Canon · Agnostic storm-review
Cross-review deliberation — map expert opinion instead of picking a winner
A multi-model method used when the goal is to map the space of expert opinion, not to crown a best answer.
How it differs from a tournament panel
| Tournament panel | Cross-review deliberation | |
|---|---|---|
| Round 1 | same prompt, answers anonymised | full standalone analyses, authors named |
| Scoring | graders rank, a winner is picked | no grading, no winner |
| Round 2 | none | each model reviews the others' actual texts |
| Output | ranked consensus, dissent discarded | agree/disagree map, dissent kept with reasons |
The three rounds
- Independent analysis. The same brief goes to each model (for example Gemini, ChatGPT and DeepSeek, with Claude as a fourth voice). Each writes a complete analysis without seeing the others.
- Cross-review by name. Each model gets the others' texts, labelled by author. For each major claim: where do you agree or disagree, and why? Then: the strongest point you missed, the biggest unaddressed risk, and where the real disagreement lies. Don't merge into consensus.
- A synthesis that keeps the dissent. A map, not an average: genuine consensus, live disputes with both sides' reasoning, new risks raised in round 2, and a recommended path. It also calls out where cross-review changed someone's opinion, which is the sign the deliberation worked.
What it buys
Opinions really do move under cross-examination. Round 2 raises issues nobody reached in round 1 — in the first run: a model-inversion attack, legal liability for a black box, and a prompt-leak paradox.
Never accept a half answer
Long answers can be cut off. The method detects truncation and asks the model to continue exactly where it stopped. It stitches the pieces together until the answer is complete.