228 lines
10 KiB
JSON
228 lines
10 KiB
JSON
{
|
||
"schema_version": "ai_agent_runtime_verifier_evidence_review_v1",
|
||
"generated_at": "2026-06-12T02:05:00+08:00",
|
||
"program_status": {
|
||
"overall_completion_percent": 99,
|
||
"current_priority": "P2",
|
||
"current_task_id": "P2-403I",
|
||
"next_task_id": "P2-403J",
|
||
"read_only_mode": true,
|
||
"runtime_authority": "runtime_verifier_evidence_review_only_no_live_execution",
|
||
"status_note": "P2-403I 已把 runtime verifier implementation 前必看的證據、redaction、rollback 與 failure receipt review 固定成只讀契約;尚未實作或執行 verifier、尚未讀 canonical target、尚未建立 rollback work item、尚未發 Telegram。"
|
||
},
|
||
"source_refs": [
|
||
"docs/evaluations/ai_agent_post_write_verifier_package_2026-06-12.json",
|
||
"docs/evaluations/ai_agent_interaction_learning_proof_2026-06-11.json",
|
||
"docs/superpowers/specs/2026-04-15-MASTER-ai-autonomous-flywheel-v2.md"
|
||
],
|
||
"review_truth": {
|
||
"review_packet_ready": true,
|
||
"runtime_verifier_implementation_allowed": false,
|
||
"post_write_verifier_execution_allowed": false,
|
||
"runtime_verifier_executed_count": 0,
|
||
"canonical_readback_allowed": false,
|
||
"canonical_readback_executed_count": 0,
|
||
"rollback_work_item_created_count": 0,
|
||
"telegram_failure_receipt_sent_count": 0,
|
||
"learning_writeback_after_verifier_count": 0,
|
||
"truth_note": "目前只建立 runtime verifier evidence implementation review;未通過 owner approval、redaction review 與 runtime gate 前,不能執行 verifier、不能讀 canonical target、不能寫 rollback work item、不能發 Telegram failure receipt。"
|
||
},
|
||
"review_package": {
|
||
"required_evidence": [
|
||
"approved_write_event_id",
|
||
"dry_run_preview_hash",
|
||
"target_write_surface",
|
||
"canonical_readback_query_plan",
|
||
"redacted_expected_diff",
|
||
"rollback_work_item_template",
|
||
"failure_receipt_template",
|
||
"redacted_evidence_refs"
|
||
],
|
||
"forbidden_evidence": [
|
||
"機密明文",
|
||
"授權標頭",
|
||
"原始工具輸出",
|
||
"私有推理內容",
|
||
"原始 Telegram 載荷",
|
||
"未脫敏 canonical 載荷"
|
||
],
|
||
"operator_meaning": "這份 review package 用來審查 runtime verifier implementation 是否可以進下一階段;它不是 verifier 實作,也不是 live write 授權。",
|
||
"approval_policy": "只有 OpenClaw owner 明確批准、Hermes redaction review 通過、NemoTron replay regression fixture 完成後,下一階段才能討論 runtime verifier implementation。",
|
||
"failure_policy": "若 readback plan、rollback template、failure receipt template 或 redaction 不完整,必須退回 P2-403H package 重修,不能啟動 runtime worker。"
|
||
},
|
||
"evidence_checks": [
|
||
{
|
||
"check_id": "canonical_readback_query_plan",
|
||
"display_name": "Canonical readback query plan",
|
||
"status": "approval_required",
|
||
"owner_agent": "hermes",
|
||
"required_evidence": "只收 query plan、target surface、expected hash 與 redacted evidence refs;不執行 live query。",
|
||
"review_question": "讀回計畫是否足以證明 post-write 結果,而且不會暴露 secret 或未脫敏 payload?",
|
||
"blocked_runtime_action": "canonical_readback_query_execution"
|
||
},
|
||
{
|
||
"check_id": "runtime_verifier_adapter_contract",
|
||
"display_name": "Runtime verifier adapter contract",
|
||
"status": "approval_required",
|
||
"owner_agent": "openclaw",
|
||
"required_evidence": "列出 adapter input、output、idempotency key、timeout、retry、rollback owner 與 no-action guard。",
|
||
