Evidence-grounded scoring
Evidence-grounded scoring is an evaluation method in which every score must be tied to specific evidence in the submission itself, such as a quoted line on a named slide. A claim without a citation does not move a score, and evidence that is missing is recorded as a finding in its own right rather than silently lowering a number.
It borrows the groundedness idea from AI reliability work and applies it to the score instead of the answer.
A number you can argue with.
An ungrounded score is unanswerable. Ask why a deck got a 6 on traction and the honest reply is “overall impression,” which no founder, judge or auditor can test. A grounded score converts the dispute into an evidence check: here is the quote, here is the slide, and either it supports the score or it does not.
The second effect is that absence becomes signal. A deck that claims traction but shows no number on any slide produces a recorded gap: usable as feedback to the founder, and as the exact thing a human judge should verify in the meeting. The third effect lands on the judge itself. When the judge is a model, forcing every finding through a citation constrains invented praise and invented criticism alike, which is why the discipline pairs naturally with LLM-as-a-judge.
No quote, no score.
Each of the six independent EvalLens judges must cite the submission for every criterion finding, and the report shows the quote next to the slide it came from. What the deck did not show is a separate section of the report, not a hidden penalty. Where judges disagree, the conflict is surfaced as score spreadwith each judge’s evidence side by side, so the human making the call can compare readings instead of adjudicating vibes. AI prepares the analysis, people decide.