UAI Supplementary

UAI's packaging model (2026 cycle, verified 2026-07-08) has three tiers with sharply different review status. Getting content into the right tier is a strategic decision, not a formatting chore. Reconfirm tier rules in the current submission instructions before finalizing.

The three tiers

Two structural consequences follow. First, because the appendix travels inside the reviewed PDF, UAI has no "the reviewers never saw my proofs" excuse — but also no mercy for identity leaks buried on appendix page 30. Second, because the ZIP is officially optional reading, nothing decision-critical may live only there.

Placement heuristics

• A result whose absence would change the accept/reject argument goes in the main part, even as a two-line proof sketch pointing to the appendix.
• A step a skeptical reviewer must verify to trust a theorem goes in the appendix — in full, with the theorem restated so the appendix reads without page-flipping.
• Anything executable goes in the ZIP, plus a pointer sentence in the paper so its existence is on the reviewed record.
• Never cite the ZIP as evidence for a claim ("see supplementary code for the full results"): reviewers are entitled to ignore it, so the claim becomes unbacked.

Appendix architecture that works here

Statisticians and probabilistic-ML researchers navigate by statement numbers. Structure for that reading style:

• One appendix section per theorem or per experiment family, in main-text order.
• Restate every assumption with its original label (A1, A2…) before using it; an appendix proof invoking an unlabeled assumption is where correctness objections start.
• Give experimental appendices the same table discipline as the body: full hyperparameter grids, per-dataset results, per-seed dispersion.
• Cross-reference forward from the body precisely ("proof in App. B.2", not "proofs in the appendix").

% Single-PDF assembly, UAI style: appendix follows references in the same document
\bibliography{refs}          % references end the main part
\clearpage
\appendix
\section{Proofs for Section 3}\label{app:proofs}
\paragraph{Assumption A2 (restated).} The mixing weights are bounded away from zero.
\begin{proof}[Proof of Theorem 3]
  ... % full argument; the body kept only the sketch
\end{proof}
\section{Experimental Protocols}\label{app:protocols}
% grids, seeds, per-dataset tables, compute

ZIP assembly under the 50 MB cap

• Ship generators and configs rather than bulky outputs; include checksums for anything regenerable.
• Structure by claim, not by code module: a reviewer looks for figure2/, not for src/utils/.
• Keep the archive as anonymous as the PDF: no .git history, no notebook author metadata, no cluster paths, no account-bound experiment-tracker URLs.
• Test the unzip-and-run path on a clean machine; a broken optional artifact is worse than no artifact, because it documents carelessness.

Pre-merge integrity checklist

Run after the last content edit, before assembling the final PDF:

• Every theorem/lemma/proposition statement in the body is character-identical to its restatement in the appendix (diff them mechanically, not by eye).
• Every assumption label used in an appendix proof is defined in the body.
• Every "App. X.Y" reference in the body resolves to an existing section after the latest appendix reshuffle.
• Appendix figures and tables are numbered in a scheme that cannot collide with the body's (e.g., Table 7 vs Table A.3 — pick one convention).
• No appendix section silently introduces a claim absent from the body; reviewers may treat body-less claims as unreviewable.
• The references section sits between main part and appendix per the current template's convention.
• Page count of the main part re-verified after the merge; late appendix edits sometimes push floats forward into page 8.
• The ZIP's README points into the appendix ("protocol details in App. C") and not vice versa for anything decision-critical.

Budgeting the 50 MB

A workable allocation when the archive is tight: code and configs are effectively free (rarely over 5 MB); reserve the bulk for whichever single item most de-risks your weakest claim — a gold-standard posterior sample set, a preprocessed data extract with checksums, or trained-model files for the headline table. Everything else ships as a generator script plus a hash of its output. If two heavyweight items compete, keep the one a reviewer cannot regenerate in minutes.

Failure patterns seen at single-PDF venues

• Overflow laundering: pushing the method description itself into the appendix to make the 8-page cap. Reviewers score clarity on the main part; an unintelligible body with a beautiful appendix fails.
• Appendix drift: editing a theorem in the body during deadline week without touching its appendix proof. Diff statement text mechanically before upload.
• Reference squeeze: trimming citations to save space, forgetting references are outside the 8-page cap and cost nothing.
• ZIP-only ablations: the reviewer asks "why no ablation?" and the answer "it was in the ZIP" wins no points, because the rules told them they could skip it.

Output format

[Tier map] main <pages> / appendix <pages> / ZIP <MB>
[Misplaced content] <items in the wrong tier, with target tier>
[Appendix integrity] statements match body? assumptions labeled?
[ZIP status] runs clean? anonymous? claim-indexed?
[Pointer sentences] <paper locations announcing appendix and ZIP contents>