Chapter 34: Adversarial Collaboration and Other Tools for Producing Less Wrong Knowledge

"The first principle is that you must not fool yourself — and you are the easiest person to fool." — Richard Feynman

Chapter Overview

Every chapter in this book has documented a failure — a way that knowledge systems go wrong. This chapter is about the fixes.

Not theoretical fixes. Not "people should be more rational" fixes. Specific, implementable institutional designs that target specific failure modes and have been tested (to varying degrees) in real settings. Some work well. Some work partially. Some haven't been tried widely enough to evaluate. And some — this is important — carry their own risks, including the overcorrection dynamic from Chapter 21.

The nine tools evaluated in this chapter are not a wish list. They are the most promising institutional innovations for reducing error in knowledge production, assessed against the failure mode framework developed throughout this book.

In this chapter, you will learn to: - Evaluate nine tools for reducing error in knowledge production - Connect each tool to specific failure modes - Assess evidence of effectiveness, including limitations - Design correction mechanisms for specific fields

🏃 Fast Track: If you're looking for specific tools to implement, use the summary table at the end of the chapter to identify tools that address your field's specific vulnerabilities (as identified by the Epistemic Health Checklist in Chapter 32).

🔬 Deep Dive: After this chapter, read the Center for Open Science's resources on pre-registration and registered reports (cos.io), and Kahneman's description of adversarial collaboration in Thinking, Fast and Slow for the original vision.

The Nine Tools

Tool 1: Adversarial Collaboration

What it is: Two researchers who disagree about a claim jointly design a study to test it. Both agree in advance on the methodology, the data, and the criteria for determining which position the evidence supports. They publish together regardless of the outcome.

What failure modes it addresses: - Consensus enforcement (Ch.14): Forces genuine engagement with opposing views rather than dismissal - Confirmation bias: Both sides design the study, preventing either from rigging it - Researcher degrees of freedom: Methodology is agreed in advance, eliminating post-hoc analytical choices

Where it's been tried: Daniel Kahneman championed adversarial collaboration as a format for resolving scientific disagreements. Several adversarial collaborations have been conducted in psychology and behavioral economics, including a notable collaboration between Kahneman himself and advocates of priming effects.

How well it works: Mixed. When both parties engage genuinely, adversarial collaboration produces higher-quality evidence than either party would produce alone — because each side designs the study to be as rigorous as possible against the other's position. The result is a study that both sides have pre-committed to accepting.

Limitations: - Willingness problem: Both parties must agree to participate. Defenders of the consensus have no incentive to enter an adversarial collaboration — they're winning under the current system. The tool is most useful when the disagreement is genuine and both sides are uncertain. - Asymmetric stakes: If one side has more at stake (career, reputation, funding), the collaboration may be undermined by unconscious or strategic behavior. - Scale: Adversarial collaboration addresses individual disputes, not systemic bias. It cannot fix a field's incentive structure.

Verdict: Excellent for resolving specific, well-defined scientific disagreements between parties who are genuinely uncertain. Not a systemic fix.

Tool 2: Pre-Registration

What it is: Before collecting data, researchers publicly register their hypothesis, methodology, sample size, and analysis plan. This prevents them from adjusting their approach after seeing the results — the practice known as "p-hacking" or exploiting researcher degrees of freedom.

What failure modes it addresses: - The replication problem (Ch.10): Reduces false positives by preventing post-hoc analytical flexibility - Incentive structures (Ch.11): Makes it harder to "find" results that aren't there - Precision without accuracy (Ch.12): Forces researchers to commit to specific predictions rather than claiming precision after the fact

Where it's been tried: Pre-registration has been adopted in clinical medicine for decades (ClinicalTrials.gov requires registration of clinical trials). In social science, the practice has grown rapidly since 2013, driven by the Open Science Framework.

How well it works: Strong evidence that pre-registration reduces false-positive rates in clinical trials. Emerging evidence that it improves the reliability of social science findings. Studies published from pre-registered analyses show smaller effect sizes than non-pre-registered studies — suggesting that non-pre-registered studies are inflated by analytical flexibility.

