"""Original response-body controls for the is_silent_failure Noul contract.""" import argparse from collections import Counter import hashlib from html import escape import json from pathlib import Path import random from .api import compile_request from .data import SPLITS, _hash, _write_dataset, split_group VERSION = "silent-failure-control-v1" QUESTION = ( "Does this response body show that the current operation failed or did not deliver " "its stated required result? Judge only the body. Maintenance or a sign-in page " "replacing the requested result counts as failure, even when a field claims success. " "A valid empty result, explicitly accepted asynchronous work, allowed partial data, " "quoted error text, or an earlier incident that has recovered does not by itself count." ) SOURCE_COMMIT = "c490e0dc7830758f84bd9d5acb806655e327113e" SOURCE_URL = ( "https://github.com/Vicente-MD/jev-resilience/blob/" + SOURCE_COMMIT + "/src/main/java/ai/jev/resilience/client/JevEvaluationService.java#L50-L55" ) SPLIT_POLICY = "whole_provider_contract_family_and_counterfactual_pairs; hash_id_splits; reserved_zh_schema_layout_and_wording_ood" PAIRS = ( "html_maintenance", "html_login", "business_rejection", "required_receipt", "empty_search", "history_current", "quoted_error", "async_job", "partial_batch", "error_count", "freshness", "complete_segments", ) RULES = { "business_rejection": ("The reservation must be confirmed for this operation to succeed.", "只有预订状态为已确认，才算本次操作完成。"), "required_receipt": ("This operation requires a receipt with a nonempty receipt_id.", "本次操作必须返回回执，且 receipt_id 不能为空。"), "empty_search": ("A completed search may return no matches; an unfinished search is not a valid result.", "检索已完成时可以没有匹配项；未完成的检索不满足要求。"), "history_current": ("Only the event marked current describes this call; previous events are historical.", "仅标为本次的事件代表当前调用；先前事件只作历史记录。"), "quoted_error": ("Return the requested document. Text inside the document or example is quoted content, not the status of this fetch.", "应返回所请求的文档。文档及示例里的文字是引用内容，不表示本次读取的状态。"), "async_job": ("Acceptance into the queue with a job_id is a valid result; completed output is not required yet.", "进入队列并获得 job_id 即满足本次请求；暂不要求任务完成后的输出。"), "partial_batch": ("Return all requested items unless partial_allowed explicitly permits a smaller nonempty result.", "应返回全部请求项；只有 partial_allowed 明确允许时，才接受少量但非空的部分结果。"), "error_count": ("All required checks in the current checks object must pass. Archive counts are from a previous run.", "本次 checks 中的全部必需检查都要通过。archive 计数属于上次运行。"), "freshness": ("The delivered revision must be at least the minimum revision; a cache hit alone does not meet this requirement.", "返回修订号必须不低于最低修订号；命中缓存本身不代表满足要求。"), "complete_segments": ("Every requested segment must be delivered; archived failures belong to an earlier request.", "必须交付全部请求片段；历史失败数属于更早的请求。"), "document": ("Return the requested document in the Document field.", "请在文档字段内返回所请求的文档。"), } CURRENT = { "html_maintenance": (("Service available; maintenance has ended.", "Service offline for maintenance; no requested document was produced."), ("维护已经结束，服务可用。", "服务正在停机维护，尚未生成所请求的文档。")), "html_login": (("Requested document returned; the sign-in link is informational.", "Sign-in is required; this page replaces the requested document."), ("已返回所请求的文档；登录链接只是页面说明。", "需要登录；当前页面取代了所请求的文档。")), } SUBJECTS = ("orchid", "granite", "cedar", "amber", "willow", "coral", "maple", "quartz") def silent_failure_request(body: str) -> dict: """Preserve the actual body string; transport status and references stay outside.""" if not isinstance(body, str) or not body.strip(): raise ValueError("body must be a nonempty stringified response body") return {"state": body, "questions": {"is_silent_failure": {"type": "noul", "instructions": QUESTION}}} def _ordered(mapping, nonce): keys = list(mapping) random.Random(nonce).shuffle(keys) return {key: mapping[key] for key in keys} def _json_body(context, pair, failed, zh, nonce, count, layout): provider, request_id, action = context receipt_id = "r-" + _hash([request_id, "receipt"])[:12] job_id = "j-" + _hash([request_id, "job"])[:12] terms = RULES[pair][zh] if pair == "business_rejection": data = {"success": True, "reservation": {"status": (("confirmed", "rejected"), ("已确认", "已拒绝"))[zh][failed], "reason": (("capacity exhausted", "预约容量已用尽")[zh] if failed else None), "receipt_id": None if failed else receipt_id}} elif pair == "required_receipt": # The visible terms establish the requirement; {} alone would be ambiguous. shells = ({}, None, {"receipt_id": ""}, {"receipt_id": None}, {"receipt_id": " "}) data = {"success": True, "receipt": shells[int(nonce[:4], 16) % len(shells)] if failed else {"receipt_id": receipt_id}} elif pair == "empty_search": data = {"success": True, "search": {"matches": [], "complete": not failed, "error": (("index unavailable; search not executed", "索引不可用，检索尚未执行")[zh] if failed else None)}} elif pair == "history_current": status = (("completed", "failed"), ("已完成", "失败"))[zh] scopes = ("previous", "current") if not zh else ("先前", "本次") events = [{"scope": scopes[0], "status": status[not failed]}, {"scope": scopes[1], "status": status[failed]}] random.Random(nonce).shuffle(events) data = {"events": events} elif pair == "quoted_error": quote = ("Example only: 'ERROR: upstream unavailable'. This line explains a diagnostic message.", "仅为示例：『错误：上游不可用』。这段话是在解释诊断信息。")[zh] data = {"fetch": {"complete": not failed}, "document": None if failed else {"title": action, "text": quote}, "documentation_example": quote, "error": (("requested document could not be read", "无法读取所请求的文档")[zh] if failed else None)} elif pair == "async_job": data = {"success": True, "job": {"status": (("queued", "rejected"), ("已入队", "已拒绝"))[zh][failed], "job_id": None if failed else job_id}, "result": None} elif pair == "partial_batch": data = {"success": True, "policy": {"partial_allowed": not failed, "requested": count}, "items": ["i-" + _hash([request_id, i])[:10] for i in range(count - 1)]} elif pair == "error_count": errors = min(count, 1 + int(nonce[:4], 16) % 3) if failed else 0 data = {"success": True, "checks": {"required": count, "passed": count - errors, "failed": errors}, "archive": {"failed": 0 if failed else count}} else: raise ValueError("unsupported JSON pair") data = _ordered(data, nonce) if layout == "flat_en": body = {"provider": provider, "request_id": request_id, "operation": action, "terms": terms, "response": data} elif layout == "envelope_en": body = {"request": {"provider": provider, "id": request_id, "operation": action}, "contract": terms, "body": data} elif layout == "entries_zh": entries = [{"栏目": key, "内容": value} for key, value in (("约定", terms), ("返回", data))] random.Random(nonce).shuffle(entries) body = {"调用信息": {"服务": provider, "标识": request_id, "动作": action}, "条目": entries} elif layout == "nested_zh": body = {"服务": provider, "调用": {"标识": request_id, "动作": action, "交付": {"约定": terms, "返回": data}}} else: raise ValueError("unsupported JSON layout") return json.dumps(_ordered(body, nonce), ensure_ascii=False, indent=2 if int(nonce[0], 16) % 2 else None) def _document_body(context, pair, failed, zh, nonce): provider, request_id, action = context labels = ("Provider", "Request", "Operation", "Requirement", "Current", "Document") if not zh else ("服务", "调用", "动作", "约定", "现在", "文档") values = (provider, request_id, action, RULES["document"][zh], CURRENT[pair][zh][failed], "" if failed else action + " / " + request_id) fields = _ordered(dict(zip(labels, values)), nonce) if zh: content = "".join(f"
{key}
{escape(value)}
" for key, value in fields.items()) return '

