Skills ai-research-reproduction

📦

ai-research-reproduction

Name: ai-research-reproduction
Author: lllllllama

Content revision r2 High Risk ⚙️ External commands📁 Filesystem access

Reproduce AI Research Repositories

Deep learning repositories often lack a clear, trustworthy reproduction path. This skill selects a minimal documented target and records execution evidence and deviations.

Supports: Claude Codex Code(CC)

⚠️ 38 Poor

Stop for confirmation before installing.

Review the plan and obtain explicit user consent before changing files.

Install with my Agent

Copy this request to your Agent. It includes the canonical Skill page and manifest.

Agent request

Review the Skillstore skill "ai-research-reproduction" from https://skillstore.io/skills/lllllllama-ai-research-reproduction.md and its manifest at https://skillstore.io/api/skills/lllllllama-ai-research-reproduction/manifest. Verify the artifact. Stop and obtain explicit user consent before installing or changing files.

Your Agent should still show its plan and request any confirmation required by the security policy.

Agent-readable resources

Use these links when an AI agent, crawler, or script needs clean context instead of reading the full page.

Markdown detail GET /skills/lllllllama-ai-research-reproduction.md Signed manifest GET /api/skills/lllllllama-ai-research-reproduction/manifest Signed lockfile GET /api/skills/lllllllama-ai-research-reproduction/lockfile

Test it

Using "ai-research-reproduction". Plan a minimal reproduction for an image-classification repository without running commands.

Expected outcome:

Selected target: documented single-image inference. Evidence lists the README section, required checkpoint, setup assumptions, and the command awaiting approval.

Using "ai-research-reproduction". Run the smallest documented evaluation and explain why it failed.

Expected outcome:

Outcome: blocked before evaluation because the checkpoint was absent. The report records the attempted command, error excerpt, asset requirement, and next safe action.

Using "ai-research-reproduction". Check whether a short training run supports the paper claim.

Expected outcome:

Outcome: partial evidence only. Startup succeeded, but dataset split and metric settings remain unverified, so no comparability claim is made.

Security Audit

High Risk

v5 • 8/8/2026 Open versioned report

Most static findings are Markdown backticks, prose, safe temporary-file handling, or fixed argument-array subprocess calls. Confirmed risks include executing README-derived commands, loading files outside the skill directory, accessing a hidden home file, and dynamically importing external sibling code.

Files scanned

2,038

Lines analyzed

Review items

False positives ignored

Confirmed security concerns (1)

High

scripts/orchestrate_repro.py:23-31

Unaudited Dynamic Module Execution

The orchestrator resolves a sibling lessons_store.py outside the skill package and executes it through importlib without provenance or integrity verification.

The code directly constructs an out-of-package path and calls exec_module. The external module is not included in the scanned file set.

Capability review items (9)

These are real local capabilities that may be expected for this skill, so they require review but are not counted as confirmed malicious behavior.

High

scripts/orchestrate_repro.py:357

Python subprocess.run

result = subprocess.run(

This executes a command extracted from an untrusted target repository README. The opt-in flag limits surprise, but there is no allowlist, sandbox, or final command approval.

High

SKILL.md:20

Path traversal sequence

`../../references/agent-operating-principles.md`, then load

The skill instructs the agent to load a path two directories above its package. This expands the trust boundary beyond marketplace-audited files.

High

SKILL.md:21

Path traversal sequence

`../../references/research-rigor-principles.md` and

The required policy is resolved outside the skill directory. External content could alter behavior without changing this audited package.

High

SKILL.md:22

Path traversal sequence

`../../references/deep-learning-experiment-principles.md` when scientific meaning, comparability, or

This directs the agent to read an out-of-package experiment policy. The referenced content is outside the report's scanned file set.

High

SKILL.md:121

Path traversal sequence

- Load `../../references/research-rigor-principles.md` before making comparability, contribution, or

The instruction repeats an out-of-package policy read before research claims. The external file is not covered by this audit.

High

SKILL.md:122

Path traversal sequence

- Load `../../references/deep-learning-experiment-principles.md` when dataset, split, metric, checkp

The agent is told to load a policy through parent-directory traversal. This creates an unaudited instruction dependency.

High

SKILL.md:123

Path traversal sequence

- Consult `~/.rigorpilot/PERSONAL_RIGOR.md` if present, under `../../references/continuous-learning-

The line references both a home-directory file and a parent-relative external policy. These reads cross the packaged skill boundary.

High

SKILL.md:123

Hidden file in home directory

- Consult `~/.rigorpilot/PERSONAL_RIGOR.md` if present, under `../../references/continuous-learning-

The skill explicitly instructs the agent to consult a hidden user-home file. Its contents may be private and are not marketplace audited.

