Add-on: Range Rules Validation (Policy Engine)

Use this skill to implement the validation stage:

• consumes extracted_clauses
• compares against known rules/policies (“Range”)
• emits structured validation results with explanations, severity, confidence, and review flags

This stage should not depend on top-k retrieval over the document as the main method.

Inputs

Collect:

• RANGE_RULE_SOURCE: db (recommended) or repo-file (YAML/JSON checked into repo).
• VALIDATION_RESULT_SET: compliant|non_compliant|uncertain.
• SEVERITY_LEVELS: default info|low|medium|high|critical.
• UNCERTAINTY_REVIEW_THRESHOLD: default 0.6.

Data Contracts

Rules

Model rules as versioned objects with:

• rule_id (stable)
• name
• description
• applies_to_clause_types
• logic (deterministic checks) and/or guidance (LLM-assisted rubric)
• version, effective_at

Validations (validations table)

Each validation record must include:

• clause_id
• rule_id / rule_name
• result (from VALIDATION_RESULT_SET)
• explanation (human-readable)
• severity
• confidence (0..1)
• review_needed (true when uncertain or confidence low)
• citations: reuse clause citation linkage; do not lose source anchoring

Validation Workflow

1. Load active ruleset (latest version or effective-at timestamp).
2. For each extracted_clause:
   • select applicable rules by clause_type
   • evaluate deterministically when possible
   • if LLM-assisted evaluation is used, enforce strict schema output and store model+prompt provenance
3. Persist one validation row per (clause, rule) evaluation.
4. Mark review_needed when:
   • result is uncertain, or
   • confidence < threshold, or
   • clause is missing a required field for the rule

Guardrails

• Do not fetch extra context via semantic retrieval unless explicitly requested; validate on clause + rule.
• If a rule cannot be applied due to missing evidence, output uncertain and set review_needed=true.
• Store rule versions used in every validation for reproducibility.

Decision Justification Rule

• Every non-trivial decision must include a concrete justification.
• Capture the alternatives considered and why they were rejected.
• State tradeoffs and residual risks for the chosen option.
• If justification is missing, treat the task as incomplete and surface it as a blocker.