What EvalLens does not do

It is not an external truth check#

What it does instead: evaluates what the deck presents and flags what is missing. It does not verify claims against the outside world.

A deck that states "14 pilot customers at $190/mo" is scored on the strength and specificity of that evidence as presented. Whether those customers exist is not something a document reader can establish. False, incomplete or unsupported claims still require evidence review and, where it matters, external validation.

The same boundary applies to completeness: missing means the deck did not cover a section, never that a claim in it is untrue.

It is not investment advice#

What it does instead: gives decision support to reviewers. It is not a recommendation to fund or pass, and the output is not shaped as one.

It is not automatic winner selection#

What it does instead: never ranks the batch. The leaderboard is built only from submitted Jury Scores and your criterion weights; the AI Total Score sits beside them, read-only.

There is no mode in which the AI picks the winner. If a ranking changed, a person changed it.

It is not a prediction of success#

What it does instead: describes the pitch today. It does not forecast whether the startup will succeed, and no claim in a report should be read as one.

The related boundary: safety is not fact-checking#

Prompt-injection safety prevents instructions inside a deck from controlling the evaluation. That is a different guarantee from the deck being truthful — see Prompt-injection safety.

What to say out loud#

Programs that adopt an AI-assisted first read do better when they state the boundary before anyone asks, rather than defending it afterwards. Three lines that hold up:

• On stage: "Every entry received a full read under identical rules, and humans made every ranking decision."
• In the rules document: a methodology statement — what the AI panel assists with, what the judges decide, and how a team can ask about its own record.
• On the submission form: plain language, so nobody discovers AI involvement after the results are announced.

Judge conflict-of-interest and recusal handling stays your policy. What the record adds is that a recusal is verifiable later, because who scored what is logged.

The honest comparison#

The awkward question is not "was AI involved". It is what the alternative actually looked like: at a large competition, entry 300 drawn by a tired volunteer at 11pm; at a hackathon expo, the average judge seeing a single-digit percentage of the field in four-minute table visits. A first read under identical rules is a claim that a volunteer process cannot make — and it is compatible with humans deciding every placement.

Next steps#

• Reproducibility — what has been benchmarked and what has not.
• Security and privacy — what happens to a deck after upload.
• Score and shortlist — where the human decision is actually made.