High

SKILL.md:123

Hidden file access

- Consult `~/.rigorpilot/PERSONAL_RIGOR.md` if present, under `../../references/continuous-learning-

The requested hidden-file read introduces user-specific, unaudited instructions into the agent context. Explicit consent and scope controls are absent.

Risk Factors

⚙️ External commands (50)

📁 Filesystem access (11)

scripts/annotate_readme.py:104 scripts/orchestrate_repro.py:12 scripts/orchestrate_repro.py:90 SKILL.md:20 SKILL.md:21 SKILL.md:22 SKILL.md:121 SKILL.md:122 SKILL.md:123 SKILL.md:123 SKILL.md:123

Audited by: codex View Audit History →

Share & cite this report

Share the versioned assessment report, neutral badge, embed card, and citations. Skillstore reports evidence without deciding whether this Skill is safe.

Open versioned report

Copy report link

https://skillstore.io/skills/lllllllama-ai-research-reproduction/audits/5?utm_source=security_passport&utm_medium=share&utm_campaign=versioned_report

Markdown badge

[![Skillstore security assessment](https://skillstore.io/badges/skills/lllllllama-ai-research-reproduction/security.svg)](https://skillstore.io/skills/lllllllama-ai-research-reproduction?utm_source=security_passport_badge)

HTML badge

<a href="https://skillstore.io/skills/lllllllama-ai-research-reproduction?utm_source=security_passport_badge"><img src="https://skillstore.io/badges/skills/lllllllama-ai-research-reproduction/security.svg" alt="Skillstore security assessment" loading="lazy"></a>

Embed card

<iframe src="https://skillstore.io/embed/skills/lllllllama-ai-research-reproduction.html" title="Skillstore Security Assessment" sandbox="allow-popups allow-popups-to-escape-sandbox" loading="lazy" referrerpolicy="no-referrer" width="420" height="180"></iframe>

Academic citations (APA · BibTeX · CFF)

APA citation

lllllllama. (2026). ai-research-reproduction security audit report (audit version 5) [Author version unspecified]. Skillstore. https://skillstore.io/skills/lllllllama-ai-research-reproduction/audits/5

BibTeX citation

@techreport{lllllllama-lllllllama-ai-research-reproduction-2026,
  author = {lllllllama},
  title = {ai-research-reproduction security audit report (audit version 5)},
  institution = {Skillstore},
  year = {2026},
  number = {5},
  url = {https://skillstore.io/skills/lllllllama-ai-research-reproduction/audits/5},
  note = {Author version unspecified}
}

CITATION.cff

cff-version: 1.2.0
message: "If you use this Skill, cite its author and this versioned security audit report."
title: "ai-research-reproduction security audit report (audit version 5)"
version: "unspecified"
type: report
authors:
  - name: "lllllllama"
date-released: "2026-08-08"
url: "https://skillstore.io/skills/lllllllama-ai-research-reproduction/audits/5"
identifiers:
  - type: other
    value: "skillstore:lllllllama-ai-research-reproduction:audit:5"
    description: "Skillstore immutable audit report identifier"

Skillstore Score

Why this score Evidence Confidence: Medium

Architecture

Maintainability

Content

Community

Spec Compliance

What You Can Build

Verify a published repository

Select and run the smallest documented inference or evaluation target, then record evidence and blockers.

Audit reproducibility before review

Compare repository guidance, observed behavior, and protocol deviations before assessing a research claim.

Prepare a team handoff

Create standardized reproduction outputs that preserve commands, assumptions, status, and next decisions.

Try These Prompts

Plan a minimal reproduction

Inspect [repository path]. Select the smallest documented inference or evaluation target. Do not execute commands. Produce a reproduction plan with evidence.

Run a documented smoke test

Reproduce [repository path] using the smallest documented target. Show the selected command before execution. Stop after the first trustworthy result.

Audit a partial reproduction

Review [repository path] and [existing output path]. Identify blockers, assumptions, protocol deviations, and the lowest-risk next action. Preserve repository semantics.

Assess training comparability

Evaluate whether [repository path] can reproduce [paper result]. Verify dataset, split, checkpoint, preprocessing, metric, and training assumptions. Require approval before fuller training.

Best Practices

Review every extracted command and its working directory before allowing execution.
Prefer the smallest documented target that can produce meaningful evidence.
Record every assumption, patch, and protocol deviation before comparing results.

Avoid

Do not run full training before environment, assets, and startup behavior are verified.
Do not change model semantics or metrics merely to obtain a successful result.
Do not treat generated reports as proof when key experimental details remain unverified.