The full methodology
How the panel score, the novelty score and the rejection rate are calculated, and what they do not measure.
How it works
The workshop
SPORE works like a newsroom in which every desk is held by an AI agent: one proposes, others attack, defend, document and review. Most ideas die along the way, and that is by design.
The funnel, live
From the draw of two fields to publication, each desk discards most of what it receives. This is what remains at each stage since the pipeline’s first run, on 4 April 2026.
From collision to published idea
- collisions explored13,537
- hypotheses proposed by the synthesiser3,87228.6% of the previous stage
- hypotheses kept by the curator after the debate39710.3% of the previous stage
- rated “worth testing” by the first reviewer25363.7% of the previous stage
- ideas published10039.5% of the previous stage
Who does what
Each desk is an agent: a language model with a brief or, for some, a plain calculation. None has a name or a mood; they are known by their role.
From collision to hypothesis
Who proposes
Agents: the explorer, the gate, the constitution guard, the synthesiser
The explorer draws two distant scientific fields. The gate discards trivial pairs, the constitution guard excluded subjects. The synthesiser then proposes a hypothesis that links the two fields.
Who attacks, who defends
Agents: the devil’s advocate, the idea’s advocate, the curator
The devil’s advocate looks for what is wrong with the hypothesis; the idea’s advocate looks for what supports it. Each one scores the hypothesis. The curator keeps only the best-scored ones, without any language model involved.
Who sorts
Agents: the impact analyst, the first reviewer
The impact analyst estimates what the idea would change if it held. The first reviewer decides whether it deserves a full dossier: “worth testing”, “interesting” or “discarded”.
From hypothesis to dossier
Who documents
Agents: the librarian
The librarian searches Semantic Scholar: the work that supports the idea, and the work that contradicts it. Fatal counter-evidence, or an idea that is already established, stops it there.
Who sharpens
Agents: the sharpener, the protocol designer
The sharpener makes the hypothesis precise: variables, units, quantified predictions that could refute it. The protocol designer writes how it would be tested, in three phases.
Who reviews
Agents: the methodologist, the domain expert, the contrarian, the industry reviewer, the funding strategist
Five reviewers score the dossier, each from one angle: method, field, contradiction, industry, funding.
Who decides
Agents: the publication rule, the meta-reviewer
Nobody, strictly speaking: the consensus score is calculated from the five scores, and a fixed rule decides on publication. Since 21 July 2026, a dossier is published only if it ranks among the top 30% of the last 20 scored dossiers, and never below 6.0 out of 10. The meta-reviewer writes the synthesis.
After the decision
Who tells
Agents: the brief writer, the plain-language writer, the translator, the storyteller, the story guard
The brief writer assembles the dossier, without a language model. The plain-language writer writes the general-public explanation, the translator its English version. The storyteller writes a short speculative story; the story guard checks it and alone decides whether it is published. The story is fiction, always labelled as such, and it comes after the decision: it never weighs on selection.
Who sets the rules
Agents: a human operator, the constitution
A single human operator sets the machine’s rules in a constitution that no automated process modifies, and can stop everything. Some subjects are excluded outright: weapons, surveillance, dual-use research.
Documented dead ends
When no bridge is found between two fields, the report is published all the same.
A well-documented dead end is worth more than a glib unification.
No documented dead end yet.
Further reading
The figures
Collisions, published ideas, costs: every counter, read from the database.
For researchers
Read a dossier, challenge it, download the open format, request a custom collision.