Adversarial Debate Prep - A Learning Method

Primary guide and scope

Use My Learning Process for the general learning process. This is the optional advanced method for a contested topic once I can explain a position and want to test it against serious opposition.
[!speech] Practice layer added 2026-08-02
The retrieval-failure diagnosis below now has a template: Speaking Reps · Speaking Rep — closed-book 2-minute reps against the position cards.

⬅️ Learning · Templates: Debate Project - Master · Debate Project - REF Note · AI Sparring Session · Worked examples: Genocide Debate - Master · Secular Democracy vs Theocracy

The method in one sentence

Pick a contested topic, build the opponent’s case first and in its strongest form, then build your answers as a numbered tree, then spar against an AI playing the opponent until the tree stops breaking — and use the breakages, not the wins, as the study signal.

Why this works, briefly

It isn’t just debate practice. It stacks four of the best-evidenced learning mechanisms at once:

• Argument mapping — structuring claims and objections as an explicit tree. In Tim van Gelder’s meta-analysis, high-intensity argument-mapping courses moved critical-thinking scores ~0.8 SD, against ~0.34 SD for ordinary critical-thinking courses and ~0.1 SD for a semester of normal study. It is the single best-supported critical-thinking intervention there is — and your C / R / O / RR notation already is one.
• Retrieval practice + the generation effect — you produce answers rather than re-read them.
• Desirable difficulty — sparring is deliberately harder than reviewing, which is why it encodes better.
• Deliberate practice with feedback — each AI round returns a specific, correctable failure.

The corollary is the part people skip: re-reading your own notes is close to worthless. The learning happens when the tree is closed and you have to rebuild it under pressure.

---

Worked practice case — morality and religion

Does Morality Require Religion - Master compresses existing research into four retrieval cues and C/R/O/RR branches. Morality and Religion - Practice supplies a ten-minute first session, hidden-answer cards, and one-objection-at-a-time sparring. REF - Objective Morality and God holds the deeper material. Start with a closed-note rep; expand research when a specific failure calls for it.

🔄 The Loop

0. Frame  →  1. Opposition-first research  →  2. Steelman  →  3. Build the tree
        →  4. Evidence briefs  →  5. AI sparring  →  6. Repair  →  7. Real debate  →  8. Post-mortem
                                        ↑___________________|
                                        (repeat 5–6 until stable)

0 · Frame the resolution before researching

Write the resolution as a proposition someone could deny, then answer three questions before opening a single source:

• Who carries the burden of proof, and on which specific elements?
• What would change my mind? (Write it down now, while you’re still honest. If nothing would, stop — you’re not researching, you’re collecting.)
• What am I not obliged to defend? Half of debate losses are accepting burdens nobody assigned you.

Your Genocide project does this well: the operative words are “intent to destroy … as such”, so every pillar has to clear the intent element, not merely the harm element. That single sentence determines what the rest of the research is for. → REF - Genocide Definitions and Key Terms

1 · Research the opposition first, and give it half your time

The rule worth adopting from competitive debate

Policy debaters treat roughly half of all prep as opposition prep — the standard failure mode is a beautiful affirmative case and no answers on the negative. Most self-directed research inverts this: 90% confirming, 10% skimming critics.

Do the opponent’s side first, before your own. Two reasons: you can’t steelman a position you’ve already spent a week arguing against, and you’ll discover which of your own planned arguments are already dead before you invest in them.

Sources, roughly in order of value: the opponent’s own words (their books, talks, prior debates — not summaries of them) → the strongest scholarly version of their case → primary documents both sides cite → then commentary. Use AI for map-making (who are the serious people, what’s the standard taxonomy of arguments) rather than for facts.

The citation trap

Never move an AI-supplied citation into a REF note without opening the primary source. Fabricated and mis-attributed references are the single most expensive failure mode in this workflow — one bad cite discredits an otherwise sound tree in front of an audience. This applies to page numbers and quoted wording, not just to whether a source exists.

2 · Steelman before you rebut

Gate: you may not write a rebuttal until the claim above it is stated in a form the opponent would endorse.

Rapoport’s rules are the working standard — re-express the position so clearly they’d say “I wish I’d put it that way”, list what you’ve learned from it, and name your points of agreement, then rebut. In practice this pays twice: it makes your rebuttals land on the real position rather than a caricature, and audiences trust someone who has visibly understood the other side.

Your Zionism debate note is the model — 27 arguments each stated at full strength before any answer.

3 · Build the tree with fixed notation

Use one notation and never deviate. Yours:

The numbering carries the structure, so the outline panel renders the whole argument as a tree. Two additions worth making:

• Mark each node’s strength — [strong] / [contested] / [weak — don't lead with this]. You will forget which is which under pressure.
• Depth-3 rule: any C where you can’t get to RR has a hole. Depth is a better completeness signal than word count.

