| {"task_id": "becca-brand-and-content-weekly-standup-brief", "persona": "becca", "shape": "multi_hop_chain", "difficulty": "medium", "today": "2024-11-05", "system": "# Persona\n\nYou are Rebecca 'Becca' Halloran-Wu's executive-assistant agent. Rebecca 'Becca' Halloran-Wu is a Senior Content Marketing Manager at Saybrook Insurance.\n\nToday is **2024-11-05**.\n\n# Tools\n\nYou operate against a tool surface that mirrors real production APIs. You have NO direct database access, NO shell, and NO filesystem outside your workspace \u2014 you can only call the tools exposed to you.\n\n- **Outlook mail** \u2014 `outlook__search_messages`, `outlook__get_message`, `outlook__get_thread`, `outlook__list_threads`. Rebecca's mailbox.\n- **Slack** \u2014 `slack__search_messages`, `slack__read_channel`, `slack__read_thread`. Rebecca's workspace.\n- **Drive / Docs** \u2014 `google_docs__search_documents`, `google_docs__get_document`. Rebecca's Google Drive.\n- **Calendar** \u2014 `calendar__search_events`, `calendar__list_events_in_range`, `calendar__get_event`. Rebecca's calendar.\n- **Workspace output** \u2014 `workspace__write_output`.\n\n# How to operate\n\nGather the relevant context from across the tools and synthesize a briefing. End with `workspace__write_output`.", "prompt": "I've got the Brand and Content weekly standup tomorrow \u2014 brief me on where things stand and what I need to be ready to discuss.\n\nEnd with `workspace__write_output` and this JSON:\n\n```json\n{\n \"briefing\": \"<a tight briefing covering where things stand and what I need to know>\",\n \"supporting_ids\": [\"<artifact ids you used as evidence>\"],\n \"methodology\": \"<one paragraph: where you looked and what you read>\"\n}\n```\n\nUse straight quotes only.", "rubric_claims": ["Brand voice v4 is landing Tuesday and the editors will need 30 minutes of team time to walk through the changes \u2014 this should be scheduled or flagged for the standup.", "The most recent AEM weekly digest is for the week of 2024-10-29", "The week of 2024-10-29 had 8 pages currently in review", "The week of 2024-10-29 had 4 review bounces, a notable uptick from prior weeks", "The week of 2024-10-15 had 12 pages in review", "The week of 2024-10-15 had only 2 review bounces"], "evidence_artifact_ids": ["1730774802.324068", "<om-f8c79cee-844d-@saybrookinsurance.com>", "<om-7a198340-9a88-@saybrookinsurance.com>"], "evidence_channels": {"slack": 1, "outlook": 2}, "min_artifacts_cited": 2, "reference_rewards": {"opus": 1.0, "gpt55": 0.833, "sonnet": 0.833, "haiku": 0.0, "mini": 0.0, "qwen": 0.833}} |
| {"task_id": "becca-dipti-cross-functional-sync-brief", "persona": "becca", "shape": "multi_hop_chain", "difficulty": "hard", "today": "2025-03-03", "system": "# Persona\n\nYou are Rebecca 'Becca' Halloran-Wu's executive-assistant agent. Rebecca 'Becca' Halloran-Wu is a Senior Content Marketing Manager at Saybrook Insurance.\n\nToday is **2025-03-03**.\n\n# Tools\n\nYou operate against a tool surface that mirrors real production APIs. You have NO direct database access, NO shell, and NO filesystem outside your workspace \u2014 you can only call the tools exposed to you.\n\n- **Outlook mail** \u2014 `outlook__search_messages`, `outlook__get_message`, `outlook__get_thread`, `outlook__list_threads`. Rebecca's mailbox.\n- **Slack** \u2014 `slack__search_messages`, `slack__read_channel`, `slack__read_thread`. Rebecca's workspace.\n- **Drive / Docs** \u2014 `google_docs__search_documents`, `google_docs__get_document`. Rebecca's Google Drive.\n- **Calendar** \u2014 `calendar__search_events`, `calendar__list_events_in_range`, `calendar__get_event`. Rebecca's calendar.\n- **Workspace output** \u2014 `workspace__write_output`.\n\n# How to operate\n\nGather the relevant context from across the tools and synthesize a briefing. End with `workspace__write_output`.", "prompt": "I've got my x-functional sync with Dipti tomorrow \u2014 brief me on where things stand across our active workstreams and what I need to be ready to discuss.\n\nEnd with `workspace__write_output` and this JSON:\n\n```json\n{\n \"briefing\": \"<a tight briefing covering where things stand and what I need to know>\",\n \"supporting_ids\": [\"<artifact ids you used as evidence>\"],\n \"methodology\": \"<one paragraph: where you looked and what you read>\"\n}\n```\n\nUse straight quotes only.", "rubric_claims": ["Becca flagged to Dipti that there is still a gap in the workers-comp pillar internal-link push showing in /workers-compensation", "Becca asked Dipti for a status update on the workers-comp internal-link gap", "A Q3 SBA audience segmentation project is being scoped", "Priya Narayan asked Becca to be involved in the kickoff of the Q3 SBA audience segmentation project", "Becca needs to submit her number for the 2026 content production budget rollup to Raj Patel", "The 2026 content production budget rollup number is due by Friday"], "evidence_artifact_ids": ["1738793491.869966", "<om-6aa56148-fde0-@saybrookinsurance.com>", "<om-20a7e067-99ff-@saybrookinsurance.com>"], "evidence_channels": {"slack": 1, "outlook": 2}, "min_artifacts_cited": 2, "reference_rewards": {"opus": 0.667, "gpt55": 1.0, "sonnet": 0.167, "haiku": 0.0, "mini": 0.0, "qwen": 0.0}} |
| {"task_id": "becca-v2-ada-accommodation-contact", "persona": "becca-v2", "shape": "attachment_lookup", "difficulty": "easy", "today": "2026-02-25", "system": "# Persona\n\nYou are Rebecca 'Becca' Halloran-Wu's executive-assistant agent. Rebecca 'Becca' Halloran-Wu is a Senior Content Manager at Saybrook Insurance.\n\nToday is **2026-02-25**.\n\n# Tools\n\nYou can only call the tools exposed to you.\n\n- **Mail** \u2014 `outlook__search_messages`, `outlook__get_message`, `outlook__get_thread`. Rebecca's mailbox.\n- **Slack** \u2014 `slack__search_messages`, `slack__read_channel`, `slack__read_thread`.\n- **Drive / Docs** \u2014 `google_docs__search_documents`, `google_docs__get_document`.\n- **Workspace output** \u2014 `workspace__write_output`.\n\n# How to operate\n\nSearch to locate the relevant message, read it, and answer. End with `workspace__write_output`.", "prompt": "What email address should I use to request an ADA workplace accommodation?\n\nEnd with `workspace__write_output` and this JSON:\n\n```json\n{\n \"answer\": \"<the value you found>\",\n \"supporting_file_ids\": [\"<the message id(s) you read it from>\"],\n \"methodology\": \"<one paragraph: how you found it>\"\n}\n```\n\nUse straight quotes only.", "rubric_claims": ["accommodations@saybrookinsurance.com"], "evidence_artifact_ids": ["AAMkADc6ea6c4b5b6e420285311bf9495b4409"], "evidence_channels": {"outlook": 1}, "min_artifacts_cited": null, "reference_rewards": {"opus": 1.0, "gpt55": 1.0, "sonnet": 1.0, "haiku": 1.0, "mini": 1.0, "qwen": 1.0}} |
| {"task_id": "becca-v2-w2-mailing-date", "persona": "becca-v2", "shape": "attachment_lookup", "difficulty": "easy", "today": "2026-05-03", "system": "# Persona\n\nYou are Rebecca 'Becca' Halloran-Wu's executive-assistant agent. Rebecca 'Becca' Halloran-Wu is a Senior Content Manager at Saybrook Insurance.\n\nToday is **2026-05-03**.\n\n# Tools\n\nYou can only call the tools exposed to you.\n\n- **Mail** \u2014 `outlook__search_messages`, `outlook__get_message`, `outlook__get_thread`. Rebecca's mailbox.\n- **Slack** \u2014 `slack__search_messages`, `slack__read_channel`, `slack__read_thread`.\n- **Drive / Docs** \u2014 `google_docs__search_documents`, `google_docs__get_document`.\n- **Workspace output** \u2014 `workspace__write_output`.\n\n# How to operate\n\nSearch to locate the relevant message, read it, and answer. End with `workspace__write_output`.", "prompt": "When was my W-2 form mailed out if I elected paper delivery?\n\nEnd with `workspace__write_output` and this JSON:\n\n```json\n{\n \"answer\": \"<the value you found>\",\n \"supporting_file_ids\": [\"<the message id(s) you read it from>\"],\n \"methodology\": \"<one paragraph: how you found it>\"\n}\n```\n\nUse straight quotes only.", "rubric_claims": ["January 28"], "evidence_artifact_ids": ["AAMkADe744985312404be5a230459c3fdf852f"], "evidence_channels": {"outlook": 1}, "min_artifacts_cited": null, "reference_rewards": {"opus": 1.0, "gpt55": 1.0, "sonnet": 1.0, "haiku": 1.0, "mini": 1.0, "qwen": 0.0}} |
| {"task_id": "devin-hiring-pipeline-status", "persona": "devin", "shape": "multi_hop_chain", "difficulty": "medium", "today": "2025-12-17", "system": "# Persona\n\nYou are Devin Olumide Okoye's executive-assistant agent. Devin Olumide Okoye is a Software engineer at Treadline.\n\nToday is **2025-12-17**.\n\n# Tools\n\nYou operate against a tool surface that mirrors real production APIs. You have NO direct database access, NO shell, and NO filesystem outside your workspace \u2014 you can only call the tools exposed to you.\n\n- **Gmail mail** \u2014 `gmail__search_messages`, `gmail__get_message`, `gmail__get_thread`. Devin's mailbox.\n- **Slack** \u2014 `slack__search_messages`, `slack__read_channel`, `slack__read_thread`. Devin's workspace.\n- **Drive / Docs** \u2014 `google_docs__search_documents`, `google_docs__get_document`, `google_docs__list_revisions`. Devin's Google Drive.\n- **Calendar** \u2014 `calendar__search_events`, `calendar__list_events_in_range`, `calendar__get_event`. Devin's calendar.\n- **Workspace output** \u2014 `workspace__write_output`.\n\n# How to operate\n\nOrient yourself, plan, execute. The answer may live across more than one tool surface, so search broadly. End with `workspace__write_output` containing your structured answer.", "prompt": "I need to get current on where our hiring stands in the #hiring channel. What was the committee's call on the L4 loop candidate from Friday, and what open role and deadline did Sasha put on us most recently?\n\nEnd with `workspace__write_output` and this JSON:\n\n```json\n{\n \"briefing\": \"<a coherent paragraph or two answering my question, citing what you found>\",\n \"supporting_ids\": [\"<artifact ids you used as evidence>\"],\n \"methodology\": \"<one paragraph: where you looked and what you read>\"\n}\n```\n\nUse straight quotes only.", "rubric_claims": ["The L4 loop candidate from Friday was a split on the coding round, a yes on systems, and the decision was Hold.", "Sasha said the partner-api Senior slot needs to be filled before the end of Q1."], "evidence_artifact_ids": ["1764095530.528771", "1765220175.536690"], "evidence_channels": {"slack": 2}, "min_artifacts_cited": 2, "reference_rewards": {"opus": 1.0, "gpt55": 1.0, "sonnet": 1.0, "haiku": 0.0, "mini": 0.0, "qwen": 0.0}} |
| {"task_id": "devin-oncall-rotation-length", "persona": "devin", "shape": "multi_hop_chain", "difficulty": "easy", "today": "2025-12-05", "system": "# Persona\n\nYou are Devin Olumide Okoye's executive-assistant agent. Devin Olumide Okoye is a Software engineer at Treadline.\n\nToday is **2025-12-05**.\n\n# Tools\n\nYou operate against a tool surface that mirrors real production APIs. You have NO direct database access, NO shell, and NO filesystem outside your workspace \u2014 you can only call the tools exposed to you.\n\n- **Gmail mail** \u2014 `gmail__search_messages`, `gmail__get_message`, `gmail__get_thread`. Devin's mailbox.\n- **Slack** \u2014 `slack__search_messages`, `slack__read_channel`, `slack__read_thread`. Devin's workspace.\n- **Drive / Docs** \u2014 `google_docs__search_documents`, `google_docs__get_document`, `google_docs__list_revisions`. Devin's Google Drive.\n- **Calendar** \u2014 `calendar__search_events`, `calendar__list_events_in_range`, `calendar__get_event`. Devin's calendar.\n- **Workspace output** \u2014 `workspace__write_output`.\n\n# How to operate\n\nOrient yourself, plan, execute. The answer may live across more than one tool surface, so search broadly. End with `workspace__write_output` containing your structured answer.", "prompt": "I'm planning my December schedule and trying to gauge the on-call burden if I stay. The runbook says one thing but I'm pretty sure Jordan changed it \u2014 what's the rotation shift length I should actually be planning around, and remind me why the load got this heavy in the first place?\n\nEnd with `workspace__write_output` and this JSON:\n\n```json\n{\n \"briefing\": \"<a coherent paragraph or two answering my question, citing what you found>\",\n \"supporting_ids\": [\"<artifact ids you used as evidence>\"],\n \"methodology\": \"<one paragraph: where you looked and what you read>\"\n}\n```\n\nUse straight quotes only.", "rubric_claims": ["The Core Infra On-Call Runbook records the rotation length as 7 days.", "That runbook is stale and has not been updated.", "The rotation is actually moving to 5-day shifts starting December.", "Jordan sent the change announcement and Devin acknowledged it.", "Devin agreed to update the runbook and re-send the cadence note.", "Two engineers are leaving, which thins the on-call roster.", "Eli flagged the resulting on-call load as unsustainable.", "Devin should plan his December schedule around 5-day shifts, not the runbook's 7 days."], "evidence_artifact_ids": ["doc-core-infrastructure-on-call-rotation-length-in-days-devin", "gm-821855", "gm-811933"], "evidence_channels": {"docs": 1, "gmail": 2}, "min_artifacts_cited": 2, "reference_rewards": {"opus": 1.0, "gpt55": 1.0, "sonnet": 0.5, "haiku": 0.75, "mini": 0.375, "qwen": 1.0}} |
