Webconf Reproducibility

Web Conference Reproducibility

Reproducibility at this venue has a problem no offline-ML venue has: the object of study mutates. Pages die, APIs close, ranking systems retrain, platform policies change what may be collected at all. A Web Conference paper is reproducible to the degree that it pins what can be pinned and measures what cannot. The 2026 CFP's sanctioned home for this material is the optional appendix — "details on reproducibility, proofs, pseudo-code" — inside the same 12-page PDF, which reviewers are not obliged to read; so the reproducibility claims go in the main 8 pages and the reproducibility mechanics go in the appendix.

Three reproducibility regimes

Most reviews go wrong when a paper claims regime-1 language ("fully reproducible") for regime-2 or regime-3 evidence. Classify every experiment in the paper into a regime and phrase its claim accordingly; the honest sentence "results on the live platform are audit-reproducible but not replay-reproducible" has never sunk a strong paper.

Pinning the decaying regime

• Date-stamp everything: crawl window, API version, model checkpoints of any third-party service used (an LLM API call in 2025 is not the same function in 2026 — record model name and version string).
• Checksum the corpus at collection time; publish per-file SHA-256 in a manifest so drift is detectable, not just suspected.
• Prefer citable snapshots: Common Crawl snapshot IDs, Internet Archive captures, Wikipedia dumps with dates — these convert a decaying corpus into a frozen one for everyone downstream.
• Separate collection from analysis in the codebase, so a future team can re-run analysis on your frozen snapshot even when re-collection is impossible.

Determinism audit for the modeling stack

# Repro header every experiment script in the artifact should share
import os, random, numpy as np, torch

SEED = int(os.environ.get("RUN_SEED", 17))
random.seed(SEED); np.random.seed(SEED); torch.manual_seed(SEED)
torch.use_deterministic_algorithms(True)          # surfaces nondeterministic ops
os.environ["CUBLAS_WORKSPACE_CONFIG"] = ":4096:8" # required by some CUDA GEMMs

# Log the things people forget to log:
#   graph/dataloader shuffling seeds, negative-sampling seeds,
#   train/val/test split hash, library versions, GPU model, wall-clock.

Web-specific nondeterminism deserves explicit lines in the appendix: crawl ordering, deduplication thresholds, timezone normalization of timestamps, and — for graph papers — node ID remapping, which silently reorders neighbor sampling.

Temporal honesty

Web data is time-indexed, and the venue's reviewers increasingly check for temporal leakage: random splits over user-item interactions or evolving graphs let the model train on the future. The reproducibility appendix should state the split rule (e.g., "train < 2025-06-01 ≤ test"), not just percentages, and the artifact should ship the split-generation code rather than opaque index files alone. If the paper uses a random split on temporal data for comparability with prior work, say so and add one temporal split as a robustness check — this one-sentence-plus-one-table addition preempts the most common modern objection.

Vignette: the benchmark that dissolved

A 2024-vintage misinformation dataset distributes tweet IDs for rehydration. By the time a team builds on it for a WWW submission, 38% of the tweets are deleted, suspended, or geo-blocked — and deletion is not random: the most-reported content vanishes first. Naively rehydrating and comparing against the original paper's numbers silently changes both the task and the class balance. The regime-honest handling, which fits in four appendix sentences plus one table column: report the rehydration date and survival rate, compare label distributions between the original and surviving corpus, rerun the strongest baseline on the surviving subset so all comparisons share one corpus, and phrase cross-paper comparisons as indicative rather than head-to-head. Reviewers do not penalize decay — it is the field's shared condition — but they increasingly penalize pretending it did not happen.

What goes where

1. Main 8 pages: one reproducibility paragraph — regime classification, snapshot/DOI pointers, seed policy, split rule. Claims a reviewer must weigh cannot hide past page 8.
2. Appendix (within the 12): hyperparameter tables, environment details, collection protocol, per-dataset manifests, negative results of tuning.
3. Artifact (see webconf-artifact-evaluation): everything executable, the manifest with checksums, and the recrawl/dead-link accounting script.

A placement corollary for review strategy: because the appendix is optional reading, a reviewer who doubts reproducibility may score the doubt without opening Appendix B. The main-text paragraph therefore needs one forward pointer with content — "seeds, environment, and the full collection protocol are in App. B; the artifact reproduces Table 2 with one script" — so the doubt has an address before it becomes a score.

Pre-submission reproducibility gate

• Every experiment classified frozen / decaying / unreplayable, phrased to match.
• Collection dates, API versions, and third-party model versions stated.
• Split rules explicit; temporal data has a temporal split somewhere.
• Seeds and repetition counts reported; variance shown where runs repeat.
• A stranger with the appendix plus artifact could rebuild the headline table or, failing that, audit the protocol end to end.

Output format

[Regimes] frozen=<experiments> decaying=<...> unreplayable=<...>
[Pinning] dates/checksums/snapshots: complete / gaps <where>
[Temporal] split rule stated? leakage risk? robustness split present?
[Determinism] seed policy + environment logged: yes/no
[Placement] claims in main text, mechanics in appendix: verified
[Honesty edits] <sentences whose reproducibility claim overshoots the regime>

Source: brycewang-stanford/Awesome-Journal-Skills → The-Web-Conference-Skills/skills/webconf-reproducibility/SKILL.md