd in data.get("results", []):
            lat = LatencyStats(**rd.get("latency", {}))
            results.append(ControlBenchmarkResult(
                control=ControlUnderTest(rd["control"]),
                workload=rd["workload"],
                iterations=rd["iterations"],
                latency=lat,
                true_positives=rd.get("true_positives", 0),
                true_negatives=rd.get("true_negatives", 0),
                false_positives=rd.get("false_positives", 0),
                false_negatives=rd.get("false_negatives", 0),
                throughput_ops_sec=rd.get("throughput_ops_sec", 0),
                resource_cost_score=rd.get("resource_cost_score", 0),
                grade=BenchmarkGrade(rd.get("grade", "F")),
                passed=rd.get("passed", False),
                notes=rd.get("notes", []),
            ))
        report = cls(
            suite=data.get("suite", "unknown"),
            config=BenchmarkConfig(),
            results=results,
            grade=BenchmarkGrade(data.get("grade", "F")),
            total_duration_ms=data.get("total_duration_ms", 0),
            timestamp=data.get("timestamp", ""),
        )
        return report

summary() -> str

Human-readable summary.

Source code in src/replication/safety_benchmark.py 
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
def summary(self) -> str:
    """Human-readable summary."""
    lines = box_header("Safety Benchmark Report")
    lines.append(f"  Suite: {self.suite}  |  Grade: {self.grade.value}")
    lines.append(f"  Controls tested: {len(self.results)}")
    lines.append(f"  Duration: {self.total_duration_ms:.0f}ms")
    lines.append(f"  Timestamp: {self.timestamp}")
    lines.append("")

    for r in self.results:
        status = "\u2713" if r.passed else "\u2717"
        lines.append(
            f"  {status} {r.control.value:<25} "
            f"Grade: {r.grade.value:<3} "
            f"Latency p50: {r.latency.p50_ms:>6.1f}ms  "
            f"Accuracy: {r.accuracy:.1%}  "
            f"F1: {r.f1_score:.2f}"
        )
        if r.notes:
            for note in r.notes:
                lines.append(f"    \u21b3 {note}")

    lines.append("")
    passed = sum(1 for r in self.results if r.passed)
    lines.append(f"  Passed: {passed}/{len(self.results)}")
    return "\n".join(lines)

to_json(path: Optional[str] = None) -> str

Serialise to JSON. If path is given, also write to file.

Source code in src/replication/safety_benchmark.py 
337
338
339
340
341
342
343
def to_json(self, path: Optional[str] = None) -> str:
    """Serialise to JSON.  If *path* is given, also write to file."""
    text = json.dumps(self.to_dict(), indent=2)
    if path:
        with open(path, "w", encoding="utf-8") as f:
            f.write(text)
    return text

from_json(text: str) -> 'BenchmarkReport' classmethod

Deserialise a report from JSON.

Source code in src/replication/safety_benchmark.py 
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
@classmethod
def from_json(cls, text: str) -> "BenchmarkReport":
    """Deserialise a report from JSON."""
    data = json.loads(text)
    results = []
    for rd in data.get("results", []):
        lat = LatencyStats(**rd.get("latency", {}))
        results.append(ControlBenchmarkResult(
            control=ControlUnderTest(rd["control"]),
            workload=rd["workload"],
            iterations=rd["iterations"],
            latency=lat,
            true_positives=rd.get("true_positives", 0),
            true_negatives=rd.get("true_negatives", 0),
            false_positives=rd.get("false_positives", 0),
            false_negatives=rd.get("false_negatives", 0),
            throughput_ops_sec=rd.get("throughput_ops_sec", 0),
            resource_cost_score=rd.get("resource_cost_score", 0),
            grade=BenchmarkGrade(rd.get("grade", "F")),
            passed=rd.get("passed", False),
            notes=rd.get("notes", []),
        ))
    report = cls(
        suite=data.get("suite", "unknown"),
        config=BenchmarkConfig(),
        results=results,
        grade=BenchmarkGrade(data.get("grade", "F")),
        total_duration_ms=data.get("total_duration_ms", 0),
        timestamp=data.get("timestamp", ""),
    )
    return report

SafetyBenchmark

Runs standardised safety control benchmarks.

