---
name: brycewang-stanford/acl-supplementary
source: https://app.decimal.ai/s/brycewang-stanford-acl-supplementary@1/SKILL.md
source_sha256: 49ebe0f1fbd3
---

# ACL Supplementary

Use this when splitting an ACL paper between body, appendix, and archive. The
governing ARR principle: **reviewers are not required to consider material in
appendices or supplements**, so anything decision-critical that lives only
there is effectively invisible.

## The ACL page anatomy

```text
[ content pages: 8 long / 4 short ]   <- the reviewed argument lives here
[ Limitations (REQUIRED, unlimited) ] <- after conclusion, outside page count
[ Ethics statement (optional) ]
[ References (unlimited) ]
[ Appendices (unlimited, same PDF) ]  <- optional reading for reviewers
+ separate .tgz/.zip archive          <- software / data supplement
```

Missing Limitations is a desk-reject condition; treating it as one throwaway
sentence is a review-stage penalty even when it passes the gate.

## What must not leave the body

- The main results table and the headline comparison.
- Task definition and enough of the method that a reviewer can judge novelty.
- At least a summary of the error analysis — a pointer-only error analysis
  reads as not having one.
- Human-evaluation design in one paragraph: raters, items, agreement.
- The experimental setup at reproduction-outline level; full grids can go down.

## What appendices are good at

- Full prompt texts and few-shot exemplars (reference them per experiment).
- Complete hyperparameter tables and search ranges.
- Per-language / per-dataset breakdowns behind an averaged headline number.
- Annotation guidelines and interface screenshots.
- Extended qualitative examples and additional ablations.
- Proofs or derivations for the occasional formal result.

## Limitations section that actually works

| Weak pattern | Stronger ACL pattern |
|---|---|
| "Results may not generalize" | Name the languages, domains, and model scales actually tested and the nearest untested regime |
| "LLMs can hallucinate" | State which conclusions depend on a specific model snapshot and API behavior |
| Silent on data | Note license constraints, demographic skew, or collection-window bias in the corpora used |
| Written last-minute | Mirrors the risks reviewers will find anyway, defusing them on your terms |

ACL's policy explicitly instructs reviewers not to punish honest limitations,
which makes this section the cheapest goodwill in the whole submission.

## Archive rules and hygiene

- Upload software/data as .tgz or .zip in the OpenReview fields; personal
  cloud-storage links are barred, and any external page must be anonymous and
  untracked.
- Strip authorship trails: git history, notebook metadata, absolute paths with
  usernames, license headers, README contact lines.
- Test the archive on a clean machine: it must unpack, the README must state
  what maps to which table, and nothing should require credentials just to read.
- Include model outputs where feasible so reviewers can verify scoring without
  compute (see `acl-reproducibility`).

## Body-vs-appendix vignette

A long paper introduces a retrieval-augmented QA method with results on six
benchmarks in three languages. Body: method figure, main table (six benchmarks
averaged + per-language block), two-paragraph error analysis, one ablation that
carries the mechanism claim. Appendix: full per-benchmark tables, prompts,
retrieval index details, remaining ablations, annotation guidelines for the
human study. Archive: code, prompts as files, and all model outputs. The test:
a reviewer who never scrolls past the references can still reconstruct and
believe every claim in the abstract.

## Appendix ordering convention that reviewers navigate well

- A: full experimental setup (models, hyperparameters, hardware, budgets).
- B: prompts and few-shot exemplars, one subsection per experiment.
- C: complete results — per-dataset, per-language, per-seed tables behind
  every averaged number in the body.
- D: annotation materials — guidelines, interface, pay, agreement detail.
- E: additional analyses and ablations, each forward-referenced from the
  body at least once (unreferenced appendix content is invisible).
- F: qualitative examples, marked as random or curated — say which.

Number tables and figures continuously with the body so the author response
can cite "Table 9" unambiguously during the discussion phase.

## Ethics statement: when to write one

Write it when the paper involves human subjects or annotators, scraped
user-generated content, demographic inference, dual-use capability, or
release of models/data with realistic misuse paths. Skip it when nothing
applies — the Responsible NLP checklist already covers the routine cases,
and a padded statement invites the very scrutiny it fails to answer.
It shares the unlimited space after the conclusion with Limitations.

## Size and dependency guardrails

- Keep the archive small enough to download on conference-hotel wifi;
  reviewers abandon multi-gigabyte supplements unopened.
- No credentials, no API keys, no gated-model weights — describe access
  paths instead of shipping secrets.
- Pin dependency versions in one environment file; "latest transformers"
  is a different codebase every cycle.
- If data cannot be shared (license, privacy), include the loader code and
  a synthetic sample with identical schema so scripts still run.

## Output format

```text
[Split status] sound / body-overloaded / appendix-dependent
[Limitations quality] substantive / ritual / missing
[Must-move-up] <decision-critical items currently below the fold>
[Archive check] <format/anonymity/clean-machine findings>
[Reviewer-blind spots] <claims visible only outside the body>
```