Panel · Template

Divergence Review.

LiveDivergence Review

Put the identical question to every specialist, then get a map of where they genuinely disagree, whether each split is about evidence, assumption or value, the crux each side turns on, and the datum or test that would settle it.

You getA map of disagreement of about four pages (Standard)

DepthsLight (three seats), Standard (four) or Full (five)

Written byAI specialists, each answering the same question

SourcesEvery claim cited or marked as a gap

Add it to your library to reuse from New → From template. Edit the saved copy in your library on the service form.

When to use it

You expect the disciplines or the markets to disagree, and you need to know where, why and what would resolve it, before you write a position that assumes they agree.

Who it is for. Leads who must walk into a room where people disagree: access, medical, regulatory, evidence and clinical operations leads preparing a position, a protocol or a negotiation.

Why a false consensus is the costly output

A position written as if everyone agreed fails in the room. AI debate makes it worse: agents drift toward agreement through sycophancy and lose the variety of stances they started with. This template fixes the question, keeps each first answer independent, shows blank where a seat said nothing, and reports the spread that remains, not only the part that closed.^{1Yao et al. Peacemaker or Troublemaker: How Sycophancy Shapes Multi-Agent Debate. arXiv:2509.23055, 27 September 2025.,5Wynn, Satija and Hadfield. Talk Isn’t Always Cheap: Understanding Failure Modes in Multi-Agent Debate. arXiv:2509.05396, 5 September 2025.}

What you get

The question as put, word for word. The same text that every seat received, so any difference comes from the expert and not from the wording.

The shape of the disagreement. How many positions exist, who holds each, and how wide the spread is, before any detail.^{3Consensus and stability measures in Delphi studies. BMC Medical Research Methodology.}

Each position, in its own terms. The seat’s genuine view from its discipline, attributed by name. Nobody is assigned to argue the other side.^{4Behavioral Scientist. Stop playing devil’s advocate and other advice for better decision-making.}

Where they diverge, and why. A table of the points of divergence, typed as evidence, assumption or value, labeled analyst judgment. A seat that did not address a point is shown blank and never as disagreement.

The cruxes, and what would settle it. At Full depth, the facts that would change a seat’s conclusion. Every report closes with a test or datum and the decision it would settle, agreed in advance in the manner of an adversarial collaboration.^{2Mellers, B., Hertwig, R. and Kahneman, D. Do frequency representations eliminate conjunction effects? An exercise in adversarial collaboration. Psychological Science, 2001;12(4):269–275.}

What is in it, by length

Section	Brief	Standard	Comprehensive
The question as putThe exact question every seat received	●	●	●
The shape of the disagreementHow many positions and how wide the spread	●	●	●
Each position, in its own termsEach seat’s view, attributed by name	●	●	●
Where they divergeThe points of divergence and who holds which side	●	●	●
Why they diverge: evidence, assumption or valueThe type of each disagreement	–	●	●
The cruxesThe facts each side turns on	–	–	●
What would settle itThe test or datum, and the decision it settles	●	●	●

A Light review has three seats and about two pages. Standard has four seats and about four pages and types each divergence. Full has five seats and adds the cruxes.

Choosing a length

Choose Standard for most questions. Choose Full when a decision hangs on one disputed fact, so you can see the crux. Choose Light for a first read.

The standards and practice behind it

The structure and the checks come from published work on expert disagreement and on the failure modes of AI debate. These are the references we drew on and what we took from each. None of their authors has reviewed the template.