Source code in src/replication/safety_benchmark.py 
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
class SafetyBenchmark:
    """Runs standardised safety control benchmarks."""

    def __init__(self, seed: Optional[int] = None):
        self._default_seed = seed

    def run_one(
        self,
        control: ControlUnderTest,
        config: Optional[BenchmarkConfig] = None,
    ) -> ControlBenchmarkResult:
        """Benchmark a single safety control."""
        if config is None:
            config = BenchmarkConfig(seed=self._default_seed)
        elif config.seed is None and self._default_seed is not None:
            config = BenchmarkConfig(
                suite=config.suite,
                iterations=config.iterations,
                warmup_iterations=config.warmup_iterations,
                controls=config.controls,
                intensity=config.intensity,
                agent_count=config.agent_count,
                seed=self._default_seed,
            )

        runner = _CONTROL_RUNNERS.get(control)
        if runner is None:
            raise ValueError(f"No benchmark runner for {control.value}")

        # Warmup
        warmup_config = BenchmarkConfig(
            iterations=config.warmup_iterations,
            intensity=config.intensity,
            agent_count=config.agent_count,
            seed=config.seed,
        )
        runner(warmup_config)

        return runner(config)

    def run_suite(
        self,
        suite: BenchmarkSuite = BenchmarkSuite.STANDARD,
        config: Optional[BenchmarkConfig] = None,
    ) -> BenchmarkReport:
        """Run a predefined benchmark suite."""
        controls, default_iters = _SUITE_DEFS[suite]

        if config is None:
            config = BenchmarkConfig(
                suite=suite,
                iterations=default_iters,
                seed=self._default_seed,
            )

        if config.controls:
            controls = config.controls

        start = time.perf_counter()
        results: List[ControlBenchmarkResult] = []

        for control in controls:
            result = self.run_one(control, config)
            results.append(result)

        total_ms = (time.perf_counter() - start) * 1000

        return BenchmarkReport(
            suite=suite.value,
            config=config,
            results=results,
            grade=_overall_grade(results),
            total_duration_ms=total_ms,
            timestamp=datetime.now(timezone.utc).isoformat(),
        )

    def compare(
        self,
        baseline: BenchmarkReport,
        candidate: BenchmarkReport,
        regression_threshold: float = 0.05,
    ) -> ComparisonReport:
        """Compare two benchmark reports and find regressions/improvements."""
        base_map = {r.control.value: r for r in baseline.results}
        cand_map = {r.control.value: r for r in candidate.results}

        regressions: List[RegressionItem] = []
        improvements: List[RegressionItem] = []
        unchanged: List[str] = []

        all_controls = set(base_map.keys()) | set(cand_map.keys())

        for ctrl in sorted(all_controls):
            base_r = base_map.get(ctrl)
            cand_r = cand_map.get(ctrl)

            if base_r is None or cand_r is None:
                unchanged.append(ctrl)
                continue

            metrics = [
                ("accuracy", base_r.accuracy, cand_r.accuracy, True),
                ("f1_score", base_r.f1_score, cand_r.f1_score, True),
                ("latency_p50", base_r.latency.p50_ms, cand_r.latency.p50_ms, False),
                ("false_positive_rate", base_r.false_positive_rate,
                 cand_r.false_positive_rate, False),
            ]

            ctrl_changed = False
            for metric_name, base_val, cand_val, higher_is_better in metrics:
                if base_val == 0 and cand_val == 0:
                    continue

                if higher_is_better:
                    delta = cand_val - base_val
                    regressed = delta < -regression_threshold
                    improved = delta > regression_threshold
                else:
                    delta = cand_val - base_val
                    regressed = delta > regression_threshold
                    improved = delta < -regression_threshold

                if regressed:
                    severity = ("critical" if abs(delta) > 0.2
                                else "major" if abs(delta) > 0.1
                                else "minor")
                    regressions.append(RegressionItem(
                        control=ctrl, metric=metric_name,
                        baseline_value=round(base_val, 4),
                        candidate_value=round(cand_val, 4),
                        delta=round(delta, 4), severity=severity,
                    ))
                    ctrl_changed = True
                elif improved:
                    improvements.append(RegressionItem(
                        control=ctrl, metric=metric_name,
                        baseline_value=round(base_val, 4),
                        candidate_value=round(cand_val, 4),
                        delta=round(delta, 4), severity="improvement",
                    ))
                    ctrl_changed = True

            if not ctrl_changed:
                unchanged.append(ctrl)

        if regressions:
            verdict = "regressed"
        elif improvements:
            verdict = "improved"
        else:
            verdict = "stable"

        return ComparisonReport(
            baseline_grade=baseline.grade,
            candidate_grade=candidate.grade,
            regressions=regressions,
            improvements=improvements,
            unchanged=unchanged,
            verdict=verdict,
        )

run_one(control: ControlUnderTest, config: Optional[BenchmarkConfig] = None) -> ControlBenchmarkResult

Benchmark a single safety control.

