# Plan Bounded Deep Learning Experiments

Exploratory training can consume budgets without producing comparable evidence. This skill ranks limited variants and records results for careful human review.

## Install

```bash
npx skillstore add lllllllama/explore-run
```

## Metadata

- Status: approved
- Slug: lllllllama-explore-run
- Skillstore revision: r2
- Version status: missing
- Tree hash: 12bd36b0c7b19461e9e706116f97f3b0a3e765c81e8cd6d8311fc440207c3807
- Author: lllllllama
- GitHub username: lllllllama
- License: MIT
- Repository: https://github.com/lllllllama/rigorpilot-skills/tree/main/skills/explore-run/
- Ref: 656be3040aef5c047555a908cd5c695d22a4a548
- Supported tools: Claude, Codex, Claude Code
- Audit status: complete
- Agent install advisory: allowed
- Manual install advisory: allowed
- Artifact signature: available
- Audit attestation: unavailable
- Human verification: not\_verified
- Risk factors: external\_commands, filesystem
- Quality score: 80
- Quality tier: silver
- Public page: https://skillstore.pages.dev/skills/lllllllama-explore-run
- Manifest: https://skillstore.pages.dev/api/skills/lllllllama-explore-run/manifest

## Capabilities

- Builds candidate variants from parameter axes, subset sizes, and short-run step limits.
- Scores variants using cost, predicted success, and expected gain.
- Prunes candidates using maximum variant and short-cycle run budgets.
- Supports custom selection weights, primary metrics, and optimization direction.
- Defines output reports for changes, comparability, top runs, and status.

## Use Cases

- Screen Training Variants: Create a small candidate matrix before committing resources to full training.
- Prioritize Idle GPU Work: Rank bounded experiments by estimated cost, success probability, and expected gain.
- Review Exploratory Evidence: Document comparability limits and top candidates before approving further validation.

## Prompt Templates

### Plan a Small Probe

```
Plan a small-subset probe for [research goal]. Use [subset size] and [step limit]. Keep the trusted baseline unchanged.
```

### Create a Variant Matrix

```
Build variants for [parameter axes]. Limit the plan to [maximum variants] and [maximum short-cycle runs]. Rank candidates conservatively.
```

### Tune Selection Priorities

```
Rank candidates using these weights: cost [value], success rate [value], and expected gain [value]. Explain the highest-ranked choices.
```

### Assess Executed Runs

```
Review results for [primary metric] with a [maximize or minimize] goal. Separate measured evidence from heuristic predictions and note unfair comparisons.
```

## Limitations

- Requires explicit researcher authorization before exploratory execution.
- Plans and ranks experiments but delegates actual training execution to other skills.
- Heuristic scores are not evidence of model quality or successful execution.
- Shared reference and writer files may require the original repository structure.

## Best Practices

- Get explicit approval before planning or launching exploratory runs.
- Start with small subsets and short cycles before using larger budgets.
- Keep exploratory state separate and report every comparability limitation.

## Anti Patterns

- Do not treat heuristic rankings as measured research results.
- Do not modify or promote the trusted baseline automatically.
- Do not claim state-of-the-art performance from bounded exploratory evidence.

## Security Audit

- Audited at: 2026-08-08T08:51:32.385\+00:00
- Summary: All 21 static findings are false positives caused by Markdown code spans, documentation links, or restrictive policy text. No malicious intent, prompt injection, command execution, or unsafe filesystem operation was found.

## Stats

- Views: 1
- Downloads: 9
- Favorites: 0
- Popularity score: 0