File size: 14,014 Bytes
d5b51d5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
283
284
285
286
287
288
289
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
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
#!/usr/bin/env python3
"""Generate synthetic LoopNet seed corpus (default: 500 records, >=40% failures)."""

from __future__ import annotations

import argparse
import json
import random
import uuid
from datetime import UTC, datetime, timedelta
from pathlib import Path

from loopnet.constants import (
    FAILURE_MODES,
    PATTERN_SLUGS,
)
from loopnet.les import les_from_trajectory, trajectory_diagnostics

ROOT = Path(__file__).resolve().parents[1]
DEFAULT_OUTPUT = ROOT / "data" / "seed" / "records.jsonl"
DEFAULT_SPLITS = ROOT / "data" / "seed" / "splits.json"

OBJECTIVES = [
    "Produce a research brief with verified citations and coverage score >= 0.85.",
    "Repair failing unit tests while preserving public API contracts.",
    "Synthesize multi-source findings into an executive summary under 500 words.",
    "Debate two solution approaches and converge on a ranked recommendation.",
    "Optimize prompt templates until rubric score exceeds 0.80 within budget.",
    "Plan and execute a data pipeline migration with zero schema regressions.",
    "Generate code patches that pass lint, type-check, and integration tests.",
    "Summarize customer feedback themes with actionable product insights.",
]

FAILURE_PROFILES: dict[str, dict] = {
    "fail.open_loop": {"shape": "flat", "termination": "budget_exhausted"},
    "fail.self_grade": {"shape": "inflate", "termination": "goal_met"},
    "fail.evaluator_drift": {"shape": "drift", "termination": "stall"},
    "fail.tau_omission": {"shape": "slow_climb", "termination": "max_iterations"},
    "fail.false_pass": {"shape": "spike_fake", "termination": "goal_met"},
    "fail.false_fail": {"shape": "reject_good", "termination": "stall"},
    "fail.oscillation": {"shape": "oscillate", "termination": "stall"},
    "fail.resource_bleed": {"shape": "costly_flat", "termination": "cost_exceeded"},
    "fail.state_corruption": {"shape": "collapse", "termination": "error"},
    "fail.orchestration_deadlock": {"shape": "stuck", "termination": "timeout"},
    "fail.meta_instability": {"shape": "decay", "termination": "error"},
    "fail.safety_bypass": {"shape": "unsafe_spike", "termination": "safety_violation"},
}

PATTERN_POOLS: dict[str, list[str]] = {
    "fail.open_loop": ["reflection-loop", "research-loop"],
    "fail.self_grade": ["reflection-loop", "critique-loop"],
    "fail.evaluator_drift": ["critique-loop", "verification-loop"],
    "fail.tau_omission": ["planning-loop", "exploration-loop"],
    "fail.false_pass": ["verification-loop", "optimization-loop"],
    "fail.false_fail": ["critique-loop", "debate-loop"],
    "fail.oscillation": ["optimization-loop", "planning-loop"],
    "fail.resource_bleed": ["exploration-loop", "multi-agent-coordination"],
    "fail.state_corruption": ["memory-augmented-loop", "multi-agent-coordination"],
    "fail.orchestration_deadlock": ["multi-agent-coordination", "debate-loop"],
    "fail.meta_instability": ["recursive-improvement-loop", "optimization-loop"],
    "fail.safety_bypass": ["safety-constrained-loop", "simulation-loop"],
}


def _uuid_record_id() -> str:
    return f"ln-{uuid.uuid4()}"


def _goal_trace(
    rng: random.Random,
    *,
    shape: str,
    iterations: int,
    goal_target: float,
    success: bool,
) -> list[float]:
    start = rng.uniform(0.15, 0.35)
    scores = [round(start, 4)]

    for step in range(1, iterations):
        prev = scores[-1]
        if shape == "flat":
            delta = rng.uniform(-0.01, 0.02)
        elif shape == "inflate":
            delta = rng.uniform(0.04, 0.09)
        elif shape == "drift":
            delta = rng.uniform(-0.03, 0.05) if step > iterations // 2 else rng.uniform(0.02, 0.06)
        elif shape == "slow_climb":
            delta = rng.uniform(0.01, 0.03)
        elif shape == "spike_fake":
            delta = rng.uniform(0.08, 0.15) if step == iterations - 1 else rng.uniform(-0.01, 0.03)
        elif shape == "reject_good":
            delta = rng.uniform(-0.04, 0.02)
        elif shape == "oscillate":
            delta = 0.07 if step % 2 == 0 else -0.06
        elif shape == "costly_flat":
            delta = rng.uniform(-0.005, 0.01)
        elif shape == "collapse":
            delta = rng.uniform(-0.12, -0.04) if step > 1 else rng.uniform(0.01, 0.04)
        elif shape == "stuck":
            delta = rng.uniform(-0.005, 0.005)
        elif shape == "decay":
            delta = rng.uniform(-0.08, -0.02)
        elif shape == "unsafe_spike":
            delta = rng.uniform(0.05, 0.12)
        else:
            delta = rng.uniform(0.02, 0.08) if success else rng.uniform(-0.03, 0.02)

