# Explore Novel Deep Learning Research Candidates

Researchers struggle to systematically explore and rank novel deep learning ideas with scientific rigor. This skill provides auditable candidate exploration with idea gating, fair comparison, and governed experiment workflows on top of current_research.

## Install

```bash
npx skillstore add lllllllama/ai-research-explore
```

## Metadata

- - Slug: lllllllama-ai-research-explore
- - Version: 1.0.0
- - Author: lllllllama
- - GitHub username: lllllllama
- - License: MIT
- - Repository: https://github.com/lllllllama/rigorpilot-skills/tree/main/skills/ai-research-explore/
- - Ref: main
- - Supported tools: Claude, Codex, Claude Code
- - Risk level: medium
- - Risk factors: external\_commands, network, env\_access, scripts, filesystem
- - Quality score: 50
- - Quality tier: warning
- - Public page: https://skillstore.pages.dev/skills/lllllllama-ai-research-explore
- - Manifest: https://skillstore.pages.dev/api/skills/lllllllama-ai-research-explore/manifest

## Capabilities

- Analyzes repository structure to understand the research codebase before candidate exploration
- Generates and ranks research candidate ideas with explicit gating criteria and score breakdowns
- Maps candidate ideas to source references from arXiv, DOI, and GitHub repositories
- Evaluates execution feasibility of proposed research candidates with bounded resource estimates
- Maintains auditable SCIENTIFIC\_CHANGELOG and COMPARABILITY\_REPORT artifacts for each campaign
- Supports cache-first source lookup with local curated literature and optional web providers

## Use Cases

- Systematic Idea Exploration: A researcher wants to explore several candidate improvements to their current model. They use this skill to generate ranked ideas with feasibility scores and scientific changelogs.
- Fair Comparison Workflows: A research team needs to compare candidate approaches against their SOTA reference with controlled variables and auditable evidence. The skill enforces frozen evaluation sources and comparability reports.
- Literature-Grounded Ideation: A PhD student wants to brainstorm novel research directions grounded in existing literature. The skill maps candidate ideas to arXiv and DOI sources for context.

## Prompt Templates

### Basic Candidate Exploration

```
I have a current_research branch exp/my-baseline with a trained checkpoint. Help me explore 3 candidate improvements using the ai-research-explore skill. Focus on architecture variations that preserve the current evaluation protocol.
```

### Campaign with SOTA Reference

```
Start a research_campaign with current_research at commit abc1234, benchmark GLUE, evaluation_source from my eval pipeline, and sota_reference paper 2301.00000. Generate candidate ideas for attention mechanism improvements with budget 4 GPU-hours.
```

### Source-Mapped Idea Generation

```
Map my current idea 'sparse attention with learned routing' to relevant sources from arXiv and GitHub. Then generate 2 derivative candidates that preserve the routing mechanism but vary the sparsity pattern.
```

### Feasibility and Fidelity Check

```
Run the execution_feasibility and implementation_fidelity passes on my top-3 ranked candidates. Report which ones are decomposable into auditable units and which need checkpoint discussion.
```

## Limitations

- Does not provide autonomous discovery or guarantee global benchmark completeness
- Novelty claims are hypotheses, not proofs - requires human verification against literature
- Requires an existing current\_research anchor \(branch, commit, checkpoint\) before use
- Network-dependent lookups require optional API tokens for authenticated access

## Best Practices

- Always establish a durable current\_research anchor \(branch, commit, or checkpoint\) before starting exploration
- Freeze the task family, dataset, benchmark, and evaluation source before generating candidate ideas
- Use cache-first source lookup and prefer local curated literature over open-ended web searches
- Write SCIENTIFIC\_CHANGELOG and COMPARABILITY\_REPORT artifacts to maintain auditable evidence

## Anti Patterns

- Do not use this skill for trusted reproduction - use ai-research-reproduction instead
- Do not skip the explicit exploration authorization step or proceed without a current\_research anchor
- Do not present exploratory gains as verified improvements or claim global benchmark completeness
- Do not silently choose between top candidates if the implementation cannot be decomposed into auditable units

## Security Audit

- - Safe to publish: true
- - Audited at: 2026-06-10T09:25:49.261\+00:00
- - Summary: The skill is a legitimate research exploration tool that uses subprocess for running internal pass scripts, hashlib for cache identity digests, and network calls to public research APIs \(GitHub, arXiv, DOI\) for metadata resolution. The static analyzer flagged 601 potential issues, but most are false positives from regex-based pattern matching. The 'weak cryptographic algorithm' findings refer to hashlib usage for non-security purposes \(cache deduplication\), not authentication. Subprocess calls execute internal pass scripts with controlled arguments, not user-controlled shell injection. Network access is limited to well-known research metadata providers. No prompt injection attempts or malicious intent detected.

## Stats

- - Views: 0
- - Downloads: 3
- - Favorites: 0
- - Popularity score: 0