Review formats
Start with Refinement for general feedback on a draft. The other formats change what reviewers do, when they see each other's responses, and how the review ends.
Quick comparison
| Format | Models | Best for | Tone |
|---|---|---|---|
| Refinement | 1 to 8 | Default. Drafts, essays, recommendations, decisions with a single best answer. | Collaborative |
| Devil's Advocate | 1 to 8 | Stress-testing an idea you already have. Plans, claims, recommendations. | Adversarial |
| Oxford Union | 2 to 6 (even) | Genuine for/against questions. Policy, ethics, contested claims. | Formal debate |
| Socratic | 2 | Deep probing of a single position. Definitions, edge cases, claims you suspect are fragile. | Cross-examination |
| Red Team | 3 (exactly) | Security, prompt injection, code/spec audit. Attacker / defender / judge. | Attack + defense |
| Delphi Panel | 3 to 8 | Forecasting, estimates, multi-perspective synthesis where anchoring bias is a risk. | Anonymous R1, revealed R2 |
| Steel-Manning | 2 to 6 | Honestly engaging with an opposing position before rebutting it. | Charitable adversary |
How to pick
A simple decision tree:
- Have a claim to test? → Devil's Advocate or Red Team
- Have a for-and-against question? → Oxford Union
- Have a single position you want to probe deeply? → Socratic
- Need a forecast or estimate? → Delphi Panel
- Need to honestly engage the strongest opposing view? → Steel-Manning
- Want feedback and improvements on a draft? → Refinement (default)
Format-specific configuration
Most formats work with defaults. A few need formatMeta in the assembly body to assign roles or sides:
| Format | Required formatMeta |
|---|---|
| Refinement | none |
| Devil's Advocate | none (uses mode: "DEVILS_ADVOCATE") |
| Oxford Union | { "sides": { "proposition": ["modelA", "modelB"], "opposition": ["modelC"] } } |
| Socratic | { "roles": { "interrogator": "modelA", "respondent": "modelB" } } |
| Red Team | { "roles": { "attacker": "modelA", "defender": "modelB", "judge": "modelC" } } |
| Delphi Panel | none |
| Steel-Manning | none (uses the model's own position as the steel-man target) |
Each format page below documents its full request body and example.
What every format produces
A completed review brings together:
- A conclusion,
finalAnswer, the summarizing model's response to your prompt. - A per-claim confidence breakdown, where the model supports it.
- A dissent preserved section, the strongest surviving counter-argument, with model attribution.
- A conviction curve, the panel's agreement score over rounds.
- The full transcript, every model's response in order, with per-turn cost and latency.
The format changes how RigorDesk organizes the panel. Available details depend on the format and the responses produced.
Format pages
- Refinement, collaborative, default
- Devil's Advocate, adversarial
- Oxford Union, proposition vs opposition
- Socratic, cross-examination
- Red Team, attacker / defender / judge
- Delphi Panel, anonymous first round
- Steel-Manning, charitable adversary