Green = funded and truly in the top f; violet = funded but not in the true top f; orange = truly top but missed; grey = not funded. The yellow bar is the score cut-off.
The science
Synthetic model: true quality q ~ N(0,1); a reviewer reports s = q + sigma e with e ~ N(0,1). The correlation is rho = 1/sqrt(1 + sigma^2). If the fraction f with the highest scores is funded, the expected true quality of the funded set is E[q | funded] = rho phi(z_f) / f with z_f = Phi^-1(1 - f) (the truncated-normal selection intensity). A partial lottery draws at random from the top 2f by score, so its mean is the mean of that larger pool, rho phi(z_2f) / (2f); a pure lottery gives 0 (the average proposal).
Honest limits. Quality is a single invented number with a Gaussian spread; real proposals have many dimensions, reviewers share biases (not just independent noise), and "quality" is not measurable after the fact. Funding rate, noise level and the 2f pool are illustrative choices. The model does not say how real panels perform; it only shows how noise reduces the benefit of ranking and what a lottery gives up in this simplified setting. Some funders have trialled partial lotteries for reasons beyond this metric (workload, fairness, risk-taking), none of which is modelled.
Checks. A headless run with 300 000 fresh proposals compares the empirical mean quality of the top fraction with the formula for all three modes; the 400-proposal board is a single noisy batch (standard error about 0.07). Scene: Blender-baked review room (tile, oak, plaster, brushed steel) in Three.js.