Vol. III · Multi-Model Reasoning · ed. 2026
Plurality
pp. i — xliv
Sign inBegin a query →
Article the First

On the use of many models, where one had been thought sufficient.

Being a treatise on structured deliberation across thirty-two large language models, the seven modes by which they may be made to reason in concert, and the cross-signals by which their findings may be trusted.
By the editors of Plurality  ·  Revised the spring of MMXXVI

A single model speaks the same way when it knows and when it guesses¹. We submit your question to a panel instead, and show you where they agree, and where they do not.

Begin a deliberation →Read § II — the modes
§ I.

On the failure of the single voice

Marginalia
cf. Article §§ III–IV, where the panel and the rounds are described in greater detail. See also fig. 1, opp.
The Failure
Confident prose, regardless of warrant.
A model speaks the same way when it knows and when it guesses. The reader cannot tell.
No exposure of disagreement.
Where a peer would have said otherwise, you receive only the answer of the one you asked.
Brittle expertise.
Each model has its own blind spots, inherited from its training. Asking only one is asking only those blind spots.
The Remedy
Independent rounds.
Models answer in parallel before they see one another, so anchoring is contained, and dissent is not silently extinguished.
Two consensus signals.
Semantic agreement and verdict alignment are required to converge. Either alone is insufficient.
Disagreement, made first-class.
Where the panel splits, the report shows the split — both arguments — rather than smoothing one away.
§ II.

The seven modes of deliberation

Index of Modes
Entry · § II.i
Quick Take.
n. fast independent views, one essay.
Best for
Fast first pass, yes/no decisions.
Rounds
1.
Pipeline
parallel—synthesize
Worked example
"Should we price by seat, by usage, or with a flat tier?"
Opus
Seat-based. Budgets are how teams buy; predictability wins enterprise.
GPT-5
Usage-based. Pain you create is pain you can charge for, and seat caps stop scaling.
Gemini
Flat tier with usage above the line. Floor for budgeting, ceiling for upside.
Synthesis
Three different shapes, no winner. Pick by your buyer: enterprise procurement (seat), self-serve growth (usage), or simplicity at low ACV (flat).
§ III.

The panel — an index of contributors

Frontier and open-weight models from ten vendors, listed alphabetically by house. Build a custom panel; or call a preset.

Anthropic
01Claude Opus 4.5q.v.
02Claude Sonnet 4.5q.v.
03Claude Haiku 4.5q.v.
OpenAI
04GPT-5q.v.
05GPT-5 miniq.v.
06o4q.v.
07o4-miniq.v.
Google
08Gemini 2.5 Proq.v.
09Gemini 2.5 Flashq.v.
10Gemini 3q.v.
DeepSeek
11DeepSeek V3q.v.
12DeepSeek R1q.v.
Meta
13Llama 4 Scoutq.v.
14Llama 4 Maverickq.v.
15Llama 3.3 70Bq.v.
Mistral
16Mistral Large 2q.v.
17Mixtral 8×22Bq.v.
18Codestralq.v.
Cohere
19Command R+q.v.
20Command Aq.v.
21Aya Expanseq.v.
xAI
22Grok 4q.v.
23Grok 3q.v.
Alibaba
24Qwen 3 Maxq.v.
25Qwen 2.5 72Bq.v.
01.AI
26Yi-Lightningq.v.
Reka
27Reka Coreq.v.
Microsoft
28Phi-4q.v.
Amazon
29Nova Proq.v.
30Nova Premierq.v.
Upstage
31Solar 1q.v.
AI21
32Jamba 1.6q.v.
Hover any entry for its detail. Models marked ¹ require API authentication.
§ IV.

Of the method, in three movements

I.
On the selecting of a panel.
The reader chooses contributors from the index of thirty-two, drawn from ten houses. Presets are provided for common occasions; a bespoke panel may be composed.
II.
On the rounds of deliberation.
The panel responds in parallel, then in critique, then in revision. Convergence is recorded after each round and made visible to the reader.
III.
On the issuing of the verdict.
Two consensus signals — semantic and verdict-aligned — must agree. Cited claims are checked against external sources. A polished report is composed.
Thus concludes the introduction

Begin your deliberation.

Bring a question of consequence; convene a panel; receive an answer that has survived the scrutiny of many.

Begin a query →
Free of charge · fifty deliberations the month