Regulatory Templates - Gate 2: Technical Validation

Overview

This sub-skill executes Gate 2 of the regulatory template workflow: validating uncertain mappings from Gate 1 and confirming all field specifications through testing.

Parent skill: regulatory-templates

Prerequisites:

• Gate 1 PASSED
• Context object with Gate 1 results

Output: Validated mappings with test results and validation rules

---

Foundational Principle

Validation is the checkpoint that prevents incorrect mappings from reaching production.

Gate 2 is the quality gate between analysis (Gate 1) and implementation (Gate 3):

• All uncertainties resolved: Gate 1 analysis ≠ Gate 2 validation. MEDIUM/LOW uncertainties often hide critical issues
• 100% mandatory validation: 95% = 5% of mandatory data could be wrong in BACEN submission
• >90% test pass rate: Test data reveals transformation bugs, data type mismatches, edge cases
• Confirmed mappings: Prevents Gate 3 from generating templates based on assumptions
• Validation rules defined: Gate 3 needs explicit validation logic for template generation

Skipping validation in Gate 2 means:

• Gate 1 assumptions become Gate 3 implementation (no verification layer)
• Uncertainties propagate to production (BACEN submission failures)
• Low-confidence mappings generate incorrect templates (compliance violations)
• No test data validation = edge cases break in production

Gate 2 is not redundant - it's the firewall between analysis and implementation.

---

When to Use

Called by: regulatory-templates skill after Gate 1 passes

Purpose: Resolve uncertainties, validate field mappings, test transformations, define validation rules

---

NO EXCEPTIONS - Validation Requirements Are Mandatory

Gate 2 validation requirements have ZERO exceptions. This is the quality firewall before template generation.

Common Pressures You Must Resist

Pressure	Your Thought	Reality
Pragmatism	"Critical uncertainties only, skip MEDIUM/LOW"	PASS criteria: ALL uncertainties resolved. MEDIUM/LOW cascade to mandatory failures
Efficiency	"88% test pass rate is excellent"	Threshold is >90%. 12% failure = edge cases that break in production
Complexity	"Validation dashboard is redundant"	Mandatory validation = 100% required. Dashboard catches missing validations
Confidence	"Mappings look correct, skip testing"	Visual inspection ≠ test validation. Test data reveals hidden bugs
Authority	"95% mandatory validation is outstanding"	100% is non-negotiable. 5% gap = 5% of mandatory data wrong in BACEN
Frustration	"Use workarounds for rejected fields"	FAIL criteria: Cannot find alternatives. Workarounds bypass validation

Validation Requirements (Non-Negotiable)

All Uncertainties Resolved:

• ✅ REQUIRED: Resolve ALL Gate 1 uncertainties (CRITICAL + MEDIUM + LOW)
• ❌ FORBIDDEN: "Fix critical only", "Skip low-priority items"
• Why: MEDIUM/LOW uncertainties often reveal systemic issues, cascade to mandatory failures

Test Data Validation:

• ✅ REQUIRED: Test pass rate >90%
• ❌ FORBIDDEN: "88% is close enough", "Skip testing, looks correct"
• Why: Test data reveals transformation bugs, data type mismatches, edge cases

Mandatory Field Validation:

• ✅ REQUIRED: 100% mandatory fields validated
• ❌ FORBIDDEN: "95% is outstanding", "Edge cases don't matter"
• Why: Each 1% gap = potential BACEN submission failure on mandatory data

Alternative Mappings:

• ✅ REQUIRED: Find alternatives for ALL rejected fields
• ❌ FORBIDDEN: "Use workarounds", "Keep original with patches"
• Why: Rejected mappings fail validation for a reason - workarounds bypass the firewall

The Bottom Line

Partial validation = no validation.

Gate 2 exists to catch what Gate 1 missed. Lowering thresholds or skipping validation defeats the purpose. Every PASS criterion exists because production incidents occurred without it.