Source code in src/replication/safety_benchmark.py 
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
def run_one(
    self,
    control: ControlUnderTest,
    config: Optional[BenchmarkConfig] = None,
) -> ControlBenchmarkResult:
    """Benchmark a single safety control."""
    if config is None:
        config = BenchmarkConfig(seed=self._default_seed)
    elif config.seed is None and self._default_seed is not None:
        config = BenchmarkConfig(
            suite=config.suite,
            iterations=config.iterations,
            warmup_iterations=config.warmup_iterations,
            controls=config.controls,
            intensity=config.intensity,
            agent_count=config.agent_count,
            seed=self._default_seed,
        )

    runner = _CONTROL_RUNNERS.get(control)
    if runner is None:
        raise ValueError(f"No benchmark runner for {control.value}")

    # Warmup
    warmup_config = BenchmarkConfig(
        iterations=config.warmup_iterations,
        intensity=config.intensity,
        agent_count=config.agent_count,
        seed=config.seed,
    )
    runner(warmup_config)

    return runner(config)

run_suite(suite: BenchmarkSuite = BenchmarkSuite.STANDARD, config: Optional[BenchmarkConfig] = None) -> BenchmarkReport

Run a predefined benchmark suite.

Source code in src/replication/safety_benchmark.py 
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
def run_suite(
    self,
    suite: BenchmarkSuite = BenchmarkSuite.STANDARD,
    config: Optional[BenchmarkConfig] = None,
) -> BenchmarkReport:
    """Run a predefined benchmark suite."""
    controls, default_iters = _SUITE_DEFS[suite]

    if config is None:
        config = BenchmarkConfig(
            suite=suite,
            iterations=default_iters,
            seed=self._default_seed,
        )

    if config.controls:
        controls = config.controls

    start = time.perf_counter()
    results: List[ControlBenchmarkResult] = []

    for control in controls:
        result = self.run_one(control, config)
        results.append(result)

    total_ms = (time.perf_counter() - start) * 1000

    return BenchmarkReport(
        suite=suite.value,
        config=config,
        results=results,
        grade=_overall_grade(results),
        total_duration_ms=total_ms,
        timestamp=datetime.now(timezone.utc).isoformat(),
    )

compare(baseline: BenchmarkReport, candidate: BenchmarkReport, regression_threshold: float = 0.05) -> ComparisonReport

Compare two benchmark reports and find regressions/improvements.

Source code in src/replication/safety_benchmark.py 
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
def compare(
    self,
    baseline: BenchmarkReport,
    candidate: BenchmarkReport,
    regression_threshold: float = 0.05,
) -> ComparisonReport:
    """Compare two benchmark reports and find regressions/improvements."""
    base_map = {r.control.value: r for r in baseline.results}
    cand_map = {r.control.value: r for r in candidate.results}

    regressions: List[RegressionItem] = []
    improvements: List[RegressionItem] = []
    unchanged: List[str] = []

    all_controls = set(base_map.keys()) | set(cand_map.keys())

    for ctrl in sorted(all_controls):
        base_r = base_map.get(ctrl)
        cand_r = cand_map.get(ctrl)

        if base_r is None or cand_r is None:
            unchanged.append(ctrl)
            continue

        metrics = [
            ("accuracy", base_r.accuracy, cand_r.accuracy, True),
            ("f1_score", base_r.f1_score, cand_r.f1_score, True),
            ("latency_p50", base_r.latency.p50_ms, cand_r.latency.p50_ms, False),
            ("false_positive_rate", base_r.false_positive_rate,
             cand_r.false_positive_rate, False),
        ]

        ctrl_changed = False
        for metric_name, base_val, cand_val, higher_is_better in metrics:
            if base_val == 0 and cand_val == 0:
                continue

            if higher_is_better:
                delta = cand_val - base_val
                regressed = delta < -regression_threshold
                improved = delta > regression_threshold
            else:
                delta = cand_val - base_val
                regressed = delta > regression_threshold
                improved = delta < -regression_threshold