4 · Evidence briefs that answer arguments

The distinction that makes REF notes useful

A brief is not an evidence category — a brief answers an argument. An evidence dump forces you to reason from raw material live, which is exactly what you can’t do under time pressure.

You already solved this: every block in REF - Genocide Definitions and Key Terms ends with a “Debate anchor” — one line saying what this evidence does in the argument. Make that mandatory. Any REF entry without an anchor is unprocessed research, not a brief.

Keep the Master / REF split: the Master is the live script you hold during the debate; REF notes hold the dense data and full source lists so they don’t bury the talking points. → Genocide Debate - Inbox for unprocessed captures.

5 · AI sparring — the distinctive step

This is the part with no established playbook, so the protocol matters more than elsewhere. Full prompts in AI Sparring Session; the principles:

The four failure modes, and the fix for each

Failure	What it looks like	Fix
Sycophancy	The AI concedes as soon as you push back, and you feel great	Instruct explicitly: “Never concede a point unless I have actually refuted it. If my answer is weak, say so and press harder.”
Softening	It argues the reasonable-centrist version, not the version you’ll actually face	Name a real opponent and their actual commitments, or paste their own words as the persona brief
Hallucinated evidence	It invents a study that supports its side	”Cite only sources you are confident exist. If you can’t source a claim, mark it [UNSOURCED] and argue it analytically.”
Context contamination	You’ve argued yourself into agreement with it over 40 messages	Start each round in a fresh context. Paste only the opponent brief and the C-list — never your rebuttals

Run several personas across rounds, because they break different things:

1. Hostile expert — knows the literature better than you; attacks your evidence.
2. Rhetorical opponent — a skilled bad-faith debater; attacks your framing, reframes your concessions as admissions, uses the Gish gallop.
3. Sympathetic skeptic — agrees with your values, doubts your conclusion; attacks your inferences. Usually the most productive.
4. The audience — asks the naive question you can’t answer in one sentence. Exposes where you’ve learned jargon instead of understanding.

Score the round on breakage, not on winning. The output you want is a list of the exact nodes that failed.

6 · Repair, then re-run

Every break is one of four things, and the fix differs:

• Evidence gap → new research, new REF entry
• Logic gap → the inference doesn’t hold; rebuild or drop the node
• Framing loss → they defined the terms; write a pre-emptive framing line
• Retrieval failure → you had the answer and couldn’t produce it → this is a practice problem, not a content problem. Don’t add notes; do more closed-book reps.

That last one is the most commonly misdiagnosed. Adding material to fix a retrieval failure makes the problem worse.

7 · The real debate

Hold the Master only. → Debate Template for the full pre-debate workup (opponent model, burdens, what not to get dragged into).

Decide your concessions in advance

Pick the two or three weakest items in your own case and plan to concede them cheerfully and early. It costs almost nothing, it’s disarming, and it protects everything else — an opponent who catches you defending an indefensible detail will use it to impeach your strong material. You did exactly this with the pre-1948 attacks list: drop the three weak entries, and the remaining seventeen become stronger, not weaker.

8 · Post-mortem within 24 hours

→ Post Debate Template. The question that matters most is not “did I win” but “which node broke that I thought was solid?” Feed it back to step 6, and log any genuine update to your own view — if a month of adversarial prep never changes your mind about anything, the loop is running as motivated reasoning.

---

🗓️ Spacing — how to schedule it

Cramming a debate project into one weekend produces a document, not knowledge. Distributed prep produces both.

---

🧭 Where this method applies beyond debating

The loop is a general engine for contested topics, and most of what’s worth learning is contested. Use it whenever a subject has a real opposition:

• Evaluating a book you suspect you agree with too easily → Process to Read a Book · Reading Hard Books
• Stress-testing a belief you hold for social rather than evidential reasons
• Preparing for a high-stakes conversation, negotiation, or interview → Persuasion · How to Have Impossible Conversations - Peter Gregory Boghossian James A Lindsay

Complementary methods in this folder

• Feynman Effect — explain it simply; catches jargon-as-understanding. Run it on any node you can’t state in one sentence.
• Active Learning · Memory Retrieval Techniques — the retrieval machinery behind the closed-book rebuilds.
• How to Use Writing to Improve Your Thinking — Atul Gawande · I’m begging you to write essays — writing is the thinking; the tree is a specialized case.
• Sympathy vs Empathy — steelmanning requires understanding a position without being captured by it. That’s the same distinction, applied to argument.

Honest limitations of this method

• It optimizes for winning arguments, which is not identical to being right. Adversarial framing makes you better at defending a position — including a wrong one. The correction is step 8: log genuine updates, and treat “I never change my mind” as a red flag rather than a sign of rigor.
• AI sparring has a ceiling. A model playing your opponent draws on the same public corpus you researched. It will rarely produce the argument that a genuinely original opponent brings. It is a drill partner, not a sparring opponent of unknown strength.
• Evidence hoarding feels like progress. Word count and source count are not the metric; tree depth and closed-book recall are.