verification-quality

STANDARD OPERATING PROCEDURE

Purpose

Assess whether delivered outputs are correct, well-grounded, and policy-compliant by tracing evidence, checking constraints, and calibrating confidence.

Trigger Conditions

• Positive: validating generated content, checking analysis against sources, or confirming policy compliance before delivery.
• Negative: code execution debugging (use functionality-audit) or style-only work (use style-audit).

Guardrails

• Confidence ceiling: Always emit Confidence: X.XX (ceiling: TYPE Y.YY) with ceilings {inference/report 0.70, research 0.85, observation/definition 0.95}.
• Evidence mapping: Link each claim to source evidence with citations or file:line references; flag ungrounded assertions.
• Constraint coverage: Verify hard/soft/inferred constraints separately; note unresolved inferences.
• Structure-first: Maintain examples/tests demonstrating grounding checks and policy validation.

Execution Phases

1. Scope & Constraints
   • Extract requirements and constraints from the request; categorize as hard, soft, inferred.
   • Identify allowed sources and policies.
2. Evidence Collection
   • Gather direct evidence (files, logs, references) and map to claims.
   • Note gaps or conflicts in evidence.
3. Assessment
   • For each claim, rate correctness, grounding strength, and policy alignment.
   • Distinguish observation vs. inference; cap confidence accordingly.
4. Reporting & Remediation
   • Summarize compliant items, violations, and uncertainties with recommended actions.
   • Provide confidence with explicit ceiling and English-only output.

Output Format

• Constraint breakdown (hard/soft/inferred) and scope.
• Claim-by-claim verification table with evidence and confidence.
• Violations or gaps with remediation steps.
• Confidence statement using ceiling syntax.

Validation Checklist

• Constraints extracted and categorized; policies identified.
• Evidence linked to each claim; gaps flagged.
• Confidence capped by evidence type; ceilings stated.
• Remediation steps proposed for violations.
• Output in English with explicit confidence ceiling.

Confidence: 0.72 (ceiling: inference 0.70) - SOP rewritten leveraging Prompt Architect confidence and constraint extraction plus Skill Forge structure-first verification.