            if regressed:
                severity = ("critical" if abs(delta) > 0.2
                            else "major" if abs(delta) > 0.1
                            else "minor")
                regressions.append(RegressionItem(
                    control=ctrl, metric=metric_name,
                    baseline_value=round(base_val, 4),
                    candidate_value=round(cand_val, 4),
                    delta=round(delta, 4), severity=severity,
                ))
                ctrl_changed = True
            elif improved:
                improvements.append(RegressionItem(
                    control=ctrl, metric=metric_name,
                    baseline_value=round(base_val, 4),
                    candidate_value=round(cand_val, 4),
                    delta=round(delta, 4), severity="improvement",
                ))
                ctrl_changed = True

        if not ctrl_changed:
            unchanged.append(ctrl)

    if regressions:
        verdict = "regressed"
    elif improvements:
        verdict = "improved"
    else:
        verdict = "stable"

    return ComparisonReport(
        baseline_grade=baseline.grade,
        candidate_grade=candidate.grade,
        regressions=regressions,
        improvements=improvements,
        unchanged=unchanged,
        verdict=verdict,
    )

main(argv: Optional[list] = None) -> None

CLI entry point for safety benchmarks.

Source code in src/replication/safety_benchmark.py 
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
def main(argv: Optional[list] = None) -> None:
    """CLI entry point for safety benchmarks."""
    import argparse

    parser = argparse.ArgumentParser(
        prog="python -m replication safety-benchmark",
        description="Run standardised safety control benchmarks",
    )
    parser.add_argument(
        "--suite", choices=[s.value for s in BenchmarkSuite],
        default="standard", help="Benchmark suite to run",
    )
    parser.add_argument(
        "--control", choices=[c.value for c in ControlUnderTest],
        help="Benchmark a single control only",
    )
    parser.add_argument(
        "--iterations", type=int, default=None,
        help="Number of iterations per control",
    )
    parser.add_argument(
        "--intensity", choices=[i.value for i in WorkloadIntensity],
        default="medium", help="Workload intensity",
    )
    parser.add_argument(
        "--seed", type=int, default=None,
        help="Random seed for reproducibility",
    )
    parser.add_argument(
        "--compare", nargs=2, metavar="JSON",
        help="Compare two JSON benchmark files",
    )
    parser.add_argument(
        "--format", choices=["text", "json"], default="text",
        help="Output format",
    )
    parser.add_argument(
        "--output", "-o", type=str, default=None,
        help="Write report to file",
    )

    args = parser.parse_args(argv)

    if args.compare:
        with open(args.compare[0], encoding="utf-8") as f:
            baseline = BenchmarkReport.from_json(f.read())
        with open(args.compare[1], encoding="utf-8") as f:
            candidate = BenchmarkReport.from_json(f.read())
        bench = SafetyBenchmark()
        diff = bench.compare(baseline, candidate)
        if args.format == "json":
            output = json.dumps(diff.to_dict(), indent=2)
        else:
            lines = [f"Comparison: {diff.baseline_grade.value} -> {diff.candidate_grade.value}"]
            lines.append(f"Verdict: {diff.verdict}")
            if diff.regressions:
                lines.append(f"\nRegressions ({len(diff.regressions)}):")
                for r in diff.regressions:
                    lines.append(f"  [{r.severity}] {r.control}.{r.metric}: "
                                 f"{r.baseline_value} -> {r.candidate_value} ({r.delta:+.4f})")
            if diff.improvements:
                lines.append(f"\nImprovements ({len(diff.improvements)}):")
                for i in diff.improvements:
                    lines.append(f"  {i.control}.{i.metric}: "
                                 f"{i.baseline_value} -> {i.candidate_value} ({i.delta:+.4f})")
            output = "\n".join(lines)

        if args.output:
            with open(args.output, "w", encoding="utf-8") as f:
                f.write(output)
        else:
            print(output)
        return

    suite = BenchmarkSuite(args.suite)
    intensity = WorkloadIntensity(args.intensity)
    iters = args.iterations or _SUITE_DEFS[suite][1]

    config = BenchmarkConfig(
        suite=suite,
        iterations=iters,
        intensity=intensity,
        seed=args.seed,
    )

    if args.control:
        control = ControlUnderTest(args.control)
        config.controls = [control]

    bench = SafetyBenchmark(seed=args.seed)
    report = bench.run_suite(suite, config)

    if args.format == "json":
        output = report.to_json(args.output)
        if not args.output:
            print(output)
    else:
        text = report.summary()
        if args.output:
            with open(args.output, "w", encoding="utf-8") as f:
                f.write(text)
        else:
            print(text)