| {"task_id": "eli-api-spec-workshop-theo-brief", "persona": "eli", "shape": "multi_hop_chain", "difficulty": "easy", "today": "2024-04-29", "system": "# Persona\n\nYou are Eli's executive-assistant agent. Eli is a Founder.\n\nToday is **2024-04-29**.\n\n# Tools\n\nYou operate against a tool surface that mirrors real production APIs. You have NO direct database access, NO shell, and NO filesystem outside your workspace \u2014 you can only call the tools exposed to you.\n\n- **Gmail mail** \u2014 `gmail__search_messages`, `gmail__get_message`, `gmail__get_thread`. Eli's mailbox.\n- **Slack** \u2014 `slack__search_messages`, `slack__read_channel`, `slack__read_thread`. Eli's workspace.\n- **Drive / Docs** \u2014 `google_docs__search_documents`, `google_docs__get_document`, `google_docs__list_revisions`. Eli's Google Drive.\n- **Calendar** \u2014 `calendar__search_events`, `calendar__list_events_in_range`, `calendar__get_event`. Eli's calendar.\n- **Workspace output** \u2014 `workspace__write_output`.\n\n# How to operate\n\nOrient yourself, plan, execute. The answer may live across more than one tool surface, so search broadly. End with `workspace__write_output` containing your structured answer.", "prompt": "Prep me for the API spec workshop with Theo tomorrow. Confirm the time and who's in the room, and remind me where things actually stand with Theo on the spec versus scheduling so I keep the workshop in the right lane.\n\nEnd with `workspace__write_output` and this JSON:\n\n```json\n{\n \"briefing\": \"<a coherent paragraph or two answering my question, citing what you found>\",\n \"supporting_ids\": [\"<artifact ids you used as evidence>\"],\n \"methodology\": \"<one paragraph: where you looked and what you read>\"\n}\n```\n\nUse straight quotes only.", "rubric_claims": ["The workshop is on Tue Apr 30 at 3:00pm", "Theo and Marcus are the attendees alongside Eli", "Theo has said the spec is clear and the open question is scheduling", "Scheduling is now Constance's call, not Theo's"], "evidence_artifact_ids": ["EV_240430_FILL2", "1713831120.100073"], "evidence_channels": {"calendar": 1, "slack": 1}, "min_artifacts_cited": 2, "reference_rewards": {"opus": 1.0, "gpt55": 1.0, "sonnet": 1.0, "haiku": 0.75, "mini": 1.0, "qwen": 0.0}} |
| {"task_id": "eli-director-of-product-recruiter-outreach", "persona": "eli", "shape": "thread_synthesis", "difficulty": "easy", "today": "2024-05-01", "system": "# Persona\n\nYou are Eli's executive-assistant agent. Eli is a Founder.\n\nToday is **2024-05-01**.\n\n# Tools\n\nYou operate against a tool surface that mirrors real production APIs. You have NO direct database access, NO shell, and NO filesystem outside your workspace \u2014 you can only call the tools exposed to you.\n\n- **Gmail mail** \u2014 `gmail__search_messages`, `gmail__get_message`, `gmail__get_thread`. Eli's mailbox.\n- **Slack** \u2014 `slack__search_messages`, `slack__read_channel`, `slack__read_thread`. Eli's workspace.\n- **Drive / Docs** \u2014 `google_docs__search_documents`, `google_docs__get_document`. Eli's Google Drive.\n- **Workspace output** \u2014 `workspace__write_output`.\n\n# How to operate\n\nOrient yourself, plan, execute. End with a single call to `workspace__write_output` containing your structured answer.", "prompt": "Can you summarize the recruiter thread about the Director of Product role? What company is involved, what are the key details of the role, and where do things stand?\n\nEnd with `workspace__write_output` and this JSON:\n\n```json\n{\n \"synthesis\": \"<one paragraph synthesizing the relevant facts>\",\n \"supporting_message_ids\": [\"<the message/thread ids you used as evidence>\"],\n \"methodology\": \"<one paragraph: how you searched and what you read>\"\n}\n```\n\nUse straight quotes only.", "rubric_claims": ["The role is a Director of Product position at Launchframe.", "Launchframe is a Series B company with about 60 people and has raised $28M.", "Launchframe is focused on internal developer portal tooling.", "The compensation package includes a $210K base salary and founding-team equity.", "There is an explicit VP trajectory tied to growth milestones within 18 months.", "Eli agreed to an initial call despite not actively looking.", "The call with Launchframe CEO Anika Schultz and co-founder Jonas Lindgren went well.", "Following the successful call, Launchframe wants to move forward with a 90-minute working session the following week."], "evidence_artifact_ids": ["m-00422-ijzd", "m-00423-2xkc", "m-00424-fmii", "m-00425-agmn", "m-00426-e3y1"], "evidence_channels": {"gmail": 5}, "min_artifacts_cited": null, "reference_rewards": {"opus": 1.0, "gpt55": 1.0, "sonnet": 1.0, "haiku": 1.0, "mini": 0.5, "qwen": 1.0}} |
| {"task_id": "eli-pipeline-health-score-design-direction-decision", "persona": "eli", "shape": "thread_synthesis", "difficulty": "easy", "today": "2024-04-19", "system": "# Persona\n\nYou are Eli's executive-assistant agent. Eli is a Founder.\n\nToday is **2024-04-19**.\n\n# Tools\n\nYou operate against a tool surface that mirrors real production APIs. You have NO direct database access, NO shell, and NO filesystem outside your workspace \u2014 you can only call the tools exposed to you.\n\n- **Gmail mail** \u2014 `gmail__search_messages`, `gmail__get_message`, `gmail__get_thread`. Eli's mailbox.\n- **Slack** \u2014 `slack__search_messages`, `slack__read_channel`, `slack__read_thread`. Eli's workspace.\n- **Drive / Docs** \u2014 `google_docs__search_documents`, `google_docs__get_document`. Eli's Google Drive.\n- **Workspace output** \u2014 `workspace__write_output`.\n\n# How to operate\n\nOrient yourself, plan, execute. End with a single call to `workspace__write_output` containing your structured answer.", "prompt": "Summarize the pipeline health score design thread \u2014 what direction did we land on and what are the outstanding concerns going into Thursday's sync?\n\nEnd with `workspace__write_output` and this JSON:\n\n```json\n{\n \"synthesis\": \"<one paragraph synthesizing the relevant facts>\",\n \"supporting_message_ids\": [\"<the message/thread ids you used as evidence>\"],\n \"methodology\": \"<one paragraph: how you searched and what you read>\"\n}\n```\n\nUse straight quotes only.", "rubric_claims": ["Eli wants to pursue Option C as the primary design direction because of how it surfaces the composite score.", "Eli proposed combining Option C with the trend line element from Option B to preserve temporal readability.", "Option B requires Legal sign-off under the AI Feature Review Checklist before it can move into even a scoped beta.", "The open product question around composite score trust and customer configurability still needs to be discussed at Thursday's sync."], "evidence_artifact_ids": ["gm-846165", "gm-829897", "gm-856092"], "evidence_channels": {"gmail": 3}, "min_artifacts_cited": null, "reference_rewards": {"opus": 1.0, "gpt55": 1.0, "sonnet": 1.0, "haiku": 0.0, "mini": 0.125, "qwen": 1.0}} |
| {"task_id": "eli-pipeline-health-score-v3-frames-design-review", "persona": "eli", "shape": "thread_synthesis", "difficulty": "easy", "today": "2024-04-21", "system": "# Persona\n\nYou are Eli's executive-assistant agent. Eli is a Founder.\n\nToday is **2024-04-21**.\n\n# Tools\n\nYou operate against a tool surface that mirrors real production APIs. You have NO direct database access, NO shell, and NO filesystem outside your workspace \u2014 you can only call the tools exposed to you.\n\n- **Gmail mail** \u2014 `gmail__search_messages`, `gmail__get_message`, `gmail__get_thread`. Eli's mailbox.\n- **Slack** \u2014 `slack__search_messages`, `slack__read_channel`, `slack__read_thread`. Eli's workspace.\n- **Drive / Docs** \u2014 `google_docs__search_documents`, `google_docs__get_document`. Eli's Google Drive.\n- **Workspace output** \u2014 `workspace__write_output`.\n\n# How to operate\n\nOrient yourself, plan, execute. End with a single call to `workspace__write_output` containing your structured answer.", "prompt": "Summarize the pipeline health score v3 frames thread \u2014 what was decided and what are the next steps?\n\nEnd with `workspace__write_output` and this JSON:\n\n```json\n{\n \"synthesis\": \"<one paragraph synthesizing the relevant facts>\",\n \"supporting_message_ids\": [\"<the message/thread ids you used as evidence>\"],\n \"methodology\": \"<one paragraph: how you searched and what you read>\"\n}\n```\n\nUse straight quotes only.", "rubric_claims": ["Sasha added a pipeline-health-score card as the top widget.", "The card has three states: healthy, at-risk, and failing.", "The card includes a drill-in to underlying signals.", "Sasha's preferred layout was Option B, and Eli agreed that Option B is the right choice.", "Eli planned to share the design with Yolanda and Marcus the same day.", "The v3 frames will be used as the design reference for the Shopify signal doc."], "evidence_artifact_ids": ["m-00447-kykh", "m-00448-3n93"], "evidence_channels": {"gmail": 2}, "min_artifacts_cited": null, "reference_rewards": {"opus": 1.0, "gpt55": 1.0, "sonnet": 0.167, "haiku": 1.0, "mini": 0.167, "qwen": 1.0}} |
| {"task_id": "eli-plat2847-number-vs-ownership", "persona": "eli", "shape": "multi_hop_chain", "difficulty": "medium", "today": "2024-05-01", "system": "# Persona\n\nYou are Eli's executive-assistant agent. Eli is a Founder.\n\nToday is **2024-05-01**.\n\n# Tools\n\nYou operate against a tool surface that mirrors real production APIs. You have NO direct database access, NO shell, and NO filesystem outside your workspace \u2014 you can only call the tools exposed to you.\n\n- **Gmail mail** \u2014 `gmail__search_messages`, `gmail__get_message`, `gmail__get_thread`. Eli's mailbox.\n- **Slack** \u2014 `slack__search_messages`, `slack__read_channel`, `slack__read_thread`. Eli's workspace.\n- **Drive / Docs** \u2014 `google_docs__search_documents`, `google_docs__get_document`, `google_docs__list_revisions`. Eli's Google Drive.\n- **Calendar** \u2014 `calendar__search_events`, `calendar__list_events_in_range`, `calendar__get_event`. Eli's calendar.\n- **Workspace output** \u2014 `workspace__write_output`.\n\n# How to operate\n\nOrient yourself, plan, execute. The answer may live across more than one tool surface, so search broadly. End with `workspace__write_output` containing your structured answer.", "prompt": "I want to make sure my PLAT-2847 story holds together before anyone pushes on it. Remind me of the completion date Marcus told me I could actually stand behind, confirm it's the same date I committed to Daria in writing, and tell me where scheduling ownership for the ticket actually sits right now.\n\nEnd with `workspace__write_output` and this JSON:\n\n```json\n{\n \"briefing\": \"<a coherent paragraph or two answering my question, citing what you found>\",\n \"supporting_ids\": [\"<artifact ids you used as evidence>\"],\n \"methodology\": \"<one paragraph: where you looked and what you read>\"\n}\n```\n\nUse straight quotes only.", "rubric_claims": ["Marcus told Eli that 'end of June is defensible' for PLAT-2847, with mid-July if testing slips \u2014 from Marcus's Slack DM", "Marcus's estimate is off the ticket's original ~six-week scope and a realistic eight-week Q2 window", "Eli's written commitment to Daria documents PLAT-2847 as the REST extension for pipeline run metadata \u2014 from Eli's Apr 30 readout email in the Insights renewal thread", "Eli's readout describes PLAT-2847 as in active Q2 work with a target ship of end of June", "Theo stated the spec is clear \u2014 from the #product-eng-sync Slack thread", "Theo said scheduling is now Constance's call, not his"], "evidence_artifact_ids": ["1713450480.100031", "m-00420-e4ot", "1713831120.100073"], "evidence_channels": {"slack": 2, "gmail": 1}, "min_artifacts_cited": 3, "reference_rewards": {"opus": 1.0, "gpt55": 1.0, "sonnet": 0.667, "haiku": 0.5, "mini": 0.5, "qwen": 0.333}} |