"review_question": "Verifier adapter 是否只驗證已批准 write,且 no-action 不會被誤加分?",
|
||
"blocked_runtime_action": "post_write_verifier_runtime_execution"
|
||
},
|
||
{
|
||
"check_id": "rollback_work_item_template",
|
||
"display_name": "Rollback work item template",
|
||
"status": "contract_ready",
|
||
"owner_agent": "hermes",
|
||
"required_evidence": "模板需含 target、diff summary、rollback owner、deadline、failure reason 與 redacted refs。",
|
||
"review_question": "Verifier failed 時是否能產生可審查 rollback 草案,而不是靜默標記成功?",
|
||
"blocked_runtime_action": "rollback_work_item_write"
|
||
},
|
||
{
|
||
"check_id": "failure_receipt_draft_redaction",
|
||
"display_name": "Failure receipt redaction",
|
||
"status": "approval_required",
|
||
"owner_agent": "openclaw",
|
||
"required_evidence": "Telegram failure receipt 只允許摘要、風險、批准連結、evidence id 與下一步。",
|
||
"review_question": "失敗通知是否不含原始載荷、私有推理內容、授權標頭或機密明文?",
|
||
"blocked_runtime_action": "telegram_send_or_receipt_write"
|
||
},
|
||
{
|
||
"check_id": "nemotron_replay_regression_fixture",
|
||
"display_name": "NemoTron replay regression fixture",
|
||
"status": "blocked_by_runtime_gate",
|
||
"owner_agent": "nemotron",
|
||
"required_evidence": "使用 sanitized replay fixture 比較 baseline 與 candidate verifier outcome,不讀 live incident payload。",
|
||
"review_question": "Verifier implementation 是否會讓 replay score 或 failure classification 變差?",
|
||
"blocked_runtime_action": "agent_replay_score_write"
|
||
}
|
||
],
|
||
"implementation_review_lanes": [
|
||
{
|
||
"lane_id": "pre_implementation_owner_review",
|
||
"display_name": "Pre-implementation owner review",
|
||
"status": "approval_required",
|
||
"owner_agent": "openclaw",
|
||
"trigger": "approve_verifier_implementation requested",
|
||
"operator_instruction": "確認 evidence checks、rollback owner、redaction 與 no-action guard 完整後,才可批准下一階段 implementation。",
|
||
"blocked_runtime_action": "runtime_verifier_implementation"
|
||
},
|
||
{
|
||
"lane_id": "evidence_redaction_review",
|
||
"display_name": "Evidence redaction review",
|
||
"status": "contract_ready",
|
||
"owner_agent": "hermes",
|
||
"trigger": "new evidence refs attached",
|
||
"operator_instruction": "檢查 evidence refs 僅含脫敏 id / hash / summary,不含機密明文、原始工具輸出或未脫敏對話。",
|
||
"blocked_runtime_action": "canonical_readback_query_execution"
|
||
},
|
||
{
|
||
"lane_id": "no_action_result_guard",
|
||
"display_name": "No-action result guard",
|
||
"status": "contract_ready",
|
||
"owner_agent": "openclaw",
|
||
"trigger": "diagnostic-only or verifier skipped",
|
||
"operator_instruction": "未執行 write 或 verifier skipped 時,必須產生 no-action evidence,不得更新 trust 或 learning score。",
|
||
"blocked_runtime_action": "runtime_learning_write"
|
||
},
|
||
{
|
||
"lane_id": "regression_replay_review",
|
||
"display_name": "Regression replay review",
|
||
"status": "blocked_by_runtime_gate",
|
||
"owner_agent": "nemotron",
|
||
"trigger": "verifier adapter candidate changed",
|
||
"operator_instruction": "NemoTron 只能跑 sanitized replay fixture;不得接 production payload 或直接改模型路由。",
|
||
"blocked_runtime_action": "agent_replay_score_write"
|
||
}
|
||
],
|
||
"operator_actions": [
|
||
{
|
||
"action_id": "review_runtime_verifier_evidence_package",
|
||
"display_name": "審查 runtime verifier evidence package",
