Content & Media Creation · lllllllama/rigorpilot-skills

paper-context-resolver

Name: paper-context-resolver
Author: lllllllama/rigorpilot-skills

Rigor Paper Context helper for README-first deep learning repo reproduction. Use only when the README and repository files leave a narrow reproduction-critical gap and the task is to resolve a specific paper detail such as dataset split, preprocessing, evaluation protocol, checkpoint mapping, or runtime assumption from primary paper sources while recording conflicts. Do not use for general paper summary, repo scanning, environment setup, command execution, title-only paper lookup, or replacing README guidance by default.

L1 Basic · Source registration onlyV1 Commit-PinnedDual-pass Audit

Control levels describe skill capabilities and suggested control measures; they do not represent quality or safety certifications, nor do they dictate usability. Usage decisions rest entirely with users and administrators. Different agent hosts apply varying constraints; do not assume hosts provide automatic protection. Third-party skill copyrights belong to their original authors. Use at your own risk.

1How It Works & Mechanism

Pure prompt-based helper: The entire skill consists solely of SKILL.md + one reference document + one Codex inventory, containing zero scripts or executable code; capability relies entirely on four discipline checklists ('When to apply / When not to apply / Inputs / Outputs') to constrain model behavior.

skills/paper-context-resolver/SKILL.md

This skill is optional. - This skill is helper-tier and should usually be orchestrator-invoked. - It supplements README-first reproduction. - It does not replace the main orchestration flow. - It does not summarize the whole paper by default.

What is happening here: Defining a hard boundary to strictly lock the skill's scope of authority. Because this helper is invoked during specific phases by the main orchestrator (ai-research-reproduction), ambiguous boundaries would cause it to be misused as a general-purpose paper summarizer—the description's 'Do not use for general paper summary' represents the natural-language formulation of this very constraint.

Trigger discipline is enforced via a 'question type whitelist': Activation is permitted exclusively for five specific question types—dataset version/split, preprocessing, evaluation protocol, checkpoint mapping, and runtime assumption—accompanied by explicit good/bad trigger examples in the reference documentation for contrast.

skills/paper-context-resolver/references/paper-assisted-reproduction.md

## Allowed question types - dataset version or split - preprocessing or postprocessing details - evaluation protocol details - checkpoint or model variant mapping - critical runtime assumptions

The Trigger discipline section of the same document provides counterexamples: A Good trigger is 'README is ambiguous about the evaluation split. Check the linked paper and record the conflict if needed.', whereas a Bad trigger is 'Summarize this paper.' / 'Find the paper for this repository title and explain it.' / 'Tell me everything the paper changed compared with prior work.'. This transforms a 'narrow question' from a slogan into a verifiable acceptance criterion.

Information sourcing follows a strict priority chain, treating Google Scholar solely as a locator rather than an authoritative source of evidence: Paper links in README > Official project page / Official paper page > Primary records on arXiv / OpenReview > Scholar (used strictly to locate primary sources).

skills/paper-context-resolver/references/paper-assisted-reproduction.md

1. paper link explicitly provided by the README 2. official project page or official paper page 3. arXiv or OpenReview primary paper record 4. Google Scholar only to help locate the primary source

What is happening here: Suppressing the common degraded pattern of treating search engine snippets as evidence, mandating that conclusions be grounded in primary literature that can be manually audited; simultaneously prioritizing links provided in the README to prevent taking reproducibility intent out of the repository authors' hands.

Conflict resolution is a hard constraint: Silent overwriting of the README is prohibited when inconsistencies exist between the README and the paper; conflicts must be logged explicitly stating what each source claims, preserving a README-first stance in the final report.

skills/paper-context-resolver/references/paper-assisted-reproduction.md

If README and paper disagree: - do not silently replace README - record the conflict - explain which source says what - preserve the README-first policy in the final report

This is the only behavioral rule in this skill with material consequences, distinguishing it from simply 'letting the model search papers online': external literature is not allowed to be elevated to a new reproduction baseline, serving only as annotated supplementary evidence.

Invocation prerequisites are explicitly written into the opening line of the reference document: consulting the paper is allowed only after inspection of the README / repository files produces a specific unanswered question.

skills/paper-context-resolver/references/paper-assisted-reproduction.md

Only invoke this step after README and repo inspection already produced a concrete unanswered question.

Acts as redundant double protection alongside rule #4 under 'When to apply' in SKILL.md ('There is already a concrete reproduction question to answer.')—preventing the model from diving into the paper right away.

The output protocol mandates distinguishing direct evidence from inferences, and explicitly providing README–paper conflict notes when applicable.

skills/paper-context-resolver/SKILL.md

## Output expectations - narrowed source list - reproduction-relevant answer only - explicit README-paper conflict note when applicable - clear distinction between direct evidence and inference

The 'narrowed source list' requirement demands returning the inventory of sources actually utilized (rather than everything retrieved), maintaining consistency with the repository's broader evidence-linked narrative.

