73 lines
4.8 KiB
JSON
73 lines
4.8 KiB
JSON
[
|
|
{
|
|
"id": "SCN-HH-001",
|
|
"category": "identity_memory",
|
|
"messages": ["ты кто?", "я назову тебя Дед, запомни", "какая погода в Сочи?"],
|
|
"must": ["identity uses stored assistant name", "rename is persisted", "follow-up keeps the same identity and task context"],
|
|
"must_not": ["model provider name presented as assistant identity", "memory confirmation replaces the next answer"]
|
|
},
|
|
{
|
|
"id": "SCN-HH-002",
|
|
"category": "current_information",
|
|
"messages": ["я сейчас в Москве", "какая погода в Питере?", "проверь цену на RTX 4090"],
|
|
"must": ["current facts use web research or an explicit source path", "location and time defaults are applied when confidence is high", "answer synthesizes evidence"],
|
|
"must_not": ["raw search snippets", "search-engine query URLs as the answer", "unsupported certainty"]
|
|
},
|
|
{
|
|
"id": "SCN-HH-003",
|
|
"category": "travel_and_places",
|
|
"messages": ["найди ближайший рейс до Сочи", "Москва", "сегодня", "найди ближайший ресторан"],
|
|
"must": ["missing high-impact origin/date is clarified once", "known context is reused", "results contain direct useful links or actionable next steps"],
|
|
"must_not": ["generic travel-site snippets presented as found tickets", "forgetting the destination or origin after clarification"]
|
|
},
|
|
{
|
|
"id": "SCN-HH-004",
|
|
"category": "multi_action_orchestration",
|
|
"messages": ["проверь погоду в Казани, найди кафе рядом и напомни вечером позвонить"],
|
|
"must": ["request is decomposed into independent bounded routes", "research and reminder agents receive separate profiles", "partial completion is reported per subtask"],
|
|
"must_not": ["one failed route suppresses unrelated work", "one monolithic agent owns all side effects"]
|
|
},
|
|
{
|
|
"id": "SCN-HH-005",
|
|
"category": "goals_reminders_tasks",
|
|
"messages": ["поставь цель подготовиться к поездке", "напомни завтра в 10 купить билет", "что осталось сделать?"],
|
|
"must": ["goal, task and reminder remain distinct", "commitment survives restart", "status is reported from durable state"],
|
|
"must_not": ["notification delivery alone treated as completion", "in-memory-only future obligation"]
|
|
},
|
|
{
|
|
"id": "SCN-HH-006",
|
|
"category": "speech_image_ocr",
|
|
"messages": ["послушай голосовое", "что на этой фотографии?", "прочитай текст на фото"],
|
|
"must": ["audio, vision and OCR are separate capability routes", "unsupported input has a useful recovery path", "confidence and unreadable regions are disclosed"],
|
|
"must_not": ["text-only model pretending to see or hear", "one unavailable modality blocks ordinary chat"]
|
|
},
|
|
{
|
|
"id": "SCN-HH-007",
|
|
"category": "creative_conversation",
|
|
"messages": ["нарисуй круг", "расскажи сказку о Колобке", "придумай шутку"],
|
|
"must": ["creative requests avoid irrelevant web search", "simple safe output can use a local fast path", "rich output uses cognition when available"],
|
|
"must_not": ["technical evidence refusal for a creative request", "provider name replacing Domovoy identity"]
|
|
},
|
|
{
|
|
"id": "SCN-HH-008",
|
|
"category": "degraded_and_conflict",
|
|
"messages": ["какая погода в Сочи?", "основной runtime недоступен", "источники противоречат друг другу"],
|
|
"must": ["unavailable, partial and conflict remain distinct", "local useful fallback is short and honest", "retry/recovery path is actionable"],
|
|
"must_not": ["generic loop responses", "invented current facts", "loss of conversation context"]
|
|
},
|
|
{
|
|
"id": "SCN-HH-009",
|
|
"category": "natural_dialogue",
|
|
"messages": ["ты кто?", "а теперь подробнее", "ладно, коротко", "дальше"],
|
|
"must": ["answers fit the requested depth without meta-commentary", "short turns contain the answer rather than boilerplate", "follow-up keeps subject and person model"],
|
|
"must_not": ["Я рядом boilerplate as every answer", "repeating capability descriptions instead of answering"]
|
|
},
|
|
{
|
|
"id": "SCN-HH-010",
|
|
"category": "authority_and_outcome",
|
|
"messages": ["купи билет", "отправь письмо", "измени проект", "не получилось ли действие?"],
|
|
"must": ["external mutation requires authority and confirmation where needed", "unknown outcome is not blindly retried", "proposal, execution and outcome are distinguishable"],
|
|
"must_not": ["model output directly executing a side effect", "duplicate purchase/message/code mutation"]
|
|
}
|
|
]
|