skills.rest
ooooooooooooooooooop
ooooooooooooooooooopCOMMUNITY
Explore all skills by this author →
GitHub

research-lab

Designs and executes reproducible controlled experiments via a Research Core runtime.

3|Updated Feb 13, 2026
Search Tags:#model comparison#evidence#benchmark#baseline#reproducibility#experiment#ablation
One-click install
npx skills add https://github.com/ooooooooooooooooooop/personal-ai --skill research-lab-ooooooooooooooooooop
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: research-lab
Source: https://github.com/ooooooooooooooooooop/personal-ai/tree/main/skills/research-lab
Command: npx skills add https://github.com/ooooooooooooooooooop/personal-ai --skill research-lab-ooooooooooooooooooop

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve? It turns falsifiable questions—such as whether one model can replace another—into structured, resumable experiments with traceable evidence instead of ad-hoc, unverifiable model comparisons. ## Core Features & Use Cases - ResearchSpec-driven experiments: Define question, hypotheses, baseline, candidates, dataset, metrics, and decision rules before any execution. - Recoverable runs: Create, execute, inspect, continue, and compare runs through research_ prefixed tools, with revision conflict handling across devices. - Evidence-bound conclusions: Every decision (SUPPORTED, REJECTED, INCONCLUSIVE, UNSUPPORTED, INVALID) must cite Evidence, Run, and Artifact digests. - Use Case: Evaluate whether Qwen 27B can replace Gemini Flash for novel state extraction by running both against a fixed test set with predefined quality, latency, and cost thresholds. ## Quick Start Ask the agent to use research-lab to turn your model comparison question into a ResearchSpec, execute the experiment with the research tools, and report a decision backed by evidence digests.

Frequently Asked Questions about research-lab

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I compare two LLMs empirically with reproducible results?▼

Define a ResearchSpec naming one model as baseline and the other as candidate, with a fixed dataset, metrics, and decision rules. Then run research_create, research_execute, research_compare, and research_evidence to get a decision backed by artifact digests.

How to evaluate whether a local model can replace a cloud model?▼

Set the cloud model as baseline and the local model as candidate, then measure quality, latency, and cost on identical inputs with thresholds fixed in decisionRules before execution. If the local model is unavailable on the device, the skill returns UNSUPPORTED instead of fabricating scores.

When should I not use a controlled experiment workflow?▼

Skip it for simple factual queries, read-only audits or inventories, low-risk reversible choices, and ordinary writing requests. The skill is only for falsifiable questions requiring real comparative execution with evidence.

Can I resume an interrupted experiment on another device?▼

Yes, call research_continue with the existing run id rather than creating a new research. Revision conflicts across devices must be surfaced explicitly; silent last-write-wins resolution is forbidden.

Why does my research return INCONCLUSIVE or INVALID?▼

INVALID means the Spec or execution was malformed; INCONCLUSIVE means the experiment ran validly but evidence was insufficient under the predefined decision rules. Add more cases or tighten thresholds and continue the same research rather than starting over.

Related Skills

Explore More Education & Research Skills→
Arize-aiArize-ai
11.3k

datasets

Reason about Phoenix dataset examples, outputs, splits, and their role in evaluators and experiments.

Official
Intermediate
mzniumzniu
7

tech_evaluation

Assess technology readiness, scalability, security, and architectural risks in a structured report.

Community
Advanced
ACGSpgpACGSpgp

llm-evaluation

Evaluate LLM applications with automated metrics, human feedback, and benchmarking.

Official
Advanced
surus-latsurus-lat
8

push-to-latamboard

Merge local benchmark results into the HuggingFace LatamBoard leaderboard dataset.

Official
Advanced