What we took	From	Where it shows
Treat sycophancy and premature consensus as a failure of AI debate, and keep each seat’s first answer independent	Peacemaker or Troublemaker: How Sycophancy Shapes Multi-Agent Debate (arXiv, 2025)^{1Yao et al. Peacemaker or Troublemaker: How Sycophancy Shapes Multi-Agent Debate. arXiv:2509.23055, 27 September 2025.}	The question held fixed and each first answer independent
Have rivals agree beforehand on the test and on what result would change their minds	Adversarial collaboration (Mellers, Hertwig and Kahneman, Psychological Science, 2001)^{2Mellers, B., Hertwig, R. and Kahneman, D. Do frequency representations eliminate conjunction effects? An exercise in adversarial collaboration. Psychological Science, 2001;12(4):269–275.}	The closing section: a test or datum and the decision it settles
Describe how far a panel moved and whether it stabilized; no single measure of convergence is agreed	A review of consensus and stability statistics in Delphi studies (BMC Medical Research Methodology)^{3Consensus and stability measures in Delphi studies. BMC Medical Research Methodology.}	The shape of the disagreement, reported as spread and not as an average
Prefer authentic dissent to assigned devil’s advocacy	Nemeth’s work on dissent, as summarized in Behavioral Scientist^{4Behavioral Scientist. Stop playing devil’s advocate and other advice for better decision-making.}	Each seat reports its own view from its discipline
Watch for models abandoning correct answers under peer pressure	Talk Isn’t Always Cheap: Understanding Failure Modes in Multi-Agent Debate (arXiv, 2025)^{5Wynn, Satija and Hadfield. Talk Isn’t Always Cheap: Understanding Failure Modes in Multi-Agent Debate. arXiv:2509.05396, 5 September 2025.}	Blank cells, never invented agreement or disagreement

What it will not do

It does not force a consensus. A report that closes every gap would be the wrong output.

It is not a vote of human experts. Every seat is an AI specialist, and the report says so.

It does not name a winner. It says what would settle the question and leaves the decision to you.

References

[1] Yao et al. Peacemaker or Troublemaker: How Sycophancy Shapes Multi-Agent Debate. arXiv:2509.23055, 27 September 2025. Abstract read on 8 October 2026.
[2] Mellers, B., Hertwig, R. and Kahneman, D. Do frequency representations eliminate conjunction effects? An exercise in adversarial collaboration. Psychological Science, 2001;12(4):269–275. Known by citation; the paper was not read, and the link was not opened.
[3] Consensus and stability measures in Delphi studies. BMC Medical Research Methodology. Known from search results; not read in full.
[4] Behavioral Scientist. Stop playing devil’s advocate and other advice for better decision-making. Secondary summary of Nemeth’s work; the original studies were not read.
[5] Wynn, Satija and Hadfield. Talk Isn’t Always Cheap: Understanding Failure Modes in Multi-Agent Debate. arXiv:2509.05396, 5 September 2025. Abstract read on 8 October 2026.

Links checked 9 October 2026

These are the sources the template drew on. None of their authors or issuers has reviewed it.

About the Panel service

Each panelist answers in their own voice, then the discussion is drawn together. You get the question and who was in the room, every position in full, what they agreed on, where they split, and what follows. Disagreements are written down rather than smoothed into a consensus nobody actually holds, because the split is usually the part you needed.

Use cases for the Divergence Review

9 ready-to-run use cases. Each opens its own page with the question, the specialists, the standards and the references.

EALiveEuropean Access DivergenceMap where five European markets will genuinely disagree on an access position, why they split and what would settle it, without forcing consensus.Divergence ReviewPublic Affairs

Open Open →

SELiveSurrogate Endpoint DivergenceMap where specialists disagree on whether a surrogate endpoint is adequate evidence of clinical benefit, and what would settle it.Divergence ReviewMedical Affairs

Open Open →

SSLiveSafety Signal DivergenceMap where specialists disagree on whether a safety signal warrants a label change, and what would settle it.Divergence ReviewPharmacovigilance

Open Open →

ECLiveExternal Control DivergenceMap where specialists disagree on whether an external control arm can support accelerated approval, and what would settle it.Divergence ReviewRegulatory