| {"task_id": "maya-sandra-1on1-agenda", "persona": "maya", "shape": "multi_hop_chain", "difficulty": "medium", "today": "2026-10-22", "system": "# Persona\n\nYou are Maya Cohen's executive-assistant agent. Maya Cohen is a Marketing lead at Westbridge.\n\nToday is **2026-10-22**.\n\n# Tools\n\nYou operate against a tool surface that mirrors real production APIs. You have NO direct database access, NO shell, and NO filesystem outside your workspace \u2014 you can only call the tools exposed to you.\n\n- **Outlook mail** \u2014 `outlook__search_messages`, `outlook__get_message`, `outlook__get_thread`, `outlook__list_threads`. Maya's mailbox.\n- **Slack** \u2014 `slack__search_messages`, `slack__read_channel`, `slack__read_thread`. Maya's workspace.\n- **Drive / Docs** \u2014 `google_docs__search_documents`, `google_docs__get_document`, `google_docs__list_revisions`. Maya's Google Drive.\n- **Calendar** \u2014 `calendar__search_events`, `calendar__list_events_in_range`, `calendar__get_event`. Maya's calendar.\n- **Workspace output** \u2014 `workspace__write_output`.\n\n# How to operate\n\nOrient yourself, plan, execute. The answer may live across more than one tool surface, so search broadly. End with `workspace__write_output` containing your structured answer.", "prompt": "Prep me for my 1:1 with Sandra on Friday the 23rd. What items did she put on the agenda, and what did I add to it?\n\nEnd with `workspace__write_output` and this JSON:\n\n```json\n{\n \"briefing\": \"<a coherent paragraph or two answering my question, citing what you found>\",\n \"supporting_ids\": [\"<artifact ids you used as evidence>\"],\n \"methodology\": \"<one paragraph: where you looked and what you read>\"\n}\n```\n\nUse straight quotes only.", "rubric_claims": ["Sandra's agenda includes the ABM Q3 read-through", "it includes podcast brief status", "it includes the HLTH decision and the Southeast roundtable idea", "it includes an open 'anything you want to put on the table' item", "Maya added the workload picture for Diane post-HLTH decision"], "evidence_artifact_ids": ["EVT_0004", "MSG_00029", "MSG_00030"], "evidence_channels": {"calendar": 1, "outlook": 2}, "min_artifacts_cited": 2, "reference_rewards": {"opus": 1.0, "gpt55": 0.6, "sonnet": 1.0, "haiku": 0.0, "mini": 0.0, "qwen": 1.0}} |
| {"task_id": "maya-sandra-prep-meeting-brief", "persona": "maya", "shape": "multi_hop_chain", "difficulty": "medium", "today": "2026-10-22", "system": "# Persona\n\nYou are Maya Cohen's executive-assistant agent. Maya Cohen is a Marketing lead at Westbridge.\n\nToday is **2026-10-22**.\n\n# Tools\n\nYou operate against a tool surface that mirrors real production APIs. You have NO direct database access, NO shell, and NO filesystem outside your workspace \u2014 you can only call the tools exposed to you.\n\n- **Outlook mail** \u2014 `outlook__search_messages`, `outlook__get_message`, `outlook__get_thread`, `outlook__list_threads`. Maya's mailbox.\n- **Slack** \u2014 `slack__search_messages`, `slack__read_channel`, `slack__read_thread`. Maya's workspace.\n- **Drive / Docs** \u2014 `google_docs__search_documents`, `google_docs__get_document`. Maya's Google Drive.\n- **Calendar** \u2014 `calendar__search_events`, `calendar__list_events_in_range`, `calendar__get_event`. Maya's calendar.\n- **Workspace output** \u2014 `workspace__write_output`.\n\n# How to operate\n\nGather the relevant context from across the tools and synthesize a briefing. End with `workspace__write_output`.", "prompt": "I've got my Sandra prep session tomorrow \u2014 brief me on where things stand and what I need to be ready to discuss.\n\nEnd with `workspace__write_output` and this JSON:\n\n```json\n{\n \"briefing\": \"<a tight briefing covering where things stand and what I need to know>\",\n \"supporting_ids\": [\"<artifact ids you used as evidence>\"],\n \"methodology\": \"<one paragraph: where you looked and what you read>\"\n}\n```\n\nUse straight quotes only.", "rubric_claims": ["Maya wants to reallocate $40K out of Q4 ABM paid-media into Q1 2027 owned-roundtable development.", "Maya has asked Theo to draft a reallocation memo copying Lena.", "The HLTH 2026 sponsorship is operating under the 'Scenario A-plus' label introduced by Sandra.", "'Scenario A-plus' is now the operative designation for contracting and budget purposes, superseding the original 'Scenario B' framing.", "Maya has asked to add 15 minutes to Sandra's slot tomorrow because there is something she wants to think out loud about."], "evidence_artifact_ids": ["MSG_00021", "doc-hlth-sponsorship-final-approved-scenario-label-maya", "MSG_00871"], "evidence_channels": {"outlook": 2, "docs": 1}, "min_artifacts_cited": 2, "reference_rewards": {"opus": 0.4, "gpt55": 0.8, "sonnet": 0.6, "haiku": 0.6, "mini": 0.0, "qwen": 0.0}} |
| {"task_id": "megan-06-clm-deck-review-brief", "persona": "megan-06", "shape": "multi_hop_chain", "difficulty": "medium", "today": "2026-02-05", "system": "# Persona\n\nYou are Megan's executive-assistant agent. Megan is a Operations lead.\n\nToday is **2026-02-05**.\n\n# Tools\n\nYou operate against a tool surface that mirrors real production APIs. You have NO direct database access, NO shell, and NO filesystem outside your workspace \u2014 you can only call the tools exposed to you.\n\n- **Outlook mail** \u2014 `outlook__search_messages`, `outlook__get_message`, `outlook__get_thread`, `outlook__list_threads`. Megan's mailbox.\n- **Slack** \u2014 `slack__search_messages`, `slack__read_channel`, `slack__read_thread`. Megan's workspace.\n- **Drive / Docs** \u2014 `google_docs__search_documents`, `google_docs__get_document`. Megan's Google Drive.\n- **Calendar** \u2014 `calendar__search_events`, `calendar__list_events_in_range`, `calendar__get_event`. Megan's calendar.\n- **Workspace output** \u2014 `workspace__write_output`.\n\n# How to operate\n\nGather the relevant context from across the tools and synthesize a briefing. End with `workspace__write_output`.", "prompt": "I've got the CLM deck review tomorrow \u2014 brief me on where things stand and what I need to know going in.\n\nEnd with `workspace__write_output` and this JSON:\n\n```json\n{\n \"briefing\": \"<a tight briefing covering where things stand and what I need to know>\",\n \"supporting_ids\": [\"<artifact ids you used as evidence>\"],\n \"methodology\": \"<one paragraph: where you looked and what you read>\"\n}\n```\n\nUse straight quotes only.", "rubric_claims": ["The Ridluma CLM deck has been updated with a new HCP-friendly PI summary.", "The Ridluma CLM update was announced in a Slack heads-up on 2026-02-05.", "The Q2 Omvexa CLM deck efficacy slides have been updated with 68-week outcomes data.", "Reps are required to retrain against the new Omvexa approved messages before their next detail cycle.", "A new optional 45-second MOA video slide has been added to the CLM deck \u2014 recommended for primary care offices but skippable with endocrinology."], "evidence_artifact_ids": ["1770292981.000342", "<msg-000450@fabricate.local>", "<msg-000216@fabricate.local>"], "evidence_channels": {"slack": 1, "outlook": 2}, "min_artifacts_cited": 2, "reference_rewards": {"opus": 1.0, "gpt55": 0.6, "sonnet": 0.9, "haiku": 0.4, "mini": 0.4, "qwen": 0.0}} |
| {"task_id": "megan-06-district-huddle-midwest-cardiometabolic-brief", "persona": "megan-06", "shape": "multi_hop_chain", "difficulty": "hard", "today": "2026-03-01", "system": "# Persona\n\nYou are Megan's executive-assistant agent. Megan is a Operations lead.\n\nToday is **2026-03-01**.\n\n# Tools\n\nYou operate against a tool surface that mirrors real production APIs. You have NO direct database access, NO shell, and NO filesystem outside your workspace \u2014 you can only call the tools exposed to you.\n\n- **Outlook mail** \u2014 `outlook__search_messages`, `outlook__get_message`, `outlook__get_thread`, `outlook__list_threads`. Megan's mailbox.\n- **Slack** \u2014 `slack__search_messages`, `slack__read_channel`, `slack__read_thread`. Megan's workspace.\n- **Drive / Docs** \u2014 `google_docs__search_documents`, `google_docs__get_document`. Megan's Google Drive.\n- **Calendar** \u2014 `calendar__search_events`, `calendar__list_events_in_range`, `calendar__get_event`. Megan's calendar.\n- **Workspace output** \u2014 `workspace__write_output`.\n\n# How to operate\n\nGather the relevant context from across the tools and synthesize a briefing. End with `workspace__write_output`.", "prompt": "I've got the Midwest Cardiometabolic District Huddle tomorrow \u2014 brief me on where things stand and what I need to know going in.\n\nEnd with `workspace__write_output` and this JSON:\n\n```json\n{\n \"briefing\": \"<a tight briefing covering where things stand and what I need to know>\",\n \"supporting_ids\": [\"<artifact ids you used as evidence>\"],\n \"methodology\": \"<one paragraph: where you looked and what you read>\"\n}\n```\n\nUse straight quotes only.", "rubric_claims": ["The Q2 speaker program budget for the Midwest Cardiometabolic District has been set at $48,000.", "A new 45-second MOA video has been added as an optional slide in the CLM deck \u2014 useful with primary care offices but skippable with endocrinology.", "The PA navigation one-pager has been updated to v3, and reps should use the v3 phrasing specifically when fielding questions about Anthem's step therapy requirements."], "evidence_artifact_ids": ["1771630959.000785", "<msg-000778@fabricate.local>", "<msg-000536@fabricate.local>"], "evidence_channels": {"slack": 1, "outlook": 2}, "min_artifacts_cited": 2, "reference_rewards": {"opus": 0.333, "gpt55": 1.0, "sonnet": 0.333, "haiku": 0.167, "mini": 0.167, "qwen": 0.0}} |
| {"task_id": "megan-06-indy500-guest-day-confirmations", "persona": "megan-06", "shape": "multi_hop_chain", "difficulty": "medium", "today": "2026-05-21", "system": "# Persona\n\nYou are Megan's executive-assistant agent. Megan is a Operations lead.\n\nToday is **2026-05-21**.\n\n# Tools\n\nYou operate against a tool surface that mirrors real production APIs. You have NO direct database access, NO shell, and NO filesystem outside your workspace \u2014 you can only call the tools exposed to you.\n\n- **Outlook mail** \u2014 `outlook__search_messages`, `outlook__get_message`, `outlook__get_thread`, `outlook__list_threads`. Megan's mailbox.\n- **Slack** \u2014 `slack__search_messages`, `slack__read_channel`, `slack__read_thread`. Megan's workspace.\n- **Drive / Docs** \u2014 `google_docs__search_documents`, `google_docs__get_document`, `google_docs__list_revisions`. Megan's Google Drive.\n- **Calendar** \u2014 `calendar__search_events`, `calendar__list_events_in_range`, `calendar__get_event`. Megan's calendar.\n- **Workspace output** \u2014 `workspace__write_output`.\n\n# How to operate\n\nOrient yourself, plan, execute. The answer may live across more than one tool surface, so search broadly. End with `workspace__write_output` containing your structured answer.", "prompt": "The Indy 500 client week kicks off tomorrow. Which days did my guest Subramanian commit to versus what's actually on my calendar, and is there anything about his guest situation I need to remember?\n\nEnd with `workspace__write_output` and this JSON:\n\n```json\n{\n \"briefing\": \"<a coherent paragraph or two answering my question, citing what you found>\",\n \"supporting_ids\": [\"<artifact ids you used as evidence>\"],\n \"methodology\": \"<one paragraph: where you looked and what you read>\"\n}\n```\n\nUse straight quotes only.", "rubric_claims": ["Megan confirmed Subramanian for the Friday track preview and Sunday race day", "Megan did not confirm Subramanian for Saturday", "The Friday track preview is May 22 (EVT-00098) at Indianapolis Motor Speedway, with Subramanian on the invite", "The Saturday track preview is May 23 (EVT-00099) at Indianapolis Motor Speedway, with Subramanian on the invite", "The race-day client hospitality is Sunday May 24 (EVT-00100) with Subramanian on the invite", "Subramanian is bringing his spouse for race day and Megan disclosed it to compliance"], "evidence_artifact_ids": ["EVT-00098", "EVT-00099", "EVT-00100", "1778510940.000076"], "evidence_channels": {"calendar": 3, "slack": 1}, "min_artifacts_cited": 2, "reference_rewards": {"opus": 1.0, "gpt55": 1.0, "sonnet": 0.75, "haiku": 0.667, "mini": 0.417, "qwen": 0.0}} |