Limitations: - Exploratory research: Pre-registration is designed for confirmatory research (testing specific hypotheses). Exploratory research — looking for patterns without a specific prediction — is essential to science but doesn't fit the pre-registration framework. Some critics worry that overemphasis on pre-registration could stifle exploration. - Compliance: Researchers can pre-register vaguely, deviate from their plan without reporting it, or pre-register after seeing their data. Enforcement mechanisms are limited. - The overcorrection risk (Ch.21): If pre-registration becomes a rigid requirement for all research, it could suppress legitimate exploratory work — the pendulum swinging from "too much flexibility" to "too little flexibility."

Verdict: One of the most effective tools available. Should be standard for confirmatory research. Should not be required for exploratory research.

Tool 3: Registered Reports

What it is: A publication format in which the journal reviews and accepts (or rejects) a study before the data are collected. The acceptance decision is based on the importance of the question and the rigor of the methodology — not on the results. If the study is accepted at Stage 1, the journal commits to publishing the results regardless of whether they support the hypothesis.

What failure modes it addresses: - Publication bias (Ch.10, Ch.11): Eliminates the bias toward positive results, because the publication decision is made before results exist - The replication problem (Ch.10): Removes the incentive to p-hack (results don't affect publication) - Incentive structures (Ch.11): Changes what is rewarded — question importance and methodological rigor rather than surprising results

Where it's been tried: Over 300 journals across multiple fields have adopted registered reports since 2013. The format has been used in psychology, neuroscience, ecology, political science, and other fields.

How well it works: Strikingly effective. Studies published as registered reports have a much higher rate of null results (approximately 55-60% report null or mixed results, compared to approximately 5-10% in traditional publications). This doesn't mean registered reports produce worse science — it means traditional publications are massively biased toward positive results. Registered reports reveal the actual distribution of findings.

Limitations: - Adoption: Despite growing, registered reports remain a small fraction of total publications. Most journals still use traditional review, and most researchers still publish traditionally. - Question selection: Journal editors may be biased toward accepting registered reports on "interesting" questions — potentially recreating the novelty bias in a different form. - Speed: The two-stage review process is slower than traditional submission.

Verdict: The single most effective institutional innovation for reducing publication bias. If one tool from this chapter were universally adopted, it should be this one.

🔄 Check Your Understanding (try to answer without scrolling up)

1. How do registered reports differ from pre-registration? What additional problem does the registered report format solve?
2. Why do registered reports produce ~55-60% null results while traditional publications produce only ~5-10%?

Verify

1. Pre-registration commits the researcher to a plan but doesn't change the publication process — journals can still reject studies with null results. Registered reports commit the journal to publishing regardless of results, eliminating publication bias in addition to researcher flexibility. 2. Not because registered report studies are worse, but because traditional publications are massively filtered — studies with null results aren't submitted (file drawer effect) or aren't accepted (publication bias). Registered reports reveal the actual distribution of findings by removing the filter.

Tool 4: Prediction Markets for Scientific Claims

What it is: Markets where researchers and others can bet on whether scientific claims will replicate. The market price reflects the aggregated probability assessment of all participants — producing a continuously updated estimate of how likely a claim is to be true.

What failure modes it addresses: - Consensus enforcement (Ch.14): Provides an anonymous mechanism for expressing doubt — you can bet against a claim without publicly challenging the consensus - Authority cascade (Ch.2): The market doesn't weight votes by prestige — a junior researcher's bet counts the same as a senior researcher's - Incentive structures (Ch.11): Creates a financial incentive for accurate assessment rather than for consensus maintenance

Where it's been tried: Small-scale prediction markets for replication outcomes have been run in psychology and economics. Research by Anna Dreber and colleagues found that prediction markets were reasonably accurate at predicting which studies would replicate.

