技能 职场通用 学术证明可复现性规范

学术证明可复现性规范

v20260724
podc-reproducibility
本指南提供学术证明的可复现性指导,确保论文的理论结论可以被独立验证。核心要求包括:撰写自包含的证明附录、在模型方框中明确声明所有假设(构建可信赖的“契约”),以及同步会议论文和完整版本,杜绝知识盲区,保证结果的严谨性。
获取技能
343 次下载
概览

PODC Reproducibility

"Reproducibility" at a proofs venue means a reader can independently verify the theorem. PODC has no artifact track and no ACM badge program, so there is nothing to package for evaluators — the deliverable that makes your result checkable is a self-contained proof and an honest model box. This skill translates the open-science instinct into what actually matters at PODC.

What "checkable" means here

A PODC result is reproducible when a competent reader, given only your submission and its full version, can:

  1. Read the model box and know exactly what network, timing, fault, adversary, randomness, and cost-measure assumptions are in force.
  2. Follow every proof to its base cases and cited lemmas without needing an external file, a private note, or a "details omitted".
  3. Confirm that each theorem holds in the stated model and nowhere relies on a stronger assumption.

The self-contained proof appendix

Because the 10-page merits budget cannot hold full proofs, the full version (submitted as the paper, and later posted to arXiv) carries them. Make it self-contained:

[Every lemma stated]     each lemma the main theorem uses appears with a full proof, or a precise
                         citation to an external theorem (statement quoted, not just referenced)
[No "it is easy to see"] for any nontrivial step; either prove it or mark it explicitly routine
[Base cases present]     inductions and recursions have their base cases proved
[Invariants proved]      each named invariant is shown to hold initially and be preserved
[Notation defined once]  a single notation table the appendix and body share
[Cross-references exact]  "by Lemma 3.2" points to the right statement in both body and full version

A proof that says "the remaining case is symmetric" is fine only if it truly is; reviewers check the "symmetric" cases that turn out not to be.

The model/assumption box is the reproducibility contract

The single most common reason a PODC result fails to reproduce is an assumption used but not declared. Treat the model box (see podc-writing-style) as a contract:

  • Every assumption the proofs use must be in the box. If a proof needs a shared coin, the box must grant one. If it needs FIFO channels, the box must say so.
  • Nothing in the box may be silently strengthened mid-proof (asynchrony becoming "eventually synchronous" without invoking GST; an oblivious adversary becoming "non-adaptive within a phase").
  • If a result holds only in a parameter regime, the box or the theorem states the regime.

Run the assumption audit from podc-experiments and confirm the box covers everything the proofs consume.

Optional simulations: transparent, not load-bearing

If the paper includes a simulation (many do not), make it reproducible as an illustration:

  • State the parameters, ranges, number of trials, and random seeds; pin the environment (language/version) in the full version or a linked bundle.
  • One-command reproduction is a nice-to-have, not a requirement — the result stands on the proof.
  • Keep any simulation repository/author-page link out of the anonymized submission (lightweight double-blind), and add it only in the camera-ready / arXiv version.
  • Never present simulation output as if it verified the theorem; label it "illustrative."

Keeping the full version in sync

  • The camera-ready (≤10 proceedings pages) and the arXiv full version must state identical theorems; only the proof detail differs. A theorem tightened in one and not the other is a correctness-record bug.
  • When you fix a proof after acceptance, propagate the fix to the arXiv version and note the revision — the community reads arXiv for the proofs.
  • Cross-reference the full version from the proceedings paper ("full proofs in the full version [arXiv:...]") so a reader can always reach the complete argument.

Common failures

  • "Proof omitted" with no full version — the result is then unverifiable; unacceptable at a proofs venue.
  • Assumption used but not in the model box — the reproducibility contract is broken.
  • Body and full version disagree on a theorem statement or constant.
  • Simulation dressed as verification — a plot presented as if it proved correctness.
  • De-anonymizing link in the submission — a simulation repo that reveals authorship under double-blind.

Output format

[Proof appendix] self-contained? every lemma proved/precisely cited? base cases + invariants present?
[Model box contract] every assumption the proofs use is declared and never silently strengthened?
[Body <-> full version] theorem statements identical; full proofs reachable from the proceedings paper?
[Simulation] absent / illustrative-with-seeds-and-ranges; no de-anonymizing links at review time?
[Fix queue] <omitted proofs to supply; undeclared assumptions; sync mismatches>
信息
Category 职场通用
Name podc-reproducibility
版本 v20260724
大小 5.24KB
更新时间 2026-07-29
语言