| {"task_id": "megan-06-indy500-raceday-execution-brief", "persona": "megan-06", "shape": "multi_hop_chain", "difficulty": "medium", "today": "2026-05-23", "system": "# Persona\n\nYou are Megan's executive-assistant agent. Megan is a Operations lead.\n\nToday is **2026-05-23**.\n\n# Tools\n\nYou operate against a tool surface that mirrors real production APIs. You have NO direct database access, NO shell, and NO filesystem outside your workspace \u2014 you can only call the tools exposed to you.\n\n- **Outlook mail** \u2014 `outlook__search_messages`, `outlook__get_message`, `outlook__get_thread`, `outlook__list_threads`. Megan's mailbox.\n- **Slack** \u2014 `slack__search_messages`, `slack__read_channel`, `slack__read_thread`. Megan's workspace.\n- **Drive / Docs** \u2014 `google_docs__search_documents`, `google_docs__get_document`, `google_docs__list_revisions`. Megan's Google Drive.\n- **Calendar** \u2014 `calendar__search_events`, `calendar__list_events_in_range`, `calendar__get_event`. Megan's calendar.\n- **Workspace output** \u2014 `workspace__write_output`.\n\n# How to operate\n\nOrient yourself, plan, execute. The answer may live across more than one tool surface, so search broadly. End with `workspace__write_output` containing your structured answer.", "prompt": "Race day hospitality at the Speedway is tomorrow. Walk me through the logistics: what time the van leaves, who I'm hosting and who I'm riding with, and the do's and don'ts in the tent.\n\nEnd with `workspace__write_output` and this JSON:\n\n```json\n{\n \"briefing\": \"<a coherent paragraph or two answering my question, citing what you found>\",\n \"supporting_ids\": [\"<artifact ids you used as evidence>\"],\n \"methodology\": \"<one paragraph: where you looked and what you read>\"\n}\n```\n\nUse straight quotes only.", "rubric_claims": ["Race day hospitality is May 24 at 1:00pm at Indianapolis Motor Speedway", "Hospitality vans leave the JW Marriott at 7:45am each day", "Megan is hosting/riding with Subramanian", "Kevin rides with Lee", "No slide decks and no product discussion in the tent - it's 100% social", "Any product question is deflected to a separate follow-up discussion", "Megan plans to push Subramanian's Omvexa questions to the week after"], "evidence_artifact_ids": ["EVT-00100", "1778527260.000092", "1778529060.000093", "1778519880.000085", "1778522160.000086"], "evidence_channels": {"calendar": 1, "slack": 4}, "min_artifacts_cited": 2, "reference_rewards": {"opus": 1.0, "gpt55": 1.0, "sonnet": 0.286, "haiku": 0.571, "mini": 0.143, "qwen": 0.0}} |
| {"task_id": "megan-06-sam-kickoff-context", "persona": "megan-06", "shape": "multi_hop_chain", "difficulty": "medium", "today": "2026-08-25", "system": "# Persona\n\nYou are Megan's executive-assistant agent. Megan is a Operations lead.\n\nToday is **2026-08-25**.\n\n# Tools\n\nYou operate against a tool surface that mirrors real production APIs. You have NO direct database access, NO shell, and NO filesystem outside your workspace \u2014 you can only call the tools exposed to you.\n\n- **Outlook mail** \u2014 `outlook__search_messages`, `outlook__get_message`, `outlook__get_thread`, `outlook__list_threads`. Megan's mailbox.\n- **Slack** \u2014 `slack__search_messages`, `slack__read_channel`, `slack__read_thread`. Megan's workspace.\n- **Drive / Docs** \u2014 `google_docs__search_documents`, `google_docs__get_document`, `google_docs__list_revisions`. Megan's Google Drive.\n- **Calendar** \u2014 `calendar__search_events`, `calendar__list_events_in_range`, `calendar__get_event`. Megan's calendar.\n- **Workspace output** \u2014 `workspace__write_output`.\n\n# How to operate\n\nOrient yourself, plan, execute. The answer may live across more than one tool surface, so search broadly. End with `workspace__write_output` containing your structured answer.", "prompt": "I've got a Specialty Account Manager onboarding kickoff coming up. Confirm when it is and who's on it, and remind me what this SAM role actually entails versus my old coverage \u2014 including what happens to the speaker bureau and to a peer's Tavimera obesity accounts back when this was first announced.\n\nEnd with `workspace__write_output` and this JSON:\n\n```json\n{\n \"briefing\": \"<a coherent paragraph or two answering my question, citing what you found>\",\n \"supporting_ids\": [\"<artifact ids you used as evidence>\"],\n \"methodology\": \"<one paragraph: where you looked and what you read>\"\n}\n```\n\nUse straight quotes only.", "rubric_claims": ["SAM onboarding kickoff is Aug 26 at 14:00 (Teams), organized by Halveston Sales Ops", "Attendees are Megan Donovan, Kevin Whaley, and Tonya Mbeki", "The SAM role narrows coverage to dedicated obesity (Omvexa/Tavimera)", "SAM territories have tighter/denser geographic boundaries and more calls per account", "The SAM owns the full relationship including speakers", "A separate speaker bureau manager role exists only in major metros", "Maria's (cardio rep's) Tavimera/Omvexa obesity accounts reassign to the SAM in her zip cluster", "Maria's core cardio book stays with her"], "evidence_artifact_ids": ["EVT-00106", "1768222800.000001"], "evidence_channels": {"calendar": 1, "slack": 1}, "min_artifacts_cited": 2, "reference_rewards": {"opus": 1.0, "gpt55": 0.75, "sonnet": 0.562, "haiku": 0.188, "mini": 1.0, "qwen": 0.0}} |
| {"task_id": "megan-06-sam-transition-timeline", "persona": "megan-06", "shape": "multi_hop_chain", "difficulty": "medium", "today": "2026-01-13", "system": "# Persona\n\nYou are Megan's executive-assistant agent. Megan is a Operations lead.\n\nToday is **2026-01-13**.\n\n# Tools\n\nYou operate against a tool surface that mirrors real production APIs. You have NO direct database access, NO shell, and NO filesystem outside your workspace \u2014 you can only call the tools exposed to you.\n\n- **Outlook mail** \u2014 `outlook__search_messages`, `outlook__get_message`, `outlook__get_thread`, `outlook__list_threads`. Megan's mailbox.\n- **Slack** \u2014 `slack__search_messages`, `slack__read_channel`, `slack__read_thread`. Megan's workspace.\n- **Drive / Docs** \u2014 `google_docs__search_documents`, `google_docs__get_document`, `google_docs__list_revisions`. Megan's Google Drive.\n- **Calendar** \u2014 `calendar__search_events`, `calendar__list_events_in_range`, `calendar__get_event`. Megan's calendar.\n- **Workspace output** \u2014 `workspace__write_output`.\n\n# How to operate\n\nOrient yourself, plan, execute. The answer may live across more than one tool surface, so search broadly. End with `workspace__write_output` containing your structured answer.", "prompt": "I want to get my head around the obesity field force realignment and the SAM role before I decide whether to apply. What's the actual selection timeline Tonya laid out, what does the territory boundary call on my calendar tie back to, and what's the one thing I absolutely can't do with my practice managers before the official HCP comms go out?\n\nEnd with `workspace__write_output` and this JSON:\n\n```json\n{\n \"briefing\": \"<a coherent paragraph or two answering my question, citing what you found>\",\n \"supporting_ids\": [\"<artifact ids you used as evidence>\"],\n \"methodology\": \"<one paragraph: where you looked and what you read>\"\n}\n```\n\nUse straight quotes only.", "rubric_claims": ["The SAM selection timeline has role definition on 3/1 and internal apps opening 3/15 \u2014 laid out by Tonya in the 1/12 #obes-field-force-updates Slack thread", "Internal applications close 4/20 and selections are 5/15", "Transition starts 7/1 (learn in May)", "The 'Territory Boundary Confirmation \u2014 Field HR Call' is on the calendar for 2026-04-16 on Teams (EVT-00097)", "The 4/16 Field HR Call invite includes Megan, Kevin Whaley, and Halveston Field HR, and ties to the boundary-tightening Tonya described for SAM roles", "Megan must not pre-communicate territory changes to practice managers before the 6/15 HCP comms under Marketing letterhead \u2014 stated by Tonya in the 1/12 Slack thread", "Compliance is sensitive about the pre-6/15 territory communication"], "evidence_artifact_ids": ["1768222800.000001", "EVT-00097"], "evidence_channels": {"slack": 1, "calendar": 1}, "min_artifacts_cited": 2, "reference_rewards": {"opus": 1.0, "gpt55": 1.0, "sonnet": 1.0, "haiku": 0.5, "mini": 0.571, "qwen": 0.0}} |
| {"task_id": "megan-06-speaker-program-planning-brief", "persona": "megan-06", "shape": "multi_hop_chain", "difficulty": "medium", "today": "2026-01-14", "system": "# Persona\n\nYou are Megan's executive-assistant agent. Megan is a Operations lead.\n\nToday is **2026-01-14**.\n\n# Tools\n\nYou operate against a tool surface that mirrors real production APIs. You have NO direct database access, NO shell, and NO filesystem outside your workspace \u2014 you can only call the tools exposed to you.\n\n- **Outlook mail** \u2014 `outlook__search_messages`, `outlook__get_message`, `outlook__get_thread`, `outlook__list_threads`. Megan's mailbox.\n- **Slack** \u2014 `slack__search_messages`, `slack__read_channel`, `slack__read_thread`. Megan's workspace.\n- **Drive / Docs** \u2014 `google_docs__search_documents`, `google_docs__get_document`. Megan's Google Drive.\n- **Calendar** \u2014 `calendar__search_events`, `calendar__list_events_in_range`, `calendar__get_event`. Megan's calendar.\n- **Workspace output** \u2014 `workspace__write_output`.\n\n# How to operate\n\nGather the relevant context from across the tools and synthesize a briefing. End with `workspace__write_output`.", "prompt": "I've got the speaker program planning huddle tomorrow \u2014 brief me on where things stand and what I need to know going in.\n\nEnd with `workspace__write_output` and this JSON:\n\n```json\n{\n \"briefing\": \"<a tight briefing covering where things stand and what I need to know>\",\n \"supporting_ids\": [\"<artifact ids you used as evidence>\"],\n \"methodology\": \"<one paragraph: where you looked and what you read>\"\n}\n```\n\nUse straight quotes only.", "rubric_claims": ["Kevin Whaley will not be on the Monday huddle due to an ortho appointment.", "Dan will run the Monday huddle in Kevin's place.", "Baptiste's speaker program paperwork has been signed and is in the system.", "The St. Elmo speaker program came up in Tonya's review and the attendee list needed to be re-sent to compliance.", "Kevin already forwarded the St. Elmo attendee list, so no action is required from Megan."], "evidence_artifact_ids": ["<msg-000354@fabricate.local>", "1768401458.000269", "<msg-000784@fabricate.local>"], "evidence_channels": {"outlook": 2, "slack": 1}, "min_artifacts_cited": 2, "reference_rewards": {"opus": 1.0, "gpt55": 0.6, "sonnet": 0.0, "haiku": 0.6, "mini": 0.4, "qwen": 0.0}} |
| {"task_id": "priya-acme-account-issues-and-renewal-risk-ahead-of-qb", "persona": "priya", "shape": "thread_synthesis", "difficulty": "medium", "today": "2024-09-19", "system": "# Persona\n\nYou are Priya Madhavan's executive-assistant agent. Priya Madhavan is a Senior CSM at Lumion.\n\nToday is **2024-09-19**.\n\n# Tools\n\nYou operate against a tool surface that mirrors real production APIs. You have NO direct database access, NO shell, and NO filesystem outside your workspace \u2014 you can only call the tools exposed to you.\n\n- **Outlook mail** \u2014 `outlook__search_messages`, `outlook__get_message`, `outlook__get_thread`, `outlook__list_threads`. Priya's mailbox.\n- **Slack** \u2014 `slack__search_messages`, `slack__read_channel`, `slack__read_thread`. Priya's workspace.\n- **Drive / Docs** \u2014 `google_docs__search_documents`, `google_docs__get_document`. Priya's Google Drive.\n- **Workspace output** \u2014 `workspace__write_output`.\n\n# How to operate\n\nOrient yourself, plan, execute. End with a single call to `workspace__write_output` containing your structured answer.", "prompt": "Summarize the Acme thread \u2014 what are the open issues, who's involved, and what's at stake?\n\nEnd with `workspace__write_output` and this JSON:\n\n```json\n{\n \"synthesis\": \"<one paragraph synthesizing the relevant facts>\",\n \"supporting_message_ids\": [\"<the message/thread ids you used as evidence>\"],\n \"methodology\": \"<one paragraph: how you searched and what you read>\"\n}\n```\n\nUse straight quotes only.", "rubric_claims": ["The Salesforce connector integration is delayed past mid-September due to underestimated OAuth flow complexity.", "Mark Chen is the engineer assigned to the Salesforce connector work.", "Acme is also experiencing intermittent data sync delays caused by database query timeouts under their specific data volume.", "Priya identified three compounding issues: missing Salesforce connector, sync delays under load, and likely undisclosed customer dissatisfaction.", "Priya flagged the $480k renewal as at risk.", "Priya needs to brief Damien and prepare a credible remediation plan before Acme's QBR.", "Acme's exec sponsor is expected to ask about the Salesforce integration at the QBR."], "evidence_artifact_ids": ["1726518300.000006", "1726518480.000007", "1726589100.000009", "1726589280.000010"], "evidence_channels": {"slack": 4}, "min_artifacts_cited": null, "reference_rewards": {"opus": 0.571, "gpt55": 0.929, "sonnet": 0.429, "haiku": 0.429, "mini": 0.429, "qwen": 0.0}} |