How well it works: Promising but limited. Prediction markets have correctly predicted replication outcomes at rates better than expert surveys. The anonymity and financial incentive create conditions where private doubts can be expressed without career risk.

Limitations: - Thin markets: Scientific prediction markets have few participants, making prices volatile and potentially manipulable. - Legal and institutional barriers: Betting on scientific outcomes raises regulatory and ethical concerns in some jurisdictions. - Gaming: Researchers who plan to conduct a replication could bet on the outcome — creating insider-trading dynamics. - Scale: Currently a research curiosity rather than an institutional tool.

Verdict: Promising concept with demonstrated accuracy in small-scale tests. Not yet ready for institutional deployment but worth continued development.

Tool 5: Red Teams

What it is: A designated group whose job is to find flaws in a plan, theory, or decision before it is implemented. Red teams argue the opposite of the institution's position, deliberately looking for weaknesses.

What failure modes it addresses: - Consensus enforcement (Ch.14): Institutionalizes dissent — red team members are expected to disagree - Einstellung (Ch.13): Forces consideration of alternatives that the institution's framework might exclude - The streetlight effect (Ch.4): Red teams can ask "what are we not looking at?"

Where it's been tried: The U.S. military has the most developed red team culture — the University of Foreign Military and Cultural Studies at Fort Leavenworth trains red team leaders. Intelligence agencies use red teams to challenge analytical conclusions. Some corporations use them for strategic planning.

How well it works: Effectiveness varies enormously with institutional culture. In organizations that genuinely value dissent, red teams improve decision quality. In organizations that treat red teams as a checkbox — performing the exercise without acting on the findings — they are theater.

Limitations: - Authority problem: Red team findings are recommendations, not decisions. If leadership ignores them, the exercise is pointless. The military's experience (Chapter 28) shows that even extensive red teaming doesn't prevent doctrinal lock-in when structural incentives override the red team's warnings. - Talent allocation: Red teams require smart, knowledgeable people who could otherwise be doing other work. If the institution assigns its weakest people to the red team, the exercise fails. - Normalization: Over time, red teams can become routine — the institutional equivalent of "have you tried turning it off and on again?" The dissent becomes performative rather than genuine.

Verdict: Valuable when institutional culture genuinely supports dissent. Counterproductive when it becomes theater. Effectiveness is entirely determined by the institution's Dissent Tolerance score (Checklist D1).

Tool 6: Independent Replication Funding

What it is: Dedicated funding for independent replication of important findings — research money allocated specifically for checking other people's work, not for producing novel results.

What failure modes it addresses: - The replication problem (Ch.10): Directly funds the activity that is currently disincentivized - Incentive structures (Ch.11): Creates a career path for replication researchers

Where it's been tried: The Dutch science funding agency NWO funded a Replication Studies program. The Education Endowment Foundation in the UK funds independent replications of educational interventions. DARPA has funded systematic replication in some areas.

How well it works: Where it has been tried, it works — independent replications catch errors, calibrate effect sizes, and build more reliable evidence bases. The problem is scale: replication funding remains a tiny fraction of total research funding in every field.

Limitations: - Opportunity cost: Money spent on replication is money not spent on novel research. Some researchers argue this slows discovery. - Political difficulty: Funding agencies are rewarded for "impact" — and replication is perceived as less impactful than discovery (even though unreliable discoveries have negative impact). - Relationship damage: Funding replications of specific researchers' work can be perceived as targeting — creating interpersonal and institutional friction.

Verdict: Essential and underfunded. Every field would benefit from dedicating 5-10% of its research funding to independent replication.

Tool 7: Prize-Based Science

What it is: Instead of funding research through grants (inputs), fund it through prizes for solving specific problems (outputs). Researchers compete to demonstrate solutions, and the prize is awarded to whoever succeeds.

