ICLR Paper Acceptance — Conference Extension v2 (Calibrated)

Two-stage leakage-safe pipeline.

• Stage 1 ensembles six reviewer-score regressors (Ridge TF-IDF, Ridge / ElasticNet on SPECTER, Ridge on MiniLM, ExtraTrees & HistGradientBoosting on SVD+metadata) into a predicted_score_mean / std / range proxy. 5-fold OOF inside training years prevents within-train leakage.
• Stage 2 combines TF-IDF + metadata + the Stage 1 proxy through a Platt-calibrated logistic regression.

Best deployable model: S2v2 TFIDF+Meta+ScoreProxy LR (v2 score) (Platt-calibrated) Decision threshold (chosen on validation): 0.240 Validation AUC: 0.6873    Test AUC: 0.6938 Calibration: Brier 0.199 (val) / 0.193 (test); ECE 0.061 (val) / 0.058 (test). Scientific embedding: allenai/specter    Lightweight embedding: sentence-transformers/all-MiniLM-L6-v2 (both frozen).

⚠️ This is a decision-support research demo, not a replacement for peer review. Example inputs are synthetic and used only to demonstrate model behavior.

Disclaimer. Acceptance depends on reviewer strictness, reviewer–paper fit, year-specific standards, rebuttals, and area-chair decisions — much of which is unobservable before submission. Treat the output as a calibrated signal, not a verdict.

Real reviewer scores are used only as Stage 1 regression target / Oracle benchmark. Reviews, rebuttals, reviewer confidence, decision text, the accepted label, and reviewer identity are excluded from the deployable model.