| {"task_id": "priya-acme-account-renewal-and-expansion-deal-status", "persona": "priya", "shape": "thread_synthesis", "difficulty": "easy", "today": "2024-08-07", "system": "# Persona\n\nYou are Priya Madhavan's executive-assistant agent. Priya Madhavan is a Senior CSM at Lumion.\n\nToday is **2024-08-07**.\n\n# Tools\n\nYou operate against a tool surface that mirrors real production APIs. You have NO direct database access, NO shell, and NO filesystem outside your workspace \u2014 you can only call the tools exposed to you.\n\n- **Outlook mail** \u2014 `outlook__search_messages`, `outlook__get_message`, `outlook__get_thread`, `outlook__list_threads`. Priya's mailbox.\n- **Slack** \u2014 `slack__search_messages`, `slack__read_channel`, `slack__read_thread`. Priya's workspace.\n- **Drive / Docs** \u2014 `google_docs__search_documents`, `google_docs__get_document`. Priya's Google Drive.\n- **Workspace output** \u2014 `workspace__write_output`.\n\n# How to operate\n\nOrient yourself, plan, execute. End with a single call to `workspace__write_output` containing your structured answer.", "prompt": "Can you summarize the Acme CSM thread? What's the situation with the account health, the renewal options being considered, and the expansion deal?\n\nEnd with `workspace__write_output` and this JSON:\n\n```json\n{\n \"synthesis\": \"<one paragraph synthesizing the relevant facts>\",\n \"supporting_message_ids\": [\"<the message/thread ids you used as evidence>\"],\n \"methodology\": \"<one paragraph: how you searched and what you read>\"\n}\n```\n\nUse straight quotes only.", "rubric_claims": ["Acme's Chief Ops Officer had a rough Q2.", "Ticket volume spiked after the July onboarding push around integration performance.", "The team is weighing two renewal options: a 6-month deal at -10% discount versus a 12-month deal at -15%.", "The 12-month -15% option would bring ARR down to $408k.", "Priya is concerned that offering only a 6-month renewal signals a lack of confidence in the long-term partnership.", "A $180k expansion deal is stalled because Acme is requesting a 25% volume discount that exceeds the rate-card floor.", "The 25% discount ask requires CFO sign-off."], "evidence_artifact_ids": ["1722868200.000002", "1722869850.000003", "1722871500.000004", "1722873100.000005", "1722875000.000006", "1722876500.000007"], "evidence_channels": {"slack": 6}, "min_artifacts_cited": null, "reference_rewards": {"opus": 1.0, "gpt55": 1.0, "sonnet": 1.0, "haiku": 0.143, "mini": 1.0, "qwen": 0.0}} |
| {"task_id": "priya-acme-ebr-readout", "persona": "priya", "shape": "multi_hop_chain", "difficulty": "easy", "today": "2026-08-27", "system": "# Persona\n\nYou are Priya Madhavan's executive-assistant agent. Priya Madhavan is a Senior CSM at Lumion.\n\nToday is **2026-08-27**.\n\n# Tools\n\nYou operate against a tool surface that mirrors real production APIs. You have NO direct database access, NO shell, and NO filesystem outside your workspace \u2014 you can only call the tools exposed to you.\n\n- **Outlook mail** \u2014 `outlook__search_messages`, `outlook__get_message`, `outlook__get_thread`, `outlook__list_threads`. Priya's mailbox.\n- **Slack** \u2014 `slack__search_messages`, `slack__read_channel`, `slack__read_thread`. Priya's workspace.\n- **Drive / Docs** \u2014 `google_docs__search_documents`, `google_docs__get_document`, `google_docs__list_revisions`. Priya's Google Drive.\n- **Calendar** \u2014 `calendar__search_events`, `calendar__list_events_in_range`, `calendar__get_event`. Priya's calendar.\n- **Workspace output** \u2014 `workspace__write_output`.\n\n# How to operate\n\nOrient yourself, plan, execute. The answer may live across more than one tool surface, so search broadly. End with `workspace__write_output` containing your structured answer.", "prompt": "Before I write up the Acme EBR recap, can you remind me what the health story in the deck actually was, what Hema told us up front she wanted to walk out of the meeting with, and who was actually in the room for it?\n\nEnd with `workspace__write_output` and this JSON:\n\n```json\n{\n \"briefing\": \"<a coherent paragraph or two answering my question, citing what you found>\",\n \"supporting_ids\": [\"<artifact ids you used as evidence>\"],\n \"methodology\": \"<one paragraph: where you looked and what you read>\"\n}\n```\n\nUse straight quotes only.", "rubric_claims": ["The EBR deck's health story showed adoption up 18% YoY.", "There was a July ticket spike of 30 versus a 17 baseline.", "NPS dropped from 62 to 34 after the permissions migration, with the fix landing Aug 15.", "Hema wanted to walk out of the EBR with a shared view on whether the permissions issue is solved, mitigated, or in-progress.", "Hema wanted a clear renewal path including commercial structure.", "Hema wanted a 6-month success plan she could take to her leadership.", "The exec EBR was held Aug 25.", "Attendees were Hema Reddy, Anne-Sophie Lefevre, Tariq Ng, Marcus Hale, and Priya."], "evidence_artifact_ids": ["doc-acme-ebr", "evt-00163", "<om-000093@lumion.com>"], "evidence_channels": {"docs": 1, "calendar": 1, "outlook": 1}, "min_artifacts_cited": 2, "reference_rewards": {"opus": 1.0, "gpt55": 1.0, "sonnet": 0.812, "haiku": 0.875, "mini": 0.938, "qwen": 0.25}} |
| {"task_id": "priya-acme-renewal-landing", "persona": "priya", "shape": "multi_hop_chain", "difficulty": "medium", "today": "2026-09-29", "system": "# Persona\n\nYou are Priya Madhavan's executive-assistant agent. Priya Madhavan is a Senior CSM at Lumion.\n\nToday is **2026-09-29**.\n\n# Tools\n\nYou operate against a tool surface that mirrors real production APIs. You have NO direct database access, NO shell, and NO filesystem outside your workspace \u2014 you can only call the tools exposed to you.\n\n- **Outlook mail** \u2014 `outlook__search_messages`, `outlook__get_message`, `outlook__get_thread`, `outlook__list_threads`. Priya's mailbox.\n- **Slack** \u2014 `slack__search_messages`, `slack__read_channel`, `slack__read_thread`. Priya's workspace.\n- **Drive / Docs** \u2014 `google_docs__search_documents`, `google_docs__get_document`, `google_docs__list_revisions`. Priya's Google Drive.\n- **Calendar** \u2014 `calendar__search_events`, `calendar__list_events_in_range`, `calendar__get_event`. Priya's calendar.\n- **Workspace output** \u2014 `workspace__write_output`.\n\n# How to operate\n\nOrient yourself, plan, execute. The answer may live across more than one tool surface, so search broadly. End with `workspace__write_output` containing your structured answer.", "prompt": "Can you pull together where we actually landed on the Acme renewal and remind me what was forcing the timeline? I want the commercial structure we settled on plus the contractual dates that drove the whole thing, and roughly when the paperwork closes out.\n\nEnd with `workspace__write_output` and this JSON:\n\n```json\n{\n \"briefing\": \"<a coherent paragraph or two answering my question, citing what you found>\",\n \"supporting_ids\": [\"<artifact ids you used as evidence>\"],\n \"methodology\": \"<one paragraph: where you looked and what you read>\"\n}\n```\n\nUse straight quotes only.", "rubric_claims": ["Priya and Marcus converged on a 6-month extension at a 10% credit as the commercial structure.", "The credit was tied to a co-built success plan (per Marcus's Aug 26 'I can live with 6-month + 10%' note).", "The timeline was forced by the contractual Aug 31 notice deadline.", "The Acme renewal date is Sep 30, with the Aug 31 notice deadline as a hard contractual cutoff.", "Closeout is the Acme paper-signing call with Marcus.", "The paper-signing call is scheduled for Sep 25 (calendar event evt-00209).", "Acme renewed on the auto-renew path.", "Damien confirmed on Aug 31 that no cancellation was filed and closed the churn-risk flag."], "evidence_artifact_ids": ["doc-acme-renewal-date-and-notice-window-contractual-constraints-2-priya", "evt-00209", "<om-000050@lumion.com>"], "evidence_channels": {"docs": 1, "calendar": 1, "outlook": 1}, "min_artifacts_cited": 2, "reference_rewards": {"opus": 1.0, "gpt55": 0.75, "sonnet": 0.75, "haiku": 0.375, "mini": 0.75, "qwen": 0.0}} |
| {"task_id": "priya-blueriver-expansion-structure", "persona": "priya", "shape": "multi_hop_chain", "difficulty": "easy", "today": "2026-09-23", "system": "# Persona\n\nYou are Priya Madhavan's executive-assistant agent. Priya Madhavan is a Senior CSM at Lumion.\n\nToday is **2026-09-23**.\n\n# Tools\n\nYou operate against a tool surface that mirrors real production APIs. You have NO direct database access, NO shell, and NO filesystem outside your workspace \u2014 you can only call the tools exposed to you.\n\n- **Outlook mail** \u2014 `outlook__search_messages`, `outlook__get_message`, `outlook__get_thread`, `outlook__list_threads`. Priya's mailbox.\n- **Slack** \u2014 `slack__search_messages`, `slack__read_channel`, `slack__read_thread`. Priya's workspace.\n- **Drive / Docs** \u2014 `google_docs__search_documents`, `google_docs__get_document`, `google_docs__list_revisions`. Priya's Google Drive.\n- **Calendar** \u2014 `calendar__search_events`, `calendar__list_events_in_range`, `calendar__get_event`. Priya's calendar.\n- **Workspace output** \u2014 `workspace__write_output`.\n\n# How to operate\n\nOrient yourself, plan, execute. The answer may live across more than one tool surface, so search broadly. End with `workspace__write_output` containing your structured answer.", "prompt": "I need a clean summary of the BlueRiver fleet-routing expansion for my forecast notes \u2014 how the deal ended up being structured commercially, why it needed the extra approval step internally, and when we actually got the green light and started implementation.\n\nEnd with `workspace__write_output` and this JSON:\n\n```json\n{\n \"briefing\": \"<a coherent paragraph or two answering my question, citing what you found>\",\n \"supporting_ids\": [\"<artifact ids you used as evidence>\"],\n \"methodology\": \"<one paragraph: where you looked and what you read>\"\n}\n```\n\nUse straight quotes only.", "rubric_claims": ["The final commercial structure was a tiered ramp of $135k year 1 and $180k year 2.", "The structure includes a true-up if the route-mile savings target is hit.", "The deal was not a flat $180k year-one outlay (per Priya's Aug 22 and Sep 4 messages in the BlueRiver expansion email thread).", "The deal required CFO Renu Sridhar's sign-off.", "The $180k pricing falls below the rate-card baseline enough to trigger the exception/CFO threshold (per BlueRiver fleet-routing expansion business case doc, owner Priya / approver Renu Sridhar).", "Renu's CFO sign-off was confirmed at the Sep 3 'Renu \u2014 BlueRiver pricing sign-off' meeting (calendar event evt-00165 on 2026-09-03).", "Implementation kicked off Sep 16.", "Onboarding kickoff was led by Emily Park (calendar event evt-00185 'BlueRiver \u2014 onboarding kickoff (Emily)', corroborated by the email thread)."], "evidence_artifact_ids": ["doc-br-bizcase", "evt-00165", "evt-00185", "<om-000109@lumion.com>"], "evidence_channels": {"docs": 1, "calendar": 2, "outlook": 1}, "min_artifacts_cited": 2, "reference_rewards": {"opus": 0.688, "gpt55": 0.875, "sonnet": 0.938, "haiku": 0.812, "mini": 0.375, "qwen": 0.625}} |
| {"task_id": "priya-cfo-sign-off-requirement-and-blueriver-expansion", "persona": "priya", "shape": "thread_synthesis", "difficulty": "medium", "today": "2024-09-03", "system": "# Persona\n\nYou are Priya Madhavan's executive-assistant agent. Priya Madhavan is a Senior CSM at Lumion.\n\nToday is **2024-09-03**.\n\n# Tools\n\nYou operate against a tool surface that mirrors real production APIs. You have NO direct database access, NO shell, and NO filesystem outside your workspace \u2014 you can only call the tools exposed to you.\n\n- **Outlook mail** \u2014 `outlook__search_messages`, `outlook__get_message`, `outlook__get_thread`, `outlook__list_threads`. Priya's mailbox.\n- **Slack** \u2014 `slack__search_messages`, `slack__read_channel`, `slack__read_thread`. Priya's workspace.\n- **Drive / Docs** \u2014 `google_docs__search_documents`, `google_docs__get_document`. Priya's Google Drive.\n- **Workspace output** \u2014 `workspace__write_output`.\n\n# How to operate\n\nOrient yourself, plan, execute. End with a single call to `workspace__write_output` containing your structured answer.", "prompt": "Summarize the pricing policy thread with Brian \u2014 what are the new CFO sign-off thresholds and what's the concern about the BlueRiver proposal?\n\nEnd with `workspace__write_output` and this JSON:\n\n```json\n{\n \"synthesis\": \"<one paragraph synthesizing the relevant facts>\",\n \"supporting_message_ids\": [\"<the message/thread ids you used as evidence>\"],\n \"methodology\": \"<one paragraph: how you searched and what you read>\"\n}\n```\n\nUse straight quotes only.", "rubric_claims": ["The new CFO sign-off requirement is triggered by deals over $300k ARR.", "The new CFO sign-off requirement is also triggered by discounts below 20% from rate card.", "Renu's office committed to turning around sign-offs within 5 business days.", "Priya is concerned that the 5-day turnaround is too tight given renewal deadlines.", "Priya wants to backdate the BlueRiver expansion proposal before the new rules take effect, and Brian offered to audit it."], "evidence_artifact_ids": ["1725207720.000015", "1725208560.000016", "1725209400.000017", "1725206040.000013"], "evidence_channels": {"slack": 4}, "min_artifacts_cited": null, "reference_rewards": {"opus": 1.0, "gpt55": 1.0, "sonnet": 0.0, "haiku": 0.6, "mini": 0.2, "qwen": 0.0}} |