What failure modes it addresses: - Incentive structures (Ch.11): Rewards outcomes rather than activity — aligns incentives with truth - Outsider access (Ch.18): Anyone can compete for a prize, regardless of credentials or institutional affiliation - Capital-sustained error (Ch.29): Prizes are paid only for results, preventing the capital-sustains-wrong-ideas dynamic

Where it's been tried: The X Prize Foundation, DARPA Grand Challenges, and various innovation prizes have used this model. The Netflix Prize ($1 million for improving the recommendation algorithm) demonstrated that open competitions can produce solutions from unexpected sources.

Limitations: - Measurability: Prize-based science works only when the goal can be clearly defined and the success criterion is measurable. This works for engineering challenges but not for basic research ("understand consciousness" is not a prize-able question). - Winner-take-all dynamics: Prizes concentrate reward on the winner, while many losing teams may have produced valuable knowledge that goes unrewarded. - Short-termism: Prizes reward specific achievements and may discourage long-term, open-ended exploration.

Verdict: Excellent for well-defined technical challenges. Not a replacement for grant-funded basic research. Best used as a complement.

Tool 8: Open Data Mandates

What it is: Requirements that researchers make their raw data publicly available, enabling independent re-analysis and verification.

What failure modes it addresses: - Process transparency (Checklist D10): Makes the evidence base inspectable - Researcher degrees of freedom: Others can check whether alternative analyses produce the same results - The replication problem (Ch.10): Enables computational replication even when full experimental replication is impractical

Where it's been tried: Multiple funding agencies and journals now require or encourage data sharing. The FAIR principles (Findable, Accessible, Interoperable, Reusable) provide a framework for data sharing.

How well it works: When implemented with genuine commitment, open data dramatically improves error detection. Several high-profile errors have been caught through re-analysis of shared data — including the Reinhart-Rogoff Excel error that influenced austerity policy (Chapter 24).

Limitations: - Privacy: Some data (medical, educational, judicial) cannot be fully shared without compromising individual privacy. De-identification is imperfect and can be reversed. - Incentive misalignment: Researchers who collected data have an incentive to delay or restrict sharing — the data represents competitive advantage. - Usability: Raw data without adequate documentation is unusable. Sharing data without sharing the codebook, cleaning procedures, and analytical pipeline is insufficient.

Verdict: Essential infrastructure for error correction. Privacy concerns are real and require careful handling, but the default should be openness with privacy protections — not closedness with sharing exceptions.

Tool 9: Post-Publication Peer Review

What it is: Mechanisms for reviewing and critiquing published work after publication — through formal comments, letters, re-analyses, online platforms (PubPeer), or structured post-publication review processes.

What failure modes it addresses: - Consensus enforcement (Ch.14): Allows challenges to published work that pre-publication peer review missed - The replication problem (Ch.10): Catches errors that the original reviewers didn't - Authority cascade (Ch.2): Enables anyone (not just the original reviewers) to evaluate published claims

Where it's been tried: PubPeer provides anonymous post-publication review for scientific papers. Some journals publish formal commentary and response formats. Social media (particularly academic Twitter) has become an informal post-publication review mechanism.

How well it works: Post-publication review has caught numerous errors, including data fabrication, statistical errors, and methodological flaws that pre-publication reviewers missed. PubPeer has been instrumental in identifying problematic papers across multiple fields.

Limitations: - Anonymity vs. accountability: Anonymous post-publication review enables junior researchers to challenge senior figures without career risk — but also enables harassment, bad-faith critiques, and personal attacks. - No institutional teeth: Post-publication critiques rarely lead to formal corrections or retractions. The system for updating the scientific record after publication is slow and ineffective. - Attention asymmetry: The original paper gets thousands of reads; the correction gets dozens. The wrong finding persists in citation networks even after it has been critiqued.

Verdict: Essential complement to pre-publication review. Currently underdeveloped and underinstitutionalized. Needs formal mechanisms for updating the scientific record when post-publication review identifies errors.

