Skip to main content
RigorDesk logoRigorDesk
Start free

Oxford Union

The classic formal debate. Models split into proposition and opposition sides, each gives an opening and a closing, and the chairman declares a winner. The format preserves tension. The verdict does not paper over disagreement.

Models: 2 to 6, even count (1 to 3 per side) · Tone: formal debate · Best for: genuinely contested questions, policy, ethics, contested claims.

When to use

Oxford Union is the right format when the question is a real "this house believes…" motion with two legitimate sides. The output is not a synthesized "best answer". It is a verdict that picks a side and explains why. RigorDesk preserves the strongest version of the other side as dissent.

  • Policy debates
  • Ethics and contested social questions
  • "Should we adopt X?" where reasonable people disagree
  • Any case where the answer genuinely is "it depends" and you need a defensible pick

If you just want pressure-testing of an idea, use Devil's Advocate. If you have one position and want it probed, use Socratic.

How it works

  1. Opening. Proposition and opposition each deliver an opening argument in alternation.
  2. Rebuttal. Each side responds to the other. The format driver controls this step, and you cannot configure it.
  3. Closing. Each side delivers a closing statement. It adds synthesis only, no new arguments.
  4. Verdict. The chairman reads the full transcript, declares a winner, and explains why. The chairman preserves the losing side's strongest argument as the "remaining tension" section.

The verdict's job is to pick a side when the evidence warrants it, not to find a center.

Sides configuration

formatMeta.sides assigns models to sides. The number of models on each side must be equal, so use 1v1, 2v2, or 3v3.

The format will reject uneven side counts. A 3-model Oxford Union assembly is invalid.

Request body

format: "OXFORD_UNION", with formatMeta.sides declaring the two sides.

curl -X POST https://rigordesk.com/api/v1/assemblies \
 -H "Authorization: Bearer $DBT_KEY" \
 -H "Content-Type: application/json" \
 -d '{
 "name": "AI Copyright Debate",
 "format": "OXFORD_UNION",
 "maxRounds": 4,
 "agreementLimit": 70,
 "models": [
 { "id": "openai/gpt-5.5", "role": "PROPOSITION" },
 { "id": "anthropic/claude-opus-4.8", "role": "PROPOSITION" },
 { "id": "x-ai/grok-4.3", "role": "OPPOSITION" },
 { "id": "google/gemini-3.1-pro-preview", "role": "OPPOSITION" }
 ],
 "formatMeta": {
 "sides": {
 "proposition": [
 "openai/gpt-5.5",
 "anthropic/claude-opus-4.8"
 ],
 "opposition": [
 "x-ai/grok-4.3",
 "google/gemini-3.1-pro-preview"
 ]
 }
 }
 }'

Then run a debate against it. The motion is your prompt:

curl -X POST https://rigordesk.com/api/v1/debates \
 -H "Authorization: Bearer $DBT_KEY" \
 -H "Content-Type: application/json" \
 -H "Idempotency-Key: $(uuidgen)" \
 -d '{
 "prompt": "This house believes AI-generated code should be copyrightable.",
 "assemblyId": "PUT_ASSEMBLY_ID_HERE"
 }'

Example prompt

"This house believes that remote work has been a net negative for engineering culture and should be ended at most companies."

What the verdict looks like

{
 "finalAnswer": "## Verdict: Proposition wins, narrowly.

The proposition makes the stronger empirical case: team-cohesion metrics, mentor-protégé relationships, and the cultural transmission of taste all measurably degrade at >60% remote participation. The opposition counters with retention numbers and individual deep-work output, both real, but neither answers the cohesion argument.

The opposition's strongest surviving argument, and the reason this is a narrow win, is that the proposition's evidence is dominated by companies that *also* underinvested in remote-work infrastructure. Companies that invested heavily (proper onboarding, intentional offsites, structured async writing) saw no degradation. So the proposition's claim is true *for the median employer* but not for the well-resourced one.

### What this verdict does NOT say

It does not say \"remote work is bad.\" It says: at most companies, run as most companies run remote today, the cultural costs are real and the productivity gains don't offset them.

### Remaining tension (the opposition's case)

For engineering orgs with senior IC density, explicit documentation culture, and intentional on-sites, the opposition's case is genuinely strong. The verdict would flip if the prompt were \"should well-resourced companies end remote work?\" instead.
",
 "totalCost": 0.71,
 "totalTokens": 19400
}

Tuning notes

  • maxRounds: 4 is the canonical Oxford Union structure, opening plus rebuttal plus closing. Don't change it.
  • agreementLimit: 70. Oxford Union is meant to expose disagreement. Don't set it too high or the debate will run all rounds without converging.
  • 2v2 is the sweet spot. 1v1 is too thin. 3v3 starts to feel like a committee.
  • Pair models with different training distributions for the most interesting cross-examination. Putting two Claude models on the same side produces less genuine disagreement than a Claude/GPT/Gemini mix.