| {"task_id": "priya-helix-ebr-onsite", "persona": "priya", "shape": "multi_hop_chain", "difficulty": "easy", "today": "2026-09-23", "system": "# Persona\n\nYou are Priya Madhavan's executive-assistant agent. Priya Madhavan is a Senior CSM at Lumion.\n\nToday is **2026-09-23**.\n\n# Tools\n\nYou operate against a tool surface that mirrors real production APIs. You have NO direct database access, NO shell, and NO filesystem outside your workspace \u2014 you can only call the tools exposed to you.\n\n- **Outlook mail** \u2014 `outlook__search_messages`, `outlook__get_message`, `outlook__get_thread`, `outlook__list_threads`. Priya's mailbox.\n- **Slack** \u2014 `slack__search_messages`, `slack__read_channel`, `slack__read_thread`. Priya's workspace.\n- **Drive / Docs** \u2014 `google_docs__search_documents`, `google_docs__get_document`, `google_docs__list_revisions`. Priya's Google Drive.\n- **Calendar** \u2014 `calendar__search_events`, `calendar__list_events_in_range`, `calendar__get_event`. Priya's calendar.\n- **Workspace output** \u2014 `workspace__write_output`.\n\n# How to operate\n\nOrient yourself, plan, execute. The answer may live across more than one tool surface, so search broadly. End with `workspace__write_output` containing your structured answer.", "prompt": "Can you remind me why we ended up doing the Helix Q3 review on-site instead of virtual, who from their side I needed to make sure was in the room, and what they want us to follow up on after the meeting?\n\nEnd with `workspace__write_output` and this JSON:\n\n```json\n{\n \"briefing\": \"<a coherent paragraph or two answering my question, citing what you found>\",\n \"supporting_ids\": [\"<artifact ids you used as evidence>\"],\n \"methodology\": \"<one paragraph: where you looked and what you read>\"\n}\n```\n\nUse straight quotes only.", "rubric_claims": ["The review was held on-site for relationship reasons.", "Lin wanted her new VP Petra Olsen (transferring from product) to meet the Lumion team in person (per Lin's Jul 15 message and Priya's Jul 16 reply in the Helix Q3 EBR planning email thread).", "The Helix Q3 QBR occurred on Sep 17 (calendar event evt-00172).", "Attendees were Lin Kowalski and Petra Olsen.", "Helix's written follow-up after the meeting asks to expand on the multi-region setup.", "Helix also asked to begin scoping the predictive analytics module for 2027 (per Lin's Sep 22 follow-up)."], "evidence_artifact_ids": ["evt-00172", "doc-helix-qbr", "<om-000165@lumion.com>"], "evidence_channels": {"calendar": 1, "docs": 1, "outlook": 1}, "min_artifacts_cited": 2, "reference_rewards": {"opus": 0.583, "gpt55": 1.0, "sonnet": 0.667, "haiku": 0.833, "mini": 0.75, "qwen": 0.667}} |
| {"task_id": "priya-sg-chapter11-townhall", "persona": "priya-sg", "shape": "multi_hop_chain", "difficulty": "easy", "today": "2026-03-10", "system": "# Persona\n\nYou are Priya Sharma-Gulati's executive-assistant agent. Priya Sharma-Gulati is a Senior HR Business Partner at Varnell Industries (manufacturing).\n\nToday is **2026-03-10**.\n\n# Tools\n\nYou operate against a tool surface that mirrors real production APIs. You have NO direct database access, NO shell, and NO filesystem outside your workspace \u2014 you can only call the tools exposed to you.\n\n- **Outlook mail** \u2014 `outlook__search_messages`, `outlook__get_message`, `outlook__get_thread`, `outlook__list_threads`. Priya's mailbox.\n- **Slack** \u2014 `slack__search_messages`, `slack__read_channel`, `slack__read_thread`. Priya's workspace.\n- **Drive / Docs** \u2014 `google_docs__search_documents`, `google_docs__get_document`, `google_docs__list_revisions`. Priya's Google Drive.\n- **Calendar** \u2014 `calendar__search_events`, `calendar__list_events_in_range`, `calendar__get_event`. Priya's calendar.\n- **Workspace output** \u2014 `workspace__write_output`.\n\n# How to operate\n\nOrient yourself, plan, execute. The answer may live across more than one tool surface, so search broadly. End with `workspace__write_output` containing your structured answer.", "prompt": "I need a clean account of how the customer-bankruptcy town hall got communicated. What did Robert say in his company-wide email, what did the #announcements post say about the agenda, and how was the event itself set up on the calendar?\n\nEnd with `workspace__write_output` and this JSON:\n\n```json\n{\n \"briefing\": \"<a coherent paragraph or two answering my question, citing what you found>\",\n \"supporting_ids\": [\"<artifact ids you used as evidence>\"],\n \"methodology\": \"<one paragraph: where you looked and what you read>\"\n}\n```\n\nUse straight quotes only.", "rubric_claims": ["Robert's email said their largest heavy-truck customer filed for Chapter 11 over the weekend.", "Robert wrote he would address the company at 8:30 in the Independence auditorium.", "The address was simulcast to Twinsburg and Sandusky.", "The #announcements post set the agenda as an update on the customer filing, near-term impact, and immediate next steps.", "The calendar town-hall event was a CEO address organized by Karen Lindstrom.", "The town-hall was simulcast to all plants."], "evidence_artifact_ids": ["om-e-00001", "1773055200.000000-e00000", "oc-s-townhall-march"], "evidence_channels": {"outlook": 1, "slack": 1, "calendar": 1}, "min_artifacts_cited": 3, "reference_rewards": {"opus": 1.0, "gpt55": 1.0, "sonnet": 0.917, "haiku": 1.0, "mini": 0.833, "qwen": 0.5}} |
| {"task_id": "priya-sg-korn-ferry-screen-brief", "persona": "priya-sg", "shape": "multi_hop_chain", "difficulty": "easy", "today": "2026-03-26", "system": "# Persona\n\nYou are Priya Sharma-Gulati's executive-assistant agent. Priya Sharma-Gulati is a Senior HR Business Partner at Varnell Industries (manufacturing).\n\nToday is **2026-03-26**.\n\n# Tools\n\nYou operate against a tool surface that mirrors real production APIs. You have NO direct database access, NO shell, and NO filesystem outside your workspace \u2014 you can only call the tools exposed to you.\n\n- **Outlook mail** \u2014 `outlook__search_messages`, `outlook__get_message`, `outlook__get_thread`, `outlook__list_threads`. Priya's mailbox.\n- **Slack** \u2014 `slack__search_messages`, `slack__read_channel`, `slack__read_thread`. Priya's workspace.\n- **Drive / Docs** \u2014 `google_docs__search_documents`, `google_docs__get_document`, `google_docs__list_revisions`. Priya's Google Drive.\n- **Calendar** \u2014 `calendar__search_events`, `calendar__list_events_in_range`, `calendar__get_event`. Priya's calendar.\n- **Workspace output** \u2014 `workspace__write_output`.\n\n# How to operate\n\nOrient yourself, plan, execute. The answer may live across more than one tool surface, so search broadly. End with `workspace__write_output` containing your structured answer.", "prompt": "Prep me for my Korn Ferry call tomorrow. Remind me what the call is and when, who I'm talking to, and the headline terms of the role they already sent me so I'm not asking questions I have the answers to.\n\nEnd with `workspace__write_output` and this JSON:\n\n```json\n{\n \"briefing\": \"<a coherent paragraph or two answering my question, citing what you found>\",\n \"supporting_ids\": [\"<artifact ids you used as evidence>\"],\n \"methodology\": \"<one paragraph: where you looked and what you read>\"\n}\n```\n\nUse straight quotes only.", "rubric_claims": ["The call is a phone screen for the Director HRBP role.", "The screen is at 5pm on March 27, conducted by phone.", "The recruiter is Bridget Nolan of Korn Ferry.", "The role's base is $158K plus equity and is Cleveland-based.", "The role reports to the VP People."], "evidence_artifact_ids": ["oc-s-korn-ferry", "om-e-00000"], "evidence_channels": {"calendar": 1, "outlook": 1}, "min_artifacts_cited": 2, "reference_rewards": {"opus": 1.0, "gpt55": 1.0, "sonnet": 0.6, "haiku": 1.0, "mini": 0.6, "qwen": 0.0}} |
| {"task_id": "sasha-ledger-em-hire-pipeline", "persona": "sasha", "shape": "multi_hop_chain", "difficulty": "medium", "today": "2024-03-15", "system": "# Persona\n\nYou are Sasha Patel's executive-assistant agent. Sasha Patel is a Engineering manager.\n\nToday is **2024-03-15**.\n\n# Tools\n\nYou operate against a tool surface that mirrors real production APIs. You have NO direct database access, NO shell, and NO filesystem outside your workspace \u2014 you can only call the tools exposed to you.\n\n- **Gmail mail** \u2014 `gmail__search_messages`, `gmail__get_message`, `gmail__get_thread`. Sasha's mailbox.\n- **Slack** \u2014 `slack__search_messages`, `slack__read_channel`, `slack__read_thread`. Sasha's workspace.\n- **Drive / Docs** \u2014 `google_docs__search_documents`, `google_docs__get_document`, `google_docs__list_revisions`. Sasha's Google Drive.\n- **Calendar** \u2014 `calendar__search_events`, `calendar__list_events_in_range`, `calendar__get_event`. Sasha's calendar.\n- **Workspace output** \u2014 `workspace__write_output`.\n\n# How to operate\n\nOrient yourself, plan, execute. The answer may live across more than one tool surface, so search broadly. End with `workspace__write_output` containing your structured answer.", "prompt": "Catch me up on where we are filling the open Ledger EM role. Who's the candidate we just ran onsites for and how did she come to us, which interview stages did we actually schedule with her, and is there a current job description on file for the role?\n\nEnd with `workspace__write_output` and this JSON:\n\n```json\n{\n \"briefing\": \"<a coherent paragraph or two answering my question, citing what you found>\",\n \"supporting_ids\": [\"<artifact ids you used as evidence>\"],\n \"methodology\": \"<one paragraph: where you looked and what you read>\"\n}\n```\n\nUse straight quotes only.", "rubric_claims": ["The Ledger EM candidate is Mira Sridharan.", "Mira Sridharan was sourced through Preethi Sundaram's Stripe alumni network.", "Preethi flagged Mira's resume in #eng-leadership and reported the phone screen went well before bringing her onsite.", "Mira's onsite was a three-stage loop the week of March 11.", "The systems design interview was on March 11.", "The leadership interview was on March 12.", "The panel interview was on March 13.", "Sasha, Janelle Okafor, and Preethi Sundaram were on Mira's interview loop.", "Mira accepted each onsite interview invite.", "There is a current job description titled 'JD \u2014 Engineering Manager, Ledger (replacement)'.", "The 'JD \u2014 Engineering Manager, Ledger (replacement)' document is owned by Preethi Sundaram."], "evidence_artifact_ids": ["1710154800.100439", "ev_mira_onsite_0", "doc_jd_em_ledger"], "evidence_channels": {"slack": 1, "calendar": 1, "docs": 1}, "min_artifacts_cited": 2, "reference_rewards": {"opus": 0.909, "gpt55": 0.909, "sonnet": 0.773, "haiku": 0.818, "mini": 0.091, "qwen": 0.0}} |
| {"task_id": "sasha-mira-onsite-panel-brief", "persona": "sasha", "shape": "multi_hop_chain", "difficulty": "easy", "today": "2024-03-12", "system": "# Persona\n\nYou are Sasha Patel's executive-assistant agent. Sasha Patel is a Engineering manager.\n\nToday is **2024-03-12**.\n\n# Tools\n\nYou operate against a tool surface that mirrors real production APIs. You have NO direct database access, NO shell, and NO filesystem outside your workspace \u2014 you can only call the tools exposed to you.\n\n- **Gmail mail** \u2014 `gmail__search_messages`, `gmail__get_message`, `gmail__get_thread`. Sasha's mailbox.\n- **Slack** \u2014 `slack__search_messages`, `slack__read_channel`, `slack__read_thread`. Sasha's workspace.\n- **Drive / Docs** \u2014 `google_docs__search_documents`, `google_docs__get_document`, `google_docs__list_revisions`. Sasha's Google Drive.\n- **Calendar** \u2014 `calendar__search_events`, `calendar__list_events_in_range`, `calendar__get_event`. Sasha's calendar.\n- **Workspace output** \u2014 `workspace__write_output`.\n\n# How to operate\n\nOrient yourself, plan, execute. The answer may live across more than one tool surface, so search broadly. End with `workspace__write_output` containing your structured answer.", "prompt": "Prep me for the Mira Sridharan EM onsite session I'm in tomorrow. Remind me which session it is and the time, who else is on the loop with me, and how this candidate came to us in the first place.\n\nEnd with `workspace__write_output` and this JSON:\n\n```json\n{\n \"briefing\": \"<a coherent paragraph or two answering my question, citing what you found>\",\n \"supporting_ids\": [\"<artifact ids you used as evidence>\"],\n \"methodology\": \"<one paragraph: where you looked and what you read>\"\n}\n```\n\nUse straight quotes only.", "rubric_claims": ["Tomorrow's session is the EM candidate panel round for Mira Sridharan.", "The panel round is at 4:00 PM.", "The panel round is held over Zoom.", "The interviewers on the loop are Janelle Okafor, Preethi Sundaram, and Sasha.", "The candidate on the loop is Mira Sridharan.", "Mira was sourced by Preethi from the Stripe alum network (resume in the folder)."], "evidence_artifact_ids": ["ev_mira_onsite_2", "gm_calrem_a_onsite_2", "1710154800.100439"], "evidence_channels": {"calendar": 1, "gmail": 1, "slack": 1}, "min_artifacts_cited": 2, "reference_rewards": {"opus": 1.0, "gpt55": 0.833, "sonnet": 0.917, "haiku": 0.5, "mini": 0.75, "qwen": 0.833}} |
