Research

How the research is made

Every round behind Corpus Cannabis is worked by three AI systems set against each other. ChatGPT and Gemini research the same frozen question independently, and neither sees the other’s report until both are filed. Claude then cross-examines both, checks every load-bearing source and writes the synthesis. After sign-off the models audit each other’s work. The friction is the point: it surfaces hallucinations (invented citations, real papers that don’t say what was claimed, numbers that drift) before anything reaches this site.

The main goal is the subject itself: cultivation and cannabis science, worked carefully enough to stand behind, and shared in public as whole rounds rather than just conclusions. A third interest runs alongside: a standing record of how well AI agents do deep research, and the kinds of mistakes they make.

How a round runs

1. 1

   Brief

   • Claude drafts
   • ChatGPT reviews
   • Gemini reviews
   • Danny freezes

   One question, the rules of evidence and 15–30 testable candidate claims. Frozen before either researcher starts.

   → brief
2. 2

   Independent research

   • ChatGPT deep research
   • Gemini deep research

   Both run the same frozen brief. Neither sees the other’s pass until both are filed. Filed verbatim, errors included.

   → raw passes
3. 3

   Verify and critique

   • Claude verifies

   Every load-bearing source gets two checks: does the citation resolve to one real paper, and does that paper say what was claimed?

   → critique
4. 4

   Correction round

   • ChatGPT answers
   • Gemini answers
   • Claude re-checks

   Researchers answer the critique. A replacement citation is unverified until it passes the same two checks.

   → corrections
5. 5

   Ledger and synthesis

   • Claude writes
   • Danny signs

   Each surviving claim gets a ledger entry with its evidence and verdict. The synthesis cites only ledger claims.

   → ledger · synthesis
6. 6

   Cross-model audit

   • ChatGPT audits
   • Gemini audits
   • Claude confirms

   A model that did not edit the report traces sampled claims back to the source. Findings are verified, then patched in the open.

   → audit

Before research starts, Claude also compares the topic with the author’s earlier teaching material, so old claims can be tested rather than repeated; that archive check stays private. Danny, the research director, picks topics, settles disputes and signs every synthesis.

Rules every round follows

• A citation that doesn’t resolve to one real, matching paper makes its claim unsourced.
• A paper that supports the direction of a finding doesn’t automatically support its size, population or causal wording.
• A corrected citation is unverified until it passes the same checks as the original.
• “Unknown” is an acceptable verdict, and “the myth is unsupported” is not the same as “the opposite is true”.
• Disagreement between researchers is recorded as content, not smoothed over.
• No single round’s scorecard ranks the models; the next round could reverse it.