MED01-P1 — Kickoff prompt

Status: v3 · 2026-09-23 · APPROVED FOR LAUNCH — ChatGPT line-edit applied (4 wording edits), Gemini approved without change. P1 brief frozen 2026-09-20 (v1.0). Git only; not mirrored to Drive. v2 → v3 (2026-09-23, ChatGPT): older-source identifier rule (many 1960–1999 papers lack DOIs); definition of “verified live”; bibliometric-count fallback wording; output-limit contingency. v1 → v2: context-isolation paragraph (ChatGPT); brief delivered as an attached file in a clean session instead of through the Drive connector; boundary rule, retrospective-source rule, and single-submission rule restated.

How to launch (Danny)

1. Same prompt, word for word, to both researchers (§M1). Any edit to the prompt applies to both or to neither.
2. Clean session for each. A new chat inside the existing ChatGPT project still inherits project context and memory — including landmark lists that were deliberately kept out of the brief. Use a session that does not inherit it:
   • ChatGPT: a chat outside the project, as a Temporary Chat or with memory / “reference chat history” turned off.
   • Gemini: a new chat with saved info / personal context off; not a Gem built for this project.
3. Attach two files to the first message rather than pointing at Drive: medical/MED01/brief.md (frozen v1.0) and medical/PROTOCOL-MEDICAL.md (v0.5). This removes the need for folder access, so the researcher cannot wander into other MED01 files, and it works in sessions where the Drive connector is off.
4. Web search / browsing on.
5. One complete submission per researcher. If a reply is cut off, ask for a continuation in the same chat; that continuation is part of the same submission.
6. Firewall: the first pass to arrive goes to Claude and into git only (raw-inputs/<model>/) — not to Drive, and not into the other researcher’s chat — until the second is filed. Then both are mirrored together, and the archive audit and editorial dispositions may follow.

The prompt (identical for ChatGPT and Gemini)

You are one of two independent researchers on MED01-P1, a historical evidence map of medical cannabis and cannabinoid research, 1960–1999. Another AI researcher is working from the same frozen brief. You will not see each other's work, and you should not try to anticipate it.

YOUR TASK SPECIFICATION is the attached file: "MED01-P1 Brief v1.0 FROZEN 2026-09-20". The second attachment, "PROTOCOL-MEDICAL v0.5", is the rulebook the brief refers to. Read the whole brief before you begin. It is authoritative: where this message and the brief differ, the brief wins.

CONTEXT ISOLATION. Treat the attached frozen brief as the sole substantive specification for this research pass. Do not rely on, reconstruct, or import conclusions from any prior conversation about MED01, the author and researcher's previous writing, archive material, editorial reviews of this brief, expected historical milestones, or another model's work. If any such prior context is visible to you, disregard it when selecting findings and sources, and say at the top of your pass that it was visible. Do not open or request any other project file. Build the evidence map independently, from the literature and the frozen brief.

Key points, all from the brief:
- This is an EVIDENCE MAP of a literature, not an efficacy review. No present-day efficacy or safety verdicts, and no advice.
- Window: primary sources published 1960-01-01 to 1999-12-31. Each publication belongs strictly to the part containing its publication year; for milestones begun in the window but resolved later, report the documented state as of 1999-12-31 and flag the rest for the next part.
- Retrospective sources (later histories, reviews, bibliometric studies about the period) are welcome, searched through the MED01-wide cutoff of 2026-09-20, and must be marked "retrospective". They may establish chronology and locate sources; what was KNOWN by 1999 must trace to evidence available by 1999. Report the contemporaneous state and any later reinterpretation separately.
- Build the A–I map from the literature FIRST. Only then answer the archive-derived questions (§6) and the public claims (§7). Those are audit probes — not a sampling frame, and not a guide to what matters or to which countries and topics deserve weight.
- Two tiers of study description: full extraction for every study behind a candidate claim and for keystone studies; one compact citation line for the wider landscape.
- At most 30 candidate claims (CAND-01 …).
- Verify sources live wherever you can. Every reference is marked "verified live" or "from memory — unverified", and "period" or "retrospective". Give a durable identifier where one exists — DOI, PMID/PMCID, trial registration, government/document number, or stable institutional URL. If no durable identifier exists, give complete bibliographic details sufficient to locate the source. Never invent or complete an identifier from memory; a wrong identifier is worse than no identifier.
- "Verified live" means both citation identity and claim support were checked: the source exists with the stated metadata, and the accessible source supports the statement attributed to it. If access prevents full support verification, state the access limitation explicitly.
- For publication counts by decade, use published bibliometric analyses where available, plus ONE reproducible database query of your own (database, exact search string, date run, result count). Never estimate, interpolate, or smooth a count. State indexing and coverage limitations explicitly.
- "None identified in my searches" is never "none exists."
- Finish with your self-flagged weak points and an exact search log: tools and databases actually used, query strings, dates run, and what you could not access.

If a platform output limit prevents delivery in one message, continue immediately in consecutively labeled parts without waiting for user input. Together those messages constitute one frozen submission. Do not revise earlier parts while continuing.

Deliver ONE complete submission, as text in this chat, in the output format of brief §5. State at the top which brief version you worked from. Once submitted, your pass is frozen as received; you will not see the other researcher's pass until both are filed.