Summary: The Tool-Failure Mode Matrix

🔗 Connection: Every tool in this chapter can be mapped onto the Epistemic Health Checklist (Chapter 32). Pre-registration and registered reports improve Dimension 2 (Replication Culture). Red teams improve Dimension 1 (Dissent Tolerance). Open data improves Dimension 10 (Process Transparency). Independent replication funding improves Dimension 3 (Incentive Alignment). A field that scores poorly on specific dimensions can use this mapping to identify which tools would produce the largest improvement.

The Overcorrection Warning

Every tool in this chapter carries a risk identified in Chapter 21: overcorrection. The pendulum dynamic predicts that the trauma of being wrong can cause a field to swing too far in the opposite direction:

• Pre-registration, if made mandatory for all research, could suppress legitimate exploratory work
• Replication demands, if taken to extremes, could paralyze fields with "replication paralysis" — no one publishes novel findings for fear of non-replication
• Open data mandates, if implemented without privacy safeguards, could harm research participants
• Red teams, if given too much authority, could create a culture of destructive criticism

The goal is calibrated correction (Chapter 21) — arriving at the right answer rather than the opposite wrong answer. Each tool should be implemented with awareness of its overcorrection risk and with mechanisms for detecting and correcting the overcorrection.

📐 Project Checkpoint

Epistemic Audit — Chapter 34 Addition: The Correction Mechanism Assessment

34A. Inventory. Which of the nine tools in this chapter exist in your field? For each, rate its maturity: (a) fully implemented, (b) partially implemented, (c) discussed but not implemented, (d) not discussed.

34B. Gap Analysis. Based on your Epistemic Health Checklist scores (Chapter 32), which tools would address your field's lowest-scoring dimensions? Rank the tools by expected impact.

34C. Design. Design one correction mechanism that would improve your field's epistemic health. Specify: what failure mode it addresses, how it would work, who would implement it, what resistance you would expect, and what overcorrection risk it carries.

34.2 Chapter Summary

Key Concepts

• Nine institutional tools for reducing error, each targeting specific failure modes
• Registered reports are the single most effective innovation for reducing publication bias
• Pre-registration is the most widely adopted and effective tool for reducing researcher degrees of freedom
• The Tool-Failure Mode Matrix maps each tool to the specific failure modes it addresses
• The overcorrection warning — every tool carries its own risk of swinging too far

Key Arguments

• Individual dissent (Chapter 33) is necessary but insufficient — systemic error requires systemic correction
• The most effective tools change incentive structures rather than relying on individual behavior change
• No single tool addresses all failure modes — fields need a portfolio of correction mechanisms tailored to their specific vulnerability profile
• Every correction mechanism can itself become a source of error (Theme 9) — overcorrection risk must be monitored

Spaced Review

Revisiting earlier material to strengthen retention.

1. (From Chapter 21 — When Correction Overcorrects) The chapter warns that every tool carries overcorrection risk. Apply the overcorrection framework to pre-registration: what would "too much pre-registration" look like? What legitimate scientific activities would be suppressed? How would you detect the overcorrection?

2. (From Chapter 11 — Incentive Structures) Registered reports change the incentive structure of publishing by making the publication decision independent of results. Apply the incentive framework: what new incentives does the registered report format create? Could these new incentives produce their own failure modes?

3. (From Chapter 22 — The Speed of Truth) If a field adopted all nine tools simultaneously, how would its Correction Speed Model profile change? Which variables would improve? Which would remain unchanged?

What's Next

In Chapter 35: The Humility Chapter, the book turns its lens inward. If you've absorbed everything in this book, you now know that failure modes trap smart people, that every field believes it's uniquely rational, and that you are currently wrong about something important. Chapter 35 provides calibration exercises that demonstrate this — not as a gotcha, but as a liberation.

Before moving on, complete the exercises and quiz to solidify your understanding.

Chapter 34 Exercises → exercises.md

Chapter 34 Quiz → quiz.md

Case Study: Registered Reports in Action — Psychology's Most Effective Reform → case-study-01.md

Case Study: Red Teams That Failed — When Institutional Design Meets Institutional Culture → case-study-02.md