Open Open →

DCLiveDTx Coverage DivergenceMap where specialists disagree on whether the evidence for a prescription digital therapeutic is enough for payer coverage.Divergence ReviewInnovation

Open Open →

TDLiveTrial Design DivergenceMap where specialists disagree on site-based versus decentralized design for a Phase 3 trial, and what would settle it.Divergence ReviewClinical Operations

Open Open →

RELiveReal-World Evidence DivergenceMap where specialists disagree on how far real-world evidence can substitute for a randomized trial, and what would settle it.Divergence ReviewMedical Affairs

Open Open →

LPLiveLegislative Proposal DivergenceMap where member states, political groups and payers will disagree on a legislative proposal, why they split and what would settle it, without forcing consensus.Divergence ReviewPublic Affairs

Open Open →

FALiveFederal and State DivergenceMap where federal agencies, Congress and the named states will disagree on a policy that affects a product class, why they split and what would settle it.Divergence ReviewPublic Affairs

Open Open →

To edit a saved template, open New → Panel → From template, or the library rail on the service form.

When to use it

You expect the disciplines or the markets to disagree, and you need to know where, why and what would resolve it, before you write a position that assumes they agree.

Who it is for. Leads who must walk into a room where people disagree: access, medical, regulatory, evidence and clinical operations leads preparing a position, a protocol or a negotiation.

Why a false consensus is the costly output

What you get

The question as put, word for word. The same text that every seat received, so any difference comes from the expert and not from the wording.

What is in it, by length

Section	Brief	Standard	Comprehensive
The question as putThe exact question every seat received	●	●	●
The shape of the disagreementHow many positions and how wide the spread	●	●	●
Each position, in its own termsEach seat’s view, attributed by name	●	●	●
Where they divergeThe points of divergence and who holds which side	●	●	●
Why they diverge: evidence, assumption or valueThe type of each disagreement	–	●	●
The cruxesThe facts each side turns on	–	–	●
What would settle itThe test or datum, and the decision it settles	●	●	●

A Light review has three seats and about two pages. Standard has four seats and about four pages and types each divergence. Full has five seats and adds the cruxes.

The standards and practice behind it

What we took	From	Where it shows
Treat sycophancy and premature consensus as a failure of AI debate, and keep each seat’s first answer independent	Peacemaker or Troublemaker: How Sycophancy Shapes Multi-Agent Debate (arXiv, 2025)^{1Yao et al. Peacemaker or Troublemaker: How Sycophancy Shapes Multi-Agent Debate. arXiv:2509.23055, 27 September 2025.}	The question held fixed and each first answer independent
Have rivals agree beforehand on the test and on what result would change their minds	Adversarial collaboration (Mellers, Hertwig and Kahneman, Psychological Science, 2001)^{2Mellers, B., Hertwig, R. and Kahneman, D. Do frequency representations eliminate conjunction effects? An exercise in adversarial collaboration. Psychological Science, 2001;12(4):269–275.}	The closing section: a test or datum and the decision it settles
Describe how far a panel moved and whether it stabilized; no single measure of convergence is agreed	A review of consensus and stability statistics in Delphi studies (BMC Medical Research Methodology)^{3Consensus and stability measures in Delphi studies. BMC Medical Research Methodology.}	The shape of the disagreement, reported as spread and not as an average
Prefer authentic dissent to assigned devil’s advocacy	Nemeth’s work on dissent, as summarized in Behavioral Scientist^{4Behavioral Scientist. Stop playing devil’s advocate and other advice for better decision-making.}	Each seat reports its own view from its discipline
Watch for models abandoning correct answers under peer pressure	Talk Isn’t Always Cheap: Understanding Failure Modes in Multi-Agent Debate (arXiv, 2025)^{5Wynn, Satija and Hadfield. Talk Isn’t Always Cheap: Understanding Failure Modes in Multi-Agent Debate. arXiv:2509.05396, 5 September 2025.}	Blank cells, never invented agreement or disagreement