Glossary · AI prepares, people decide

The words behind a defensible evaluation.

Plain definitions for the terms organizers, program leads and investors meet when AI enters application review. Each page opens with a self-contained definition, then shows how the concept behaves on a real batch, with our own numbers where we have them: six independent judges, a spread threshold of 3.0, and 1,000+ evaluation runs.

The map follows the work, not the alphabet: what comes in, what breaks, what fixes it, and what the rules allow.

The map

Eight terms, four stages.

Cluster 1

What comes in, and how to sort it

Deck triageThe first-pass gate that routes every deck to advance, hold or decline, with a recorded reason for each.Read the definition

Cluster 2

What breaks in panel scoring

Score spreadThe gap between the highest and lowest judge on one criterion, and why it should be surfaced, not averaged.Read the definition

Cluster 3

What fixes it

LLM as a judgeA language model scoring work against criteria, and what changes when the work is a person's application.Read the definition AI judge panelSeveral AI reviewers scoring independently. Worth more than one judge only if they can actually disagree.Read the definition Evidence-grounded scoringEvery score tied to a quote on a named slide. No citation, no score, and missing proof logged as a finding.Read the definition Deterministic aggregationJudge scores combined with fixed math outside the model, so the same inputs always give the same result.Read the definition Advisory scoreAn AI number that informs the decision but has no path into the leaderboard, by architecture.Read the definition

Cluster 4

What the rules allow

ERC AI guidelinesThe ERC's 2026 rules: no delegated merit assessment, no uploading proposals, and what that leaves open.Read the definition

After the definitions

The glossary defines. The trust pages prove.

A definition tells you what a concept is. Whether a product actually behaves that way is a measurement question, and that is what the trust section is for.

The proofMethodologyThe glossary defines the vocabulary. The methodology page shows the system built from it.The measurementsConsistency and reliabilityRepeat-run behavior of the pipeline, with the spread and reproducibility numbers measured.The storiesNewsroomWhat these concepts look like when they meet real batches of decks.