|
||
"action_type": "review",
|
||
"status": "ready_for_owner",
|
||
"owner_agent": "openclaw",
|
||
"operator_instruction": "檢查 readback、rollback、failure receipt、redaction 與 no-action guard 是否齊備。",
|
||
"blocked_runtime_action": "post_write_verifier_runtime_execution"
|
||
},
|
||
{
|
||
"action_id": "collect_missing_redacted_evidence",
|
||
"display_name": "補齊脫敏證據",
|
||
"action_type": "collect_evidence",
|
||
"status": "ready_for_owner",
|
||
"owner_agent": "hermes",
|
||
"operator_instruction": "補 evidence id、hash、summary 與 owner refs;不得補機密明文或原始載荷。",
|
||
"blocked_runtime_action": "canonical_readback_query_execution"
|
||
},
|
||
{
|
||
"action_id": "approve_runtime_verifier_implementation",
|
||
"display_name": "批准 runtime verifier implementation",
|
||
"action_type": "approve_implementation",
|
||
"status": "approval_required",
|
||
"owner_agent": "openclaw",
|
||
"operator_instruction": "只批准下一階段 implementation review 進入實作;不是批准 live execution 或 live write。",
|
||
"blocked_runtime_action": "runtime_verifier_implementation"
|
||
},
|
||
{
|
||
"action_id": "reject_runtime_verifier_implementation",
|
||
"display_name": "退回 runtime verifier implementation",
|
||
"action_type": "reject_or_rework",
|
||
"status": "ready_for_owner",
|
||
"owner_agent": "nemotron",
|
||
"operator_instruction": "若 replay regression、redaction 或 rollback lane 不完整,退回 P2-403H package 重修。",
|
||
"blocked_runtime_action": "runtime_learning_write"
|
||
}
|
||
],
|
||
"approval_boundaries": {
|
||
"runtime_verifier_implementation_allowed": false,
|
||
"post_write_verifier_execution_allowed": false,
|
||
"canonical_readback_allowed": false,
|
||
"rollback_work_item_write_allowed": false,
|
||
"telegram_failure_receipt_send_allowed": false,
|
||
"learning_writeback_allowed": false,
|
||
"km_write_allowed": false,
|
||
"playbook_trust_write_allowed": false,
|
||
"timeline_learning_write_allowed": false,
|
||
"agent_replay_score_write_allowed": false,
|
||
"runtime_worker_allowed": false,
|
||
"secret_plaintext_allowed": false
|
||
},
|
||
"display_redaction_contract": {
|
||
"redaction_required": true,
|
||
"raw_payload_display_allowed": false,
|
||
"private_reasoning_display_allowed": false,
|
||
"secret_value_display_allowed": false,
|
||
"allowed_frontend_content": [
|
||
"evidence check",
|
||
"implementation review lane",
|
||
"operator action",
|
||
"required evidence count",
|
||
"forbidden evidence count",
|
||
"blocked runtime action",
|
||
"approval policy",
|
||
"failure policy"
|
||
],
|
||
"forbidden_frontend_content": [
|
||
"機密明文",
|
||
"授權標頭",
|
||
"原始工具輸出",
|
||
"私有推理內容",
|
||
"原始 Telegram 載荷",
|
||
"未脫敏 canonical 載荷"
|
||
],
|
||
"frontend_display_policy": "治理頁只顯示 evidence check、implementation lane、operator action、blocked runtime action 與 policy;不顯示機密明文、授權標頭、原始工具輸出、私有推理內容、原始 Telegram 載荷或未脫敏 canonical 載荷。"
|
||
},
|
||
"rollups": {
|
||
"evidence_check_count": 5,
|
||
"implementation_review_lane_count": 4,
|
||
"operator_action_count": 4,
|
||
"approval_required_action_ids": [
|
||
"approve_runtime_verifier_implementation"
|
||
],
|
||
"blocked_runtime_action_count": 7,
|
||
"required_evidence_count": 8,
|
||
"forbidden_evidence_count": 6,
|
||
"live_verifier_execution_count": 0
|
||
}
|
||
}
|