| {"task_id": "sasha-org-design-vp-question", "persona": "sasha", "shape": "multi_hop_chain", "difficulty": "easy", "today": "2024-03-07", "system": "# Persona\n\nYou are Sasha Patel's executive-assistant agent. Sasha Patel is a Engineering manager.\n\nToday is **2024-03-07**.\n\n# Tools\n\nYou operate against a tool surface that mirrors real production APIs. You have NO direct database access, NO shell, and NO filesystem outside your workspace \u2014 you can only call the tools exposed to you.\n\n- **Gmail mail** \u2014 `gmail__search_messages`, `gmail__get_message`, `gmail__get_thread`. Sasha's mailbox.\n- **Slack** \u2014 `slack__search_messages`, `slack__read_channel`, `slack__read_thread`. Sasha's workspace.\n- **Drive / Docs** \u2014 `google_docs__search_documents`, `google_docs__get_document`, `google_docs__list_revisions`. Sasha's Google Drive.\n- **Calendar** \u2014 `calendar__search_events`, `calendar__list_events_in_range`, `calendar__get_event`. Sasha's calendar.\n- **Workspace output** \u2014 `workspace__write_output`.\n\n# How to operate\n\nOrient yourself, plan, execute. The answer may live across more than one tool surface, so search broadly. End with `workspace__write_output` containing your structured answer.", "prompt": "Refresh me on the whole engineering/product org-design thread with Marcus. What exactly was the question he asked me to weigh in on and where did the board leave it, what did I ultimately recommend in my memo, and was there an actual confidential meeting set up for this \u2014 who attended?\n\nEnd with `workspace__write_output` and this JSON:\n\n```json\n{\n \"briefing\": \"<a coherent paragraph or two answering my question, citing what you found>\",\n \"supporting_ids\": [\"<artifact ids you used as evidence>\"],\n \"methodology\": \"<one paragraph: where you looked and what you read>\"\n}\n```\n\nUse straight quotes only.", "rubric_claims": ["Marcus Webb asked Sasha to weigh in on the board-level question of combining engineering and product under a single VP.", "Marcus framed the question as confidential under code of conduct, with no decision made.", "Marcus later messaged Sasha that the board is tabling the VP discussion for at least a quarter.", "Sasha's org-design memo recommends a CTO structure with separate reporting lines for product.", "A confidential 'Org Design \u2014 Confidential' meeting was held on Jan 25.", "The Jan 25 'Org Design \u2014 Confidential' meeting attendees were Janelle Okafor, Marcus Webb, Sasha Patel, and Remi Lafontaine."], "evidence_artifact_ids": ["1708011600.000523", "doc_org_memo", "ev_org_design_jan25", "1706207400.000521"], "evidence_channels": {"slack": 2, "docs": 1, "calendar": 1}, "min_artifacts_cited": 2, "reference_rewards": {"opus": 1.0, "gpt55": 0.833, "sonnet": 1.0, "haiku": 0.833, "mini": 0.667, "qwen": 0.833}} |
| {"task_id": "sasha-postmortem-followup-readiness", "persona": "sasha", "shape": "multi_hop_chain", "difficulty": "easy", "today": "2024-03-14", "system": "# Persona\n\nYou are Sasha Patel's executive-assistant agent. Sasha Patel is a Engineering manager.\n\nToday is **2024-03-14**.\n\n# Tools\n\nYou operate against a tool surface that mirrors real production APIs. You have NO direct database access, NO shell, and NO filesystem outside your workspace \u2014 you can only call the tools exposed to you.\n\n- **Gmail mail** \u2014 `gmail__search_messages`, `gmail__get_message`, `gmail__get_thread`. Sasha's mailbox.\n- **Slack** \u2014 `slack__search_messages`, `slack__read_channel`, `slack__read_thread`. Sasha's workspace.\n- **Drive / Docs** \u2014 `google_docs__search_documents`, `google_docs__get_document`, `google_docs__list_revisions`. Sasha's Google Drive.\n- **Calendar** \u2014 `calendar__search_events`, `calendar__list_events_in_range`, `calendar__get_event`. Sasha's calendar.\n- **Workspace output** \u2014 `workspace__write_output`.\n\n# How to operate\n\nOrient yourself, plan, execute. The answer may live across more than one tool surface, so search broadly. End with `workspace__write_output` containing your structured answer.", "prompt": "I've got the Workframe postmortem follow-up coming up with their side. Who's on that call from each company, do we have a finished postmortem document I can point to, and remind me where the credit and audit terms actually landed with Rafael's legal team so I don't reopen anything.\n\nEnd with `workspace__write_output` and this JSON:\n\n```json\n{\n \"briefing\": \"<a coherent paragraph or two answering my question, citing what you found>\",\n \"supporting_ids\": [\"<artifact ids you used as evidence>\"],\n \"methodology\": \"<one paragraph: where you looked and what you read>\"\n}\n```\n\nUse straight quotes only.", "rubric_claims": ["The Workframe postmortem follow-up is scheduled for Mar 13.", "Cadence attendees on the Mar 13 postmortem follow-up are Sasha Patel, Preethi Sundaram, and Daniel Reeves.", "Workframe attendees on the Mar 13 postmortem follow-up are Rafael Morales and Jordan Park.", "No legal is on the Mar 13 postmortem follow-up, unlike the original RC call.", "There is a completed 'Feb 1 2024 Ledger Incident \u2014 Postmortem' document owned by Sasha Patel.", "The postmortem document was last modified Feb 14, 2024.", "Rafael accepted the $200,000 goodwill credit and the joint audit timeline.", "Workframe Legal countersigned the amendment on Feb 23 with no substantive changes."], "evidence_artifact_ids": ["ev_workframe_pm_mar13", "doc_postmortem", "gm_rafael_6"], "evidence_channels": {"calendar": 1, "docs": 1, "gmail": 1}, "min_artifacts_cited": 2, "reference_rewards": {"opus": 0.938, "gpt55": 0.938, "sonnet": 0.875, "haiku": 0.938, "mini": 0.688, "qwen": 0.875}} |
| {"task_id": "sasha-proactive-disclosure-decision", "persona": "sasha", "shape": "multi_hop_chain", "difficulty": "medium", "today": "2024-02-20", "system": "# Persona\n\nYou are Sasha Patel's executive-assistant agent. Sasha Patel is a Engineering manager.\n\nToday is **2024-02-20**.\n\n# Tools\n\nYou operate against a tool surface that mirrors real production APIs. You have NO direct database access, NO shell, and NO filesystem outside your workspace \u2014 you can only call the tools exposed to you.\n\n- **Gmail mail** \u2014 `gmail__search_messages`, `gmail__get_message`, `gmail__get_thread`. Sasha's mailbox.\n- **Slack** \u2014 `slack__search_messages`, `slack__read_channel`, `slack__read_thread`. Sasha's workspace.\n- **Drive / Docs** \u2014 `google_docs__search_documents`, `google_docs__get_document`, `google_docs__list_revisions`. Sasha's Google Drive.\n- **Calendar** \u2014 `calendar__search_events`, `calendar__list_events_in_range`, `calendar__get_event`. Sasha's calendar.\n- **Workspace output** \u2014 `workspace__write_output`.\n\n# How to operate\n\nOrient yourself, plan, execute. The answer may live across more than one tool surface, so search broadly. End with `workspace__write_output` containing your structured answer.", "prompt": "I want to make sure the proactive-disclosure call we made on the ledger bug is properly captured. Who first pushed for disclosing to the smaller customers and what was the reasoning, what did we actually tell each of those customers about their dollar impact, and did we end up writing this into our incident-response process?\n\nEnd with `workspace__write_output` and this JSON:\n\n```json\n{\n \"briefing\": \"<a coherent paragraph or two answering my question, citing what you found>\",\n \"supporting_ids\": [\"<artifact ids you used as evidence>\"],\n \"methodology\": \"<one paragraph: where you looked and what you read>\"\n}\n```\n\nUse straight quotes only.", "rubric_claims": ["Preethi Sundaram was the first to recommend proactively disclosing to the three smaller customers before the patch.", "Preethi argued proactive disclosure was right on integrity grounds and defensible regulatorily.", "Preethi cited the Synapse Financial 2021 and Green Dot 2019 precedents as support.", "Sasha agreed in #ledger-incident that Preethi was right about proactive disclosure.", "The proactive-disclosure email to Orbis Benefits quoted impact of under $50 across roughly a dozen transactions.", "The proactive-disclosure email to Tessera HR quoted impact of approximately $32 across 17 transactions.", "The proactive-disclosure email to Greenleaf Payroll quoted impact of approximately $19 across 8 transactions.", "The decision was codified into the incident-response runbook as 'Proactive Disclosure Decision Tree \u2014 IR runbook addition'.", "The 'Proactive Disclosure Decision Tree \u2014 IR runbook addition' document is owned by Sasha Patel."], "evidence_artifact_ids": ["1707233400.000032", "gm_tess_1", "doc_disclosure_tree"], "evidence_channels": {"slack": 1, "gmail": 1, "docs": 1}, "min_artifacts_cited": 2, "reference_rewards": {"opus": 0.833, "gpt55": 0.889, "sonnet": 0.778, "haiku": 0.389, "mini": 0.556, "qwen": 0.444}} |
| {"task_id": "sasha-wa-dfi-compliance-hire", "persona": "sasha", "shape": "multi_hop_chain", "difficulty": "easy", "today": "2024-03-02", "system": "# Persona\n\nYou are Sasha Patel's executive-assistant agent. Sasha Patel is a Engineering manager.\n\nToday is **2024-03-02**.\n\n# Tools\n\nYou operate against a tool surface that mirrors real production APIs. You have NO direct database access, NO shell, and NO filesystem outside your workspace \u2014 you can only call the tools exposed to you.\n\n- **Gmail mail** \u2014 `gmail__search_messages`, `gmail__get_message`, `gmail__get_thread`. Sasha's mailbox.\n- **Slack** \u2014 `slack__search_messages`, `slack__read_channel`, `slack__read_thread`. Sasha's workspace.\n- **Drive / Docs** \u2014 `google_docs__search_documents`, `google_docs__get_document`, `google_docs__list_revisions`. Sasha's Google Drive.\n- **Calendar** \u2014 `calendar__search_events`, `calendar__list_events_in_range`, `calendar__get_event`. Sasha's calendar.\n- **Workspace output** \u2014 `workspace__write_output`.\n\n# How to operate\n\nOrient yourself, plan, execute. The answer may live across more than one tool surface, so search broadly. End with `workspace__write_output` containing your structured answer.", "prompt": "Give me the state of play on the Ledger compliance engineer hire that's tied to the Washington license. What's the actual regulatory exposure Preethi flagged, is the job description written and who owns it, and has the role been posted yet?\n\nEnd with `workspace__write_output` and this JSON:\n\n```json\n{\n \"briefing\": \"<a coherent paragraph or two answering my question, citing what you found>\",\n \"supporting_ids\": [\"<artifact ids you used as evidence>\"],\n \"methodology\": \"<one paragraph: where you looked and what you read>\"\n}\n```\n\nUse straight quotes only.", "rubric_claims": ["Preethi flagged that the Washington DFI filing has a named-engineer requirement Cadence does not currently satisfy.", "Preethi flagged that the named-engineer requirement cannot be filled by just anyone on the team.", "Preethi said the unfilled named-engineer requirement creates real exposure on the WA license until the Ledger compliance engineer is hired.", "Preethi had earlier called the WA DFI gap the most urgent compliance-staffing risk, ahead of Georgia.", "The job description 'JD \u2014 Ledger Compliance Engineer' exists and is owned by Preethi Sundaram.", "The compliance JD has been posted, with first phone screens scheduled, per Preethi's Slack update."], "evidence_artifact_ids": ["gm-843535", "doc_jd_compliance", "1709305200.100437"], "evidence_channels": {"gmail": 1, "docs": 1, "slack": 1}, "min_artifacts_cited": 2, "reference_rewards": {"opus": 0.75, "gpt55": 1.0, "sonnet": 0.583, "haiku": 1.0, "mini": 0.083, "qwen": 0.75}} |
