skills.rest
Lingjie-wang
Lingjie-wangCOMMUNITY
Explore all skills by this author →
WebsiteGitHub

result-to-claim

Evaluates experiment results against intended research claims and routes to next actions.

Updated Jul 9, 2026
Search Tags:#ablation#wandb#evidence verification#research workflow#codex mcp#experiment evaluation#research claims
One-click install
npx skills add https://github.com/Lingjie-wang/autoRL --skill result-to-claim-lingjie-wang
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: result-to-claim
Source: https://github.com/Lingjie-wang/autoRL/tree/main/Auto-claude-code-research-in-sleep/skills/result-to-claim
Command: npx skills add https://github.com/Lingjie-wang/autoRL --skill result-to-claim-lingjie-wang

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve? After experiments finish, researchers must decide what the numbers actually support before writing a paper or running ablations. This Skill provides an objective verdict on whether results support, partially support, or fail the intended claim, preventing inflated claims and post-hoc rationalization. ## Core Features & Use Cases - Deterministic evidence pre-check: Verifies that every cited number actually exists in its result file before any model judgment, catching hallucinated evidence early. - Codex MCP jury evaluation: Sends collected results, baselines, and caveats to Codex for a structured verdict (claim_supported, missing evidence, suggested revisions, confidence). - Automatic routing: Based on the verdict, routes to postmortem recording, supplementary experiments, ablation planning, or paper writing, and updates the research wiki with experiment nodes and support edges. - Use Case: After a training run completes on W&B, invoke this Skill to collect metrics, verify cited numbers, get an objective judgment on your claim, and record the verdict in findings.md and the research wiki. ## Quick Start Evaluate whether my completed experiments support the intended claim by running the result-to-claim gate on the latest W&B run.

Frequently Asked Questions about result-to-claim

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

FAQPage Schema
How do I check if my experiment results support my research claim?▼

Collect results from W&B, experiment logs, or training logs, then submit them with the intended claim and baselines to a Codex MCP evaluation. The verdict returns claim_supported (yes, partial, or no), missing evidence, and suggested claim revisions.

How to detect hallucinated evidence in research claims?▼

Build a claims list mapping each cited number to its source file, then run the deterministic evidence_check.py script before any model judgment. Claims with value_not_found or path_missing status are flagged as hallucinated evidence and marked unsupported immediately.

What happens when experiment results only partially support a claim?▼

A partial verdict triggers updating the working claim to reflect what is supported, recording the gap in findings.md, and designing supplementary experiments. Multiple partial rounds on the same claim suggest narrowing the claim scope or switching ideas.

Can this evaluation run without Codex MCP available?▼

Yes, if the Codex MCP call fails, the fallback judgment is made directly and marked as pending Codex review so the pipeline is not blocked. The deterministic evidence pre-check also runs independently of any model call.

When should I run a result-to-claim evaluation in my research workflow?▼

Run it after main experiments complete, before committing to claims in a paper or review response, and before running ablations. It should run once per result set, not on a schedule, since the verdict only changes when results change.

Related Skills

Explore More Education & Research Skills→
OpenLAIROpenLAIR
1.1k

aris-result-to-claim

Evaluates experiment results against intended claims and routes to next research action.

Official
Advanced
LambenthanLambenthan
77

exp-eval

Judge completed experiments and update wiki claims with verdicts.

Community
Advanced
qcmuuqcmuu

weights-and-biases

Track experiment runs, log metrics, and manage hyperparameters with Weights & Biases.

Community
Advanced
vstorm-covstorm-co
1.0k

research-methodology

Structure literature reviews with search strategy, source evaluation, and note-taking templates.

Official
Advanced