Ground

Intent

Verify that claims are supported by evidence from reliable sources. This is the foundation of the Grounded Agency principle - no assertion should be made without evidence anchors.

Success criteria:

• Claim evaluated against available evidence
• Evidence sources identified and referenced
• Grounding strength assessed (strong, moderate, weak, ungrounded)
• Evidence gaps explicitly documented

Compatible schemas:

• schemas/output_schema.yaml

Inputs

Procedure

1. Parse the claim: Understand what is being asserted

   • Identify the core assertion
   • Note any implicit sub-claims
   • Determine what evidence would support it

2. Identify evidence sources: Determine where to look

   • Use provided sources if available
   • Identify additional relevant sources
   • Prioritize authoritative sources

3. Search for evidence: Look for supporting or contradicting data

   • Read relevant files and documentation
   • Execute commands to gather data
   • Note both supporting and contradicting evidence

4. Evaluate evidence quality: Assess reliability of found evidence

   • Consider source authority
   • Check for recency and relevance
   • Note any bias or limitations

5. Assess grounding strength: Determine how well claim is supported

   • Strong: multiple authoritative sources agree
   • Moderate: one authoritative source or multiple secondary
   • Weak: indirect or partial evidence
   • Ungrounded: no supporting evidence found

6. Document gaps: Note missing evidence

   • What evidence would strengthen grounding?
   • What sources could not be checked?
   • What assumptions remain unverified?

Output Contract

Return a structured object:

grounded: boolean  # Whether claim meets required_strength
claim: string  # The claim being evaluated
strength: string  # strong, moderate, weak, ungrounded
evidence:
  - source: string  # Evidence source
    type: string  # file, url, command, observation
    content: string  # Relevant excerpt
    supports: boolean  # Supports or contradicts claim
    authority: string  # high, medium, low
contradictions:
  - source: string
    content: string
    severity: string  # How seriously it undermines claim
gaps:
  - description: string  # What evidence is missing
    would_provide: string  # What it would establish
confidence: 0..1
evidence_anchors: ["file:line", "url", "command:output"]
assumptions: []

Field Definitions

Examples

Example 1: Ground a Technical Claim

Input:

claim: "The payment processor uses Stripe API v2023-10-16"
sources:
  - "Gemfile"
  - "config/initializers/stripe.rb"
required_strength: "strong"

Output:

grounded: true
claim: "The payment processor uses Stripe API v2023-10-16"
strength: "strong"
evidence:
  - source: "Gemfile"
    type: "file"
    content: "gem 'stripe', '~> 10.0'"
    supports: true
    authority: "high"
  - source: "config/initializers/stripe.rb"
    type: "file"
    content: "Stripe.api_version = '2023-10-16'"
    supports: true
    authority: "high"
  - source: "Gemfile.lock"
    type: "file"
    content: "stripe (10.1.0)"
    supports: true
    authority: "high"
contradictions: []
gaps:
  - description: "Runtime verification of API version"
    would_provide: "Confirmation that configured version is actually used"
confidence: 0.95
evidence_anchors:
  - "Gemfile:42"
  - "config/initializers/stripe.rb:3"
  - "Gemfile.lock:156"
assumptions:
  - "Configuration is loaded at runtime"
  - "No environment-specific overrides"

Example 2: Partial Grounding

Input:

claim: "User authentication requires 2FA for all accounts"
sources:
  - "app/models/user.rb"
  - "config/devise.yml"
required_strength: "strong"

Output:

grounded: false
claim: "User authentication requires 2FA for all accounts"
strength: "weak"
evidence:
  - source: "app/models/user.rb"
    type: "file"
    content: "has_one :two_factor_auth"
    supports: true
    authority: "medium"
  - source: "config/devise.yml"
    type: "file"
    content: "two_factor_authentication: enabled: true"
    supports: true
    authority: "high"
contradictions:
  - source: "app/controllers/sessions_controller.rb"
    content: "skip_2fa_for_api_tokens: true"
    severity: "high"
  - source: "app/models/user.rb"
    content: "scope :without_2fa, -> { where(two_factor_enabled: false) }"
    severity: "medium"
gaps:
  - description: "Policy enforcement mechanism"
    would_provide: "Evidence that 2FA is mandatory, not optional"
  - description: "Admin user configuration"
    would_provide: "Whether admin accounts have different requirements"
confidence: 0.40
evidence_anchors:
  - "app/models/user.rb:15"
  - "config/devise.yml:8"
  - "app/controllers/sessions_controller.rb:23"
assumptions:
  - "2FA support exists but enforcement unclear"

Example 3: Ungrounded Claim

Input:

claim: "The system processes 10,000 requests per second"
sources:
  - "README.md"
  - "docs/architecture.md"
required_strength: "moderate"

Output:

grounded: false
claim: "The system processes 10,000 requests per second"
strength: "ungrounded"
evidence:
  - source: "README.md"
    type: "file"
    content: "High-performance API server"
    supports: false
    authority: "low"
contradictions: []
gaps:
  - description: "Load test results"
    would_provide: "Measured throughput under load"
  - description: "Production metrics"
    would_provide: "Actual request rates and latencies"
  - description: "Benchmark documentation"
    would_provide: "Performance claims with methodology"
confidence: 0.1
evidence_anchors:
  - "README.md:5"
  - "docs/architecture.md (no relevant content)"
assumptions:
  - "No performance data available in provided sources"

Verification

• Claim is clearly stated
• At least one evidence source was checked
• Strength assessment is consistent with evidence
• Contradictions are noted if found
• Evidence anchors reference specific locations

Verification tools: Read (to verify evidence excerpts)

Safety Constraints

• mutation: false
• requires_checkpoint: false
• requires_approval: false
• risk: low

Capability-specific rules:

• Never claim something is grounded without evidence
• Note contradicting evidence even if claim seems grounded
• Flag when sources may be outdated
• Do not fabricate evidence

Composition Patterns

Commonly follows:

• generate - Ground generated content
• predict - Ground prediction assumptions
• explain - Ground explanation premises

Commonly precedes:

• verify - Grounding enables verification
• audit - Grounding supports audit trails
• plan - Only proceed with grounded information

Anti-patterns:

• Never skip grounding for important assertions
• Avoid accepting claims without evidence in planning

Workflow references:

• See reference/workflow_catalog.yaml#world_model_build for grounding in world models
• Core principle of Grounded Agency - all skills should ground claims