Evidence Gate is a lightweight validation layer that enforces structured evidence obligations before an AI agent presents a conclusion or executes a high-impact action.
Install & Download
1. ClawHub CLI
The fastest way to install a skill directly from the registry.
npx clawhub@latest install evidence-gate
2. Manual Installation
Copy the skill folder to one of these locations
~/.openclaw/skills/ <project>/skills/ Priority: Workspace > Local > Bundled
3. Prompt Installation
Copy this prompt to OpenClaw to install it automatically.
Help me install evidence-gate using Clawhub. If Clawhub is not installed, install it first (npm i -g clawhub).
Prefer to download?
Get the raw skill files in a ZIP archive.
What is Evidence Gate?
Evidence Gate is a specialized tool designed to bring accountability and structure to AI agent decision-making. Instead of relying on vague caution, this Openclaw Skills integration forces agents to define exactly what evidence must exist before a claim—such as a root-cause diagnosis or a production configuration change—is considered responsible. By treating tentative conclusions as unverified until they pass through the gate, it ensures that AI outputs are auditable and grounded in verifiable data.
The skill operates as a non-blocking, single-pass validator. It doesn't replace existing workflows but rather inserts a checkpoint to evaluate known evidence against specific requirements. It is particularly effective for preventing hallucination-driven actions in sensitive environments like SRE, security, and software development, where the cost of an incorrect definitive statement is high.
Evidence Gate Use Cases
- Validating root-cause diagnoses in SRE and system debugging to ensure they aren't based on a single log line.
- Gating high-impact infrastructure actions like database deletions, rollbacks, or production configuration changes.
- Verifying safety assertions before an agent claims a specific operation or code change is safe to execute.
- Strengthening research conclusions by requiring source support and checking for contradictory evidence.
- Transitioning from automatic execution to human-in-the-loop approval when evidence thresholds are not met.
How Evidence Gate Works
- The agent reaches a tentative conclusion or identifies a high-risk action and invokes the skill.
- The skill normalizes the candidate claim and defines 2-5 specific, externally checkable evidence obligations.
- It evaluates the evidence currently available within the context, marking requirements as satisfied, missing, or conflicting.
- A structured verdict (PASS, SOFT_PASS, BLOCK, or CONFLICT) is generated based on the sufficiency of the evidence.
- The skill provides a safe next-step policy, suggesting either proceeding, downgrading the claim's strength, or gathering specific missing data.
- The agent receives the JSON output and adjusts its behavior or communication style according to the gate's guidance.
Evidence Gate Setup
To integrate this component of Openclaw Skills into your agent environment, ensure your agent has access to the core protocol definitions. You can initialize the skill via the CLI or include it in your agent's toolset configuration:
# Example of invoking the gate manually for testing
/evidence-gate "The root cause is a nil dereference in request parsing"
# Or for a high-impact action
/evidence-gate "Safe to delete the staging database"
Ensure that references/verdict-schema.json and the associated input/output templates are accessible in your skill's directory for proper schema validation.
Evidence Gate Data Schema & Taxonomy
The skill utilizes a structured JSON schema to ensure machine-readability and consistency across different AI platforms.
| Field | Description |
|---|---|
gate_required |
Boolean indicating if the claim warrants formal evidence gating. |
evidence_requirements |
A list of concrete, checkable obligations generated for the claim. |
verdict |
One of four states: PASS, SOFT_PASS, BLOCK, or CONFLICT. |
missing_evidence |
Specific artifacts or data points that were not found in the current context. |
fallback_behavior |
Guidance on how to downgrade the claim (e.g., from 'settled' to 'provisional'). |
next_actions |
Recommended steps for the caller to gather missing evidence. |
Evidence Gate Advanced Features
- Intelligent inference of risk levels and execution modes based on the claim's natural language.
- Safe downgrade guidance that suggests specific wording to avoid overstating certainty when evidence is thin.
- Support for multiple verdict states allowing for nuanced 'SOFT_PASS' results where low-risk continuation is permitted.
- Stateless execution design, making it ideal for multi-agent systems and distributed Openclaw Skills workflows.
- Configurable integration policies to override default trigger heuristics for specific domain requirements.
SKILL.md
Loading
Related Openclaw Skills
AI Skill Creator Attach Info
A specialized utility for Openclaw Skills that automatically injects essential metadata like Author, Date, and Version into newly created skill files.
GitHub Repo Teardown
A professional-grade analysis tool that transforms complex GitHub repositories into comprehensive product and engineering teardown reports.
GitHub CLI Skill
A powerful integration for Openclaw Skills that leverages the GitHub CLI to manage repositories, pull requests, and CI/CD pipelines directly via terminal commands.
Podcast Summarize
A professional-grade AI skill designed to convert podcast episodes from various sources into structured, actionable text summaries.
content-researcher
A powerful batch research tool that automates searching, collecting, and summarizing articles or news into structured material libraries.
LaTeX Revision Tracker
A professional-grade workflow for tracking, managing, and revising large-scale LaTeX academic papers with precision and version control.
Featured*