2Core Capabilities

01Narrowed evidence retrieval for reproduction-critical gaps (answers exclusively: dataset split / preprocessing / evaluation protocol / checkpoint mapping / runtime assumption)

02Priority-based paper source resolution (README links prioritized; Scholar downgraded to a locator tool)

03Log-driven resolution of README–paper conflicts (silent overwriting prohibited)

04Trigger discipline: Good vs. bad example contrasts, explicitly rejecting requests such as 'summarize paper' or 'find paper by title and explain'

05Default exclusions list (full-text summaries / novelty explanations / related work discussions / guessing 'best settings' without reproduction basis)

06Tiered evidence output: Explicit separation of direct evidence vs. inferences, alongside a narrowed source list

07Addressable helper stage for main orchestrators (integratable into skill chains via --include-paper-gap; recorded as a blocked stage when questions or sources are missing)

3External Dependencies

Kind	Dependency / Endpoint
network	一手论文站点（指令层依赖，非代码调用：由宿主模型的检索/阅读能力访问）

4Recommended Governance & Evidence

Suggested Controls for Safe Operation

For Agent Runtime

Can be auto-loaded

For Administrators

Evidence & Triggers (File Paths & Excerpts)

SKILL.md: Pure prompt instructions and rule text; no script execution or external side effects.

Audit Notes: Zero scripts, zero network calls, zero credential access, and zero local disk writes (a second token scan across the directory yielded zero matches for URL/curl/wget/subprocess/os.environ/open(, and no code files exist within the directory). It functions purely as a discourse rule: strictly governing 'when to consult papers, to what depth, and how conflicts must be documented.' The only requirements placed on the user are screening for potentially adversarial or misleading content in parsed paper text (a host-level generic risk, not introduced by this skill) and preventing confusion between arXiv claims and README instructions—a failure mode this skill explicitly mitigates via its conflict handling rules.

5Second-Pass Verification

[ok] skill 路径（任务书标『待定位』） — 实际相对路径 skills/paper-context-resolver（SKILL.md 在该目录根）。仓库内另无同名副本；.claude/skills 之类标准位置不存在于本仓库。
[ok] 零脚本 / 零网络 / 零凭证 / 零写入 — 全目录 token 扫描零命中；文件清单三个文件全为 markdown/yaml；SKILL.md 全文无 bash 代码块、无工具调用指令。
[ok] 『不静默覆盖 README』规则真实存在 — references/paper-assisted-reproduction.md 的 Conflict rule 段逐字含 'do not silently replace README' 与 'preserve the README-first policy in the final report'。
[ok] 取源优先级链存在且 Scholar 被降级 — Source order 四行逐字核对无误；第 4 条限定词 'only to help locate the primary source' 明确排除把 Scholar 当证据。
[ok] helper 定位与仓库机制一致（注册表 + 编排器） — skill-registry.json tier/lane='helper'、output_mode.kind='paper_gap'；orchestrate_repro.py plan_skill_chain() 在 include_paper_gap 时追加本 slug，并在缺参数时记 blocked 阶段（'A narrow paper question and authoritative paper source were not supplied.'）。
[ok] 触发边界（不该被误触发的场景） — tests/trigger_cases.json 中本 slug 在数十条用例里出现在 forbidden 列表，仅两条用例把 'paper-context-resolver' 列为 expected_top；与 description 的 'Use only when…' 一致。
[ok] Codex 清单无 policy.allow_implicit_invocation 字段 — agents/openai.yaml 仅有 interface 块（display_name/short_description/default_prompt），没有 policy 块——即 Codex 侧未关闭隐式调用；这与 grill-me 范式不同，属事实记录，非缺陷。
[ok] README 帮助项定位 — README.md:160 'Two helpers support orchestration: `repo-intake-and-plan` and `paper-context-resolver`.'；README 顶部 badge 还把 paper-context-resolver 作为 Skillselion Top 100 展示项。

6Verdict & Scenarios

Transforms 'checking papers' into a bounded evidence-gathering action: Whitelisted question types + strict sourcing priority chain inherently suppress unconstrained model-generated literature reviews.

Auditable conflict handling: Silent overwriting is prohibited when the README contradicts the paper; claims from both sides must be recorded—the skill's strictest rule and core value-add over generic paper search prompts.

Zero code and zero network access, portable across hosts with virtually no attack surface: The three files combined contain no executable statements, requiring no trust in external scripts.

Clear division of labor with sister helpers, explicitly addressable by orchestrators: One scans the repository while the other inspects the paper without overlap; both expose missing parameters as blocked stages rather than making wild guesses.

Explicitly tiered output: Distinguishes direct evidence from inferences, mitigating the risk of citing speculative paper reasoning as established fact.

Best for: Ideal for researchers or agent orchestration workflows reproducing deep learning repositories based on a README who are blocked by a specific reproduction detail (split / preprocessing / evaluation protocol / checkpoint mapping / runtime assumption); also well-suited as an optional stage called by a broader reproduction agent (--include-paper-gap). Maximum value is achieved when the target repository is an official paper implementation and the README explicitly links to the paper.

Not for: Not suitable for users seeking to quickly grasp a paper (general paper reading/summarization tools should be used; Default exclusions explicitly bar full-text summaries and related work discussions); not suitable for scenarios lacking prior README/repository inspection with only a paper title available (reference: 'Only invoke this step after README and repo inspection already produced a concrete unanswered question.'); not suitable for tasks requiring verification of the paper's original metrics themselves (the skill explicitly disclaims paper-level reproduction assertions; orchestrator output specifies 'this result alone does not establish paper-level reproduction'); nor is it intended for downloading/parsing PDFs or executing code.

Install Agent-Ready

① Verified Mirror 2026-09-15

Option A · Direct DownloadDownload paper-context-resolver.tar.gz

sha256: e6d72d22dd174312…

Option B · Copy Guide for Agent

install.json

Agents automatically download and verify mirrors from this JSON guide.

② Upstream GitHub

GitHub Repo ↗

Pinned to commit bd91195ad7.