        scores.append(round(max(0.0, min(1.0, prev + delta)), 4))

    if success and scores[-1] < goal_target:
        scores[-1] = round(rng.uniform(goal_target, min(1.0, goal_target + 0.08)), 4)

    return scores


def _build_trajectory(
    rng: random.Random,
    *,
    shape: str,
    iterations: int,
    goal_target: float,
    success: bool,
    failure_mode: str | None,
) -> list[dict]:
    goal_scores = _goal_trace(
        rng,
        shape=shape,
        iterations=iterations,
        goal_target=goal_target,
        success=success,
    )
    trajectory = []
    base_latency = rng.uniform(8.0, 25.0)

    for index, goal_score in enumerate(goal_scores, start=1):
        latency = round(base_latency * rng.uniform(0.8, 1.4), 3)
        cost = round(rng.uniform(0.03, 0.18), 4)
        if shape == "costly_flat":
            cost = round(rng.uniform(0.12, 0.28), 4)

        failure_codes: list[str] = []
        safety_events = 0
        human_intervention = False

        if failure_mode and index >= max(2, iterations - 2):
            failure_codes.append(failure_mode)
        if failure_mode == "fail.safety_bypass" and index >= iterations - 1:
            safety_events = rng.randint(1, 3)
        if failure_mode == "fail.orchestration_deadlock" and index >= iterations - 1:
            latency = round(base_latency * 6, 3)
        if rng.random() < 0.08:
            human_intervention = True

        trajectory.append(
            {
                "iteration": index,
                "goal_score": goal_score,
                "primary_quality": round(
                    max(0.0, min(1.0, goal_score + rng.uniform(-0.05, 0.05))), 4
                ),
                "cost_usd": cost,
                "latency_seconds": latency,
                "tokens": rng.randint(800, 6000),
                "failure_codes": failure_codes,
                "safety_events": safety_events,
                "human_intervention": human_intervention,
            }
        )

    return trajectory


def _loop_spec_snapshot(
    rng: random.Random,
    *,
    loop_name: str,
    patterns: list[str],
    max_iterations: int,
) -> dict:
    worker_count = rng.randint(1, 4)
    evaluator_count = rng.randint(1, 3)
    return {
        "loop_name": loop_name,
        "version": "1.0.0",
        "workers": [{"id": f"worker-{i}"} for i in range(1, worker_count + 1)],
        "evaluators": [{"id": f"evaluator-{i}"} for i in range(1, evaluator_count + 1)],
        "optimization_strategy": {"type": "prompt_refinement", "max_steps": max_iterations},
        "extensions": {"patterns": patterns},
    }


def _make_record(
    rng: random.Random,
    *,
    split: str,
    force_failure: bool | None = None,
    created_at: datetime,
) -> dict:
    goal_target = round(rng.choice([0.75, 0.80, 0.85, 0.90]), 2)
    max_iterations = rng.randint(4, 12)

    if force_failure is True:
        outcome = "failure"
    elif force_failure is False:
        outcome = "success" if rng.random() < 0.85 else "partial"
    else:
        outcome = rng.choices(["failure", "success", "partial"], weights=[0.42, 0.48, 0.10])[0]

    failure_mode: str | None = None
    failure_modes: list[str] = []
    shape = "climb"
    termination_reason = "goal_met"

    if outcome == "failure":
        failure_mode = rng.choice(FAILURE_MODES)
        failure_modes = [failure_mode]
        if rng.random() < 0.25:
            secondary = rng.choice([mode for mode in FAILURE_MODES if mode != failure_mode])
            failure_modes.append(secondary)
        profile = FAILURE_PROFILES[failure_mode]
        shape = profile["shape"]
        termination_reason = profile["termination"]
    elif outcome == "partial":
        failure_mode = None
        failure_modes = [rng.choice(FAILURE_MODES)]
        shape = "slow_climb"
        termination_reason = rng.choice(["stall", "budget_exhausted", "human_stop"])
    else:
        shape = "climb"
        termination_reason = "goal_met"

    iterations = (
        rng.randint(2, max_iterations - 1)
        if outcome == "failure"
        else rng.randint(3, max_iterations)
    )
    if termination_reason == "max_iterations":
        iterations = max_iterations

    patterns = PATTERN_POOLS.get(failure_mode or "", [])
    if not patterns:
        patterns = [rng.choice(PATTERN_SLUGS)]
        if rng.random() < 0.35:
            patterns.append(rng.choice([p for p in PATTERN_SLUGS if p not in patterns]))

