Create a compact horizontal 4-panel vector academic figure in NeurIPS style: white background, thin borders, mild rounded corners, minimal decoration, tightly packed, no infographic/poster look. Title: Fixed Evidence, Controlled Failures, and Contract-Based Payoff. Color scheme: blue setup, orange failures, purple contract, green payoff. Panels left to right with arrows: (a) Fixed Evidence Setup: schematic prediction heatmap + P(ignite) color bar; dashed divider; blue 3D feature cube with axes C,Y,X,T; arrow to candidate heads H1,H2,…,Hk; bottom boxes: Held fixed: inputs, split, predictions/features, post-processing; Changed later: reading or selector. (b) Fixed-Output Reading Instability: subtitle outputs fixed within each entry; small heatmap; 3 mini-cards: Strict, T+S, Union, each with a tiny grid schematic and minimal legend; bottom anchor Strict F1 0.86% -> Union F1 51.46%; italic note same prediction record; only the reading changes. (c) Fixed-Feature Selection Regret: blue feature cube; candidate heads H1,H2,…,Hk; two selector boxes PR-AUC selector chooses h_R and Union-F1 selector chooses h_D; formula Regret = Union-F1(h_D) - Union-F1(h_R); bottom anchor up to 88.10 pp union-F1 regret. (d) Contract Before Claim: formula C = (T, M, Λ, S, A); list T Task, M Metric/evidence view, Λ Tolerance, S Scope, A Adaptation family; bottom two green payoff boxes: occupancy/spread — clearest decision-task gap and retrieval/regression — narrower gaps, ties, different orderings.. Emphasize a cohesive composition, cinematic lighting, vivid details, strong subject focus, and layered depth.

The image shows a black and white diagram depicting fixed evidence, controlled failures, and contract-based payoff. The diagram is composed of several boxes connected by arrows, with text written on them.