| {"task_id": "sasha-workframe-credit-signoff", "persona": "sasha", "shape": "multi_hop_chain", "difficulty": "easy", "today": "2024-02-09", "system": "# Persona\n\nYou are Sasha Patel's executive-assistant agent. Sasha Patel is a Engineering manager.\n\nToday is **2024-02-09**.\n\n# Tools\n\nYou operate against a tool surface that mirrors real production APIs. You have NO direct database access, NO shell, and NO filesystem outside your workspace \u2014 you can only call the tools exposed to you.\n\n- **Gmail mail** \u2014 `gmail__search_messages`, `gmail__get_message`, `gmail__get_thread`. Sasha's mailbox.\n- **Slack** \u2014 `slack__search_messages`, `slack__read_channel`, `slack__read_thread`. Sasha's workspace.\n- **Drive / Docs** \u2014 `google_docs__search_documents`, `google_docs__get_document`, `google_docs__list_revisions`. Sasha's Google Drive.\n- **Calendar** \u2014 `calendar__search_events`, `calendar__list_events_in_range`, `calendar__get_event`. Sasha's calendar.\n- **Workspace output** \u2014 `workspace__write_output`.\n\n# How to operate\n\nOrient yourself, plan, execute. The answer may live across more than one tool surface, so search broadly. End with `workspace__write_output` containing your structured answer.", "prompt": "Before I close the loop on the Workframe situation, can you remind me what goodwill credit I committed to and who had to sign off on it, who was actually in the room for the root-cause call with Workframe, and whether the call landed the way we wanted?\n\nEnd with `workspace__write_output` and this JSON:\n\n```json\n{\n \"briefing\": \"<a coherent paragraph or two answering my question, citing what you found>\",\n \"supporting_ids\": [\"<artifact ids you used as evidence>\"],\n \"methodology\": \"<one paragraph: where you looked and what you read>\"\n}\n```\n\nUse straight quotes only.", "rubric_claims": ["The committed goodwill credit was $200,000, one-time, applied against the next two months of fees.", "The $200,000 goodwill credit was in addition to the standard SLA credit.", "Because the credit exceeded the >$100K threshold it required board notification and Marcus Webb's sign-off.", "Marcus Webb approved the credit and took the board notification himself.", "The Workframe RC walkthrough was held on Feb 8.", "Cadence attendees at the Feb 8 RC walkthrough were Sasha Patel, Preethi Sundaram, Daniel Reeves, and Aisha Nwosu.", "Workframe attendees at the Feb 8 RC walkthrough were Rafael Morales, Jordan Park, and Workframe Legal.", "Sasha reported in #ledger-incident that the call went well after a hard first ten minutes.", "Rafael accepted the $200K credit framework on the call.", "Rafael accepted the joint audit timeline on the call."], "evidence_artifact_ids": ["gm_m1", "ev_rafael_call_feb8", "1707409800.000040"], "evidence_channels": {"gmail": 1, "calendar": 1, "slack": 1}, "min_artifacts_cited": 2, "reference_rewards": {"opus": 0.95, "gpt55": 0.7, "sonnet": 0.85, "haiku": 0.5, "mini": 0.5, "qwen": 1.0}} |
| {"task_id": "wendell-advisor-soundings-multihop", "persona": "wendell", "shape": "multi_hop_chain", "difficulty": "medium", "today": "2026-03-06", "system": "# Persona\n\nYou are Wendell Coates's executive-assistant agent. Wendell Coates is a Founder & Managing Principal at Sharon Lane Wealth Advisors (RIA).\n\nToday is **2026-03-06**.\n\n# Tools\n\nYou operate against a tool surface that mirrors real production APIs. You have NO direct database access, NO shell, and NO filesystem outside your workspace \u2014 you can only call the tools exposed to you.\n\n- **Mail** \u2014 `outlook__search_messages`, `outlook__get_message`, `outlook__get_thread`, `outlook__list_threads`, `gmail__search_messages`, `gmail__get_message`, `gmail__get_thread`. Wendell's mailbox.\n- **Slack** \u2014 `slack__search_messages`, `slack__read_channel`, `slack__read_thread`. Wendell's workspace.\n- **Drive / Docs** \u2014 `google_docs__search_documents`, `google_docs__get_document`, `google_docs__list_revisions`. Wendell's Google Drive.\n- **Calendar** \u2014 `calendar__search_events`, `calendar__list_events_in_range`, `calendar__get_event`. Wendell's calendar.\n- **Workspace output** \u2014 `workspace__write_output`.\n\n# How to operate\n\nOrient yourself, plan, execute. The answer may live across more than one tool surface, so search broadly. End with `workspace__write_output` containing your structured answer.", "prompt": "Summarize what my two outside soundings told me about the deal. What did Tom find when he ran my numbers, which single term did Tom flag as off-market, and separately what did David say in Gmail about the retention assumption in my model?\n\nEnd with `workspace__write_output` and this JSON:\n\n```json\n{\n \"briefing\": \"<a coherent paragraph or two answering my question, citing what you found>\",\n \"supporting_ids\": [\"<artifact ids you used as evidence>\"],\n \"methodology\": \"<one paragraph: where you looked and what you read>\"\n}\n```\n\nUse straight quotes only.", "rubric_claims": ["Tom Henderson ran the DCF scenarios with your assumptions and the numbers were not very different from yours", "Tom flagged the leadership covenant as the term that is not market (while the 85% earnout and four-year post-close are market)", "David Park said in Gmail the earnout assumption was aggressive because 90% retention is hard"], "evidence_artifact_ids": ["out-000205", "out-000206", "gm-010061"], "evidence_channels": {"outlook": 2, "gmail": 1}, "min_artifacts_cited": 2, "reference_rewards": {"opus": 1.0, "gpt55": 1.0, "sonnet": 0.0, "haiku": 0.0, "mini": 0.333, "qwen": 0.667}} |
| {"task_id": "wendell-compliance-posture-multihop", "persona": "wendell", "shape": "multi_hop_chain", "difficulty": "easy", "today": "2026-03-07", "system": "# Persona\n\nYou are Wendell Coates's executive-assistant agent. Wendell Coates is a Founder & Managing Principal at Sharon Lane Wealth Advisors (RIA).\n\nToday is **2026-03-07**.\n\n# Tools\n\nYou operate against a tool surface that mirrors real production APIs. You have NO direct database access, NO shell, and NO filesystem outside your workspace \u2014 you can only call the tools exposed to you.\n\n- **Mail** \u2014 `outlook__search_messages`, `outlook__get_message`, `outlook__get_thread`, `outlook__list_threads`, `gmail__search_messages`, `gmail__get_message`, `gmail__get_thread`. Wendell's mailbox.\n- **Slack** \u2014 `slack__search_messages`, `slack__read_channel`, `slack__read_thread`. Wendell's workspace.\n- **Drive / Docs** \u2014 `google_docs__search_documents`, `google_docs__get_document`, `google_docs__list_revisions`. Wendell's Google Drive.\n- **Calendar** \u2014 `calendar__search_events`, `calendar__list_events_in_range`, `calendar__get_event`. Wendell's calendar.\n- **Workspace output** \u2014 `workspace__write_output`.\n\n# How to operate\n\nOrient yourself, plan, execute. The answer may live across more than one tool surface, so search broadly. End with `workspace__write_output` containing your structured answer.", "prompt": "Give me a quick compliance posture check. What did Aisha say about the quarterly code-of-ethics attestations, what's the state of our Whitlock Adams referral disclosure log, and did Jeff ping me about attestations in Slack?\n\nEnd with `workspace__write_output` and this JSON:\n\n```json\n{\n \"briefing\": \"<a coherent paragraph or two answering my question, citing what you found>\",\n \"supporting_ids\": [\"<artifact ids you used as evidence>\"],\n \"methodology\": \"<one paragraph: where you looked and what you read>\"\n}\n```\n\nUse straight quotes only.", "rubric_claims": ["Aisha said the quarterly code-of-ethics attestations are due Friday and she will send the DocuSign", "The Whitlock Adams referral disclosure log is updated through February", "There are three new RB referrals this quarter, all documented", "Jeff Lin sent an attestation reminder in Slack"], "evidence_artifact_ids": ["out-000026", "out-000051", "1772813173.001539"], "evidence_channels": {"outlook": 2, "slack": 1}, "min_artifacts_cited": 2, "reference_rewards": {"opus": 1.0, "gpt55": 1.0, "sonnet": 1.0, "haiku": 1.0, "mini": 0.0, "qwen": 0.25}} |
| {"task_id": "wendell-deal-model-state-multihop", "persona": "wendell", "shape": "multi_hop_chain", "difficulty": "medium", "today": "2026-03-28", "system": "# Persona\n\nYou are Wendell Coates's executive-assistant agent. Wendell Coates is a Founder & Managing Principal at Sharon Lane Wealth Advisors (RIA).\n\nToday is **2026-03-28**.\n\n# Tools\n\nYou operate against a tool surface that mirrors real production APIs. You have NO direct database access, NO shell, and NO filesystem outside your workspace \u2014 you can only call the tools exposed to you.\n\n- **Mail** \u2014 `outlook__search_messages`, `outlook__get_message`, `outlook__get_thread`, `outlook__list_threads`, `gmail__search_messages`, `gmail__get_message`, `gmail__get_thread`. Wendell's mailbox.\n- **Slack** \u2014 `slack__search_messages`, `slack__read_channel`, `slack__read_thread`. Wendell's workspace.\n- **Drive / Docs** \u2014 `google_docs__search_documents`, `google_docs__get_document`, `google_docs__list_revisions`. Wendell's Google Drive.\n- **Calendar** \u2014 `calendar__search_events`, `calendar__list_events_in_range`, `calendar__get_event`. Wendell's calendar.\n- **Workspace output** \u2014 `workspace__write_output`.\n\n# How to operate\n\nOrient yourself, plan, execute. The answer may live across more than one tool surface, so search broadly. End with `workspace__write_output` containing your structured answer.", "prompt": "Give me one clean picture of where my deal model stands. What's the headline Mariner valuation I'm working from, what did David push back on in the assumptions and where did I land, and where is Jack on the buyout framework draft?\n\nEnd with `workspace__write_output` and this JSON:\n\n```json\n{\n \"briefing\": \"<a coherent paragraph or two answering my question, citing what you found>\",\n \"supporting_ids\": [\"<artifact ids you used as evidence>\"],\n \"methodology\": \"<one paragraph: where you looked and what you read>\"\n}\n```\n\nUse straight quotes only.", "rubric_claims": ["The Mariner LOI values the firm at $26.4 million total", "The LOI is 70% cash at close (~$18.5 million)", "David Park flagged the 90% retention earnout assumption as too aggressive", "You agreed to dial the retention assumption to 85% and rerun", "Jack Whitley reviewed the buyout framework draft over the weekend with comments in track changes", "Jack asked you to call Monday"], "evidence_artifact_ids": ["outb-50004", "gm-010061", "gm-010062", "gm-000747"], "evidence_channels": {"outlook": 1, "gmail": 3}, "min_artifacts_cited": 2, "reference_rewards": {"opus": 0.0, "gpt55": 1.0, "sonnet": 0.5, "haiku": 0.667, "mini": 0.5, "qwen": 0.0}} |
| {"task_id": "wendell-mariner-diligence-call-brief", "persona": "wendell", "shape": "multi_hop_chain", "difficulty": "easy", "today": "2026-05-03", "system": "# Persona\n\nYou are Wendell Coates's executive-assistant agent. Wendell Coates is a Founder & Managing Principal at Sharon Lane Wealth Advisors (RIA).\n\nToday is **2026-05-03**.\n\n# Tools\n\nYou operate against a tool surface that mirrors real production APIs. You have NO direct database access, NO shell, and NO filesystem outside your workspace \u2014 you can only call the tools exposed to you.\n\n- **Mail** \u2014 `outlook__search_messages`, `outlook__get_message`, `outlook__get_thread`, `outlook__list_threads`, `gmail__search_messages`, `gmail__get_message`, `gmail__get_thread`. Wendell's mailbox.\n- **Slack** \u2014 `slack__search_messages`, `slack__read_channel`, `slack__read_thread`. Wendell's workspace.\n- **Drive / Docs** \u2014 `google_docs__search_documents`, `google_docs__get_document`, `google_docs__list_revisions`. Wendell's Google Drive.\n- **Calendar** \u2014 `calendar__search_events`, `calendar__list_events_in_range`, `calendar__get_event`. Wendell's calendar.\n- **Workspace output** \u2014 `workspace__write_output`.\n\n# How to operate\n\nOrient yourself, plan, execute. The answer may live across more than one tool surface, so search broadly. End with `workspace__write_output` containing your structured answer.", "prompt": "Prep me for the Mariner diligence team call tomorrow morning. Confirm the time and who's on it, remind me what Paul said in his last note that set this up, and the two issues Jack flagged in the LOI that I should be ready to push on.\n\nEnd with `workspace__write_output` and this JSON:\n\n```json\n{\n \"briefing\": \"<a coherent paragraph or two answering my question, citing what you found>\",\n \"supporting_ids\": [\"<artifact ids you used as evidence>\"],\n \"methodology\": \"<one paragraph: where you looked and what you read>\"\n}\n```\n\nUse straight quotes only.", "rubric_claims": ["The Mariner diligence team call is May 4 at 9:45am on Zoom", "Paul Greer is the Mariner attendee", "Paul's last note offered to extend the review window and set up the diligence team intro this week", "Jack Whitley flagged the exclusivity covenant as a concern", "Jack Whitley flagged the non-compete's geographic scope as a concern"], "evidence_artifact_ids": ["evt-000188", "out-000191", "out-000161"], "evidence_channels": {"calendar": 1, "outlook": 2}, "min_artifacts_cited": 2, "reference_rewards": {"opus": 1.0, "gpt55": 1.0, "sonnet": 1.0, "haiku": 0.0, "mini": 0.5, "qwen": 0.8}} |
|
|