    loop_name = f"{patterns[0].replace('-loop', '')}-{rng.randint(1000, 9999)}"
    objective = rng.choice(OBJECTIVES)

    trajectory = _build_trajectory(
        rng,
        shape=shape,
        iterations=iterations,
        goal_target=goal_target,
        success=outcome == "success",
        failure_mode=failure_mode,
    )

    diag = trajectory_diagnostics(trajectory)
    les_observed = les_from_trajectory(
        trajectory,
        goal_target=goal_target,
        outcome=outcome,
        failure_mode=failure_mode,
        max_iterations_budget=max_iterations,
    )

    record: dict = {
        "record_id": _uuid_record_id(),
        "schema_version": "ln/record-v1",
        "spec_pins": {"lss": "lss@1.0.0", "les": "les@1.0.0"},
        "created_at": created_at.isoformat().replace("+00:00", "Z"),
        "source": "synthetic",
        "split": split,
        "patterns": patterns,
        "loop_name": loop_name,
        "objective": objective,
        "loop_spec": _loop_spec_snapshot(
            rng,
            loop_name=loop_name,
            patterns=patterns,
            max_iterations=max_iterations,
        ),
        "outcome": outcome,
        "termination_reason": termination_reason,
        "trajectory": trajectory,
        "les_observed": les_observed,
        "metadata": {
            "iteration_count": diag["iteration_count"],
            "cost_total_usd": diag["cost_total_usd"],
            "goal_target": goal_target,
            "goal_final": diag["goal_final"],
            "worker_count": 0,
            "evaluator_count": 0,
            "max_iterations_budget": max_iterations,
            "regression_count": diag["regression_count"],
            "tags": ["synthetic", "v0.1"],
        },
        "redaction": {"level": "none", "fields_removed": []},
    }

    spec = record["loop_spec"]
    record["metadata"]["worker_count"] = len(spec["workers"])
    record["metadata"]["evaluator_count"] = len(spec["evaluators"])

    if failure_modes:
        record["failure_modes"] = failure_modes
    if failure_mode:
        record["failure_mode"] = failure_mode

    return record


def _assign_splits(count: int, rng: random.Random) -> list[str]:
    train_n = int(count * 0.8)
    val_n = int(count * 0.1)
    test_n = count - train_n - val_n
    splits = ["train"] * train_n + ["val"] * val_n + ["test"] * test_n
    rng.shuffle(splits)
    return splits


def generate_corpus(
    *,
    count: int = 500,
    seed: int = 42,
    failure_ratio: float = 0.42,
) -> list[dict]:
    rng = random.Random(seed)
    splits = _assign_splits(count, rng)
    failure_count = int(count * failure_ratio)
    failure_flags = [True] * failure_count + [False] * (count - failure_count)
    rng.shuffle(failure_flags)

    base_time = datetime(2026, 1, 1, tzinfo=UTC)
    records: list[dict] = []

    for index in range(count):
        created_at = base_time + timedelta(hours=index * 3)
        records.append(
            _make_record(
                rng,
                split=splits[index],
                force_failure=failure_flags[index],
                created_at=created_at,
            )
        )

    return records


def write_jsonl(records: list[dict], path: Path) -> None:
    path.parent.mkdir(parents=True, exist_ok=True)
    with path.open("w", encoding="utf-8") as handle:
        for record in records:
            handle.write(json.dumps(record, separators=(",", ":")) + "\n")


def write_splits(records: list[dict], path: Path) -> None:
    splits = {
        "train": [r["record_id"] for r in records if r["split"] == "train"],
        "val": [r["record_id"] for r in records if r["split"] == "val"],
        "test": [r["record_id"] for r in records if r["split"] == "test"],
    }
    path.parent.mkdir(parents=True, exist_ok=True)
    with path.open("w", encoding="utf-8") as handle:
        json.dump(splits, handle, indent=2)
        handle.write("\n")


def main(argv: list[str] | None = None) -> int:
    parser = argparse.ArgumentParser(description=__doc__)
    parser.add_argument("--count", type=int, default=500)
    parser.add_argument("--seed", type=int, default=42)
    parser.add_argument("--failure-ratio", type=float, default=0.42)
    parser.add_argument("--output", type=Path, default=DEFAULT_OUTPUT)
    parser.add_argument("--splits", type=Path, default=DEFAULT_SPLITS)
    args = parser.parse_args(argv)

    records = generate_corpus(
        count=args.count,
        seed=args.seed,
        failure_ratio=args.failure_ratio,
    )
    write_jsonl(records, args.output)
    write_splits(records, args.splits)

    failures = sum(1 for record in records if record["outcome"] == "failure")
    print(
        f"Wrote {len(records)} records to {args.output} "
        f"({failures} failures, {failures / len(records):.1%})"
    )
    print(f"Wrote split manifest to {args.splits}")
    return 0


if __name__ == "__main__":
    raise SystemExit(main())