调用回执

' + content + '
' content = "".join(f"

{key}: {escape(value)}

" for key, value in fields.items()) return '

Service response

' + content + '
' def _text_body(context, pair, failed, zh, nonce, count): provider, request_id, action = context if pair == "freshness": names = ("Minimum revision", "Delivered revision", "Cache") if not zh else ("最低修订号", "返回修订号", "缓存") values = (str(count), str(count - 1 if failed else count + int(nonce[:2], 16) % 2), ("hit", "命中")[zh]) else: names = ("Requested segments", "Delivered segments", "Archived failed segments") if not zh else ("请求片段数", "交付片段数", "历史失败数") values = (str(count), str(count - 1 if failed else count), str(count + 1)) base_names = ("Provider", "Request", "Operation", "Requirement", "Claim") if not zh else ("服务", "调用", "动作", "约定", "声明") fields = dict(zip(base_names, (provider, request_id, action, RULES[pair][zh], ("ready", "已就绪")[zh]))) fields.update(zip(names, values)) separator, heading = (" => ", "返回单") if zh else (": ", "Response record") return heading + "\n" + "\n".join(key + separator + value for key, value in _ordered(fields, nonce).items()) def generate(groups=400, seed=42, ood_groups=80): if type(groups) is not int or type(ood_groups) is not int or not 1 <= ood_groups < groups: raise ValueError("require integer groups > ood_groups >= 1") if type(seed) is not int: raise ValueError("seed must be an integer") rows = [] for index in range(groups): code = _hash([VERSION, seed, index])[:16] provider = f"service-{code}.example" group_id = "silent:" + code zh = index >= groups - ood_groups split = "ood" if zh else split_group(group_id, seed) subject = SUBJECTS[index % len(SUBJECTS)] + "-" + code[:6] for pair_index, pair in enumerate(PAIRS): nonce = _hash([code, pair_index])[:16] request_id = "call-" + nonce action = ("Requested resource " + subject) if not zh else ("请求资源：" + subject) context = (provider, request_id, action) count = 3 + int(nonce[2:6], 16) % 7 if pair.startswith("html_"): fmt, layout = "html", "definition_zh" if zh else "paragraph_en" elif pair in ("freshness", "complete_segments"): fmt, layout = "text", "receipt_zh" if zh else "record_en" else: fmt = "json" layout = ("entries_zh", "nested_zh")[index % 2] if zh else ("flat_en", "envelope_en")[index % 2] for failed in (False, True): if fmt == "html": body = _document_body(context, pair, failed, zh, nonce) elif fmt == "text": body = _text_body(context, pair, failed, zh, nonce, count) else: body = _json_body(context, pair, failed, zh, nonce, count, layout) compiled = compile_request(**silent_failure_request(body))[0] rows.append({"id": "sf-" + _hash([nonce, failed])[:24], "group_id": group_id, "split": split, "source": VERSION, **{key: compiled[key] for key in ("state", "question", "kind", "options")}, "target": [0.0, 1.0] if failed else [1.0, 0.0], "metadata": {"family": "evidence", "domain": VERSION, "question_id": "is_silent_failure", "template_id": f"{VERSION}:{layout}:{pair}", "pair": pair, "body_format": fmt, "layout": layout, "entity_ids": [provider, request_id], "provenance": {"type": "synthetic", "generator_version": VERSION, "seed": seed, "group_index": index, "variant": 2 * pair_index + int(failed), "license": "CC0-1.0", "split_policy": SPLIT_POLICY}}}) random.Random(seed).shuffle(rows) return rows def build_dataset(output_dir, groups=400, seed=42, ood_groups=80): output = Path(output_dir) if output.is_symlink() or (output.exists() and (not output.is_dir() or any(output.iterdir()))): raise ValueError("Choose a new empty silent-failure directory; existing corpora are never overwritten") root = Path(__file__).resolve().parents[1] sources = ("jev/silent_failure_data.py", "jev/api.py", "jev/data.py", "reports/silent-failure-control-v1/verify.py") config = {"type": "synthetic", "generator_version": VERSION, "groups": groups, "ood_groups": ood_groups, "seed": seed, "license": "CC0-1.0", "split_policy": SPLIT_POLICY, "source_contract": {"repository": "Vicente-MD/jev-resilience", "commit": SOURCE_COMMIT, "url": SOURCE_URL, "source_code_imported": False, "source_examples_imported": False, "question_independently_authored": True}, "runtime_contract": {"state": "stringified response body only", "question_id": "is_silent_failure", "type": "noul", "transport_status_is_model_input": False, "judge_transport_fallback_is_training_target": False}, "source_files_sha256": {name: hashlib.sha256((root / name).read_bytes()).hexdigest() for name in sources}} rows = generate(groups, seed, ood_groups) manifest = _write_dataset(rows, output, config) manifest.update(pair_count=groups * len(PAIRS), records_per_group=24, label_counts=dict(Counter("yes" if row["target"][1] else "no" for row in rows)), format_counts=dict(Counter(row["metadata"]["body_format"] for row in rows)), family_counts={split: len({row["group_id"] for row in rows if row["split"] == split}) for split in SPLITS}, training_performed=False, model_inference_performed=False, frozen_training_datasets_modified=False, scope="Controlled original response-body contracts, not arbitrary API correctness or a measured model capability.") (output / "manifest.json").write_text(json.dumps(manifest, ensure_ascii=False, indent=2) + "\n", encoding="utf-8") return manifest def main(): parser = argparse.ArgumentParser(description=__doc__) parser.add_argument("--output-dir", required=True, type=Path) parser.add_argument("--groups", type=int, default=400) parser.add_argument("--ood-groups", type=int, default=80) parser.add_argument("--seed", type=int, default=42) args = parser.parse_args() print(json.dumps(build_dataset(args.output_dir, args.groups, args.seed, args.ood_groups), indent=2)) if __name__ == "__main__": main()