If you're tempted to skip ANY validation, ask yourself: Am I willing to defend this shortcut during a BACEN audit?

---

Rationalization Table - Know the Excuses

Every rationalization below has been used to justify skipping validation. ALL are invalid.

Excuse	Why It's Wrong	Correct Response
"Critical uncertainties only, MEDIUM/LOW can wait"	ALL uncertainties = all 8. MEDIUM cascade to mandatory failures	Resolve ALL uncertainties
"88% is excellent, 2% gap is edge cases"	>90% threshold exists for production edge cases	Fix to reach >90%
"Validation dashboard is redundant with Gate 1"	Gate 1 = mapping, Gate 2 = validation. Different purposes	Run dashboard, ensure 100%
"Mappings look correct, testing is busywork"	Visual inspection missed bugs testing would catch	Run test data validation
"95% is outstanding, 5% isn't worth 2 hours"	100% is binary requirement. 95% ≠ 100%	Fix to reach 100%
"Rejected fields can use workarounds"	Workarounds bypass validation layer	Find valid alternatives
"Gate 2 rarely finds issues after 50 templates"	Experience doesn't exempt from validation	Run full validation
"Following spirit not letter"	Validation thresholds ARE the spirit	Meet all thresholds
"Being pragmatic vs dogmatic"	Thresholds prevent regulatory incidents	Rigor is pragmatism
"Fix in next sprint if issues arise"	Regulatory submissions are final, can't patch	Fix now before Gate 3

If You Find Yourself Making These Excuses

STOP. You are rationalizing.

The validation exists to prevent these exact thoughts from allowing errors into production. If validation seems "redundant," that's evidence it's working - catching what analysis missed.

---

Gate 2 Process

Check for Template-Specific Validation Rules

Check for template-specific sub-skill at skills/regulatory-{template}/SKILL.md containing:

• Validation rules (VR001, VR002...), business rules (BR001, BR002...)
• Format rules, test data with expected outputs

Agent Dispatch with Gate 1 Context

Dispatch: Task(subagent_type: "finops-analyzer", model: "opus")

CRITICAL: ⚠️ DO NOT MAKE MCP API CALLS - use Gate 1 context ONLY

Prompt structure:

Section	Content
Context	Full Gate 1 context (field mappings, uncertainties)
Uncertain Mappings	For each: field_code, current_mapping, doubt, confidence, action_needed
Validation Tasks	1. Use Gate 1 mapping 2. Validate transformations 3. Check business logic 4. Confirm data types 5. Mark CONFIRMED/REJECTED
Output	Per field: code, resolution, alternative (if rejected), test_result

Output: Field resolutions + validation rules + cross-field logic + test data

---

Validation Process

⚠️ All validation uses Gate 1 context ONLY - no MCP API calls.

1. Field Validation

Per uncertain field: field_code, original_doubt, validation_steps (5), resolution (confirmed/rejected), transformation, test_data (input/expected/actual/status)

2. Validation Rules Definition

Rule Type	Example	Formula
field_format	CNPJ 8 digits	length(field_001) == 8
cross_field	CPF/CNPJ check	length(field_001) IN (11, 14)
date_range	Within period	field_020 >= period_start AND field_020 <= period_end

3. Test Results Documentation

Per test: field, test_name, input, transformation, output, expected, passed (true/false)

Example: Field 001 CNPJ extraction: "12345678000190" → slice:':8' → "12345678" ✓

---

Capture Gate 2 Response

Merge with Gate 1: validated_mappings[], validation_rules[], all_uncertainties_resolved, test_summary (total/passed/failed/success_rate)

---

Red Flags - STOP Immediately

If you catch yourself thinking ANY of these, STOP and re-read the NO EXCEPTIONS section:

Partial Resolution

• "Resolve critical only, skip MEDIUM/LOW"
• "Fix most uncertainties, good enough"
• "ALL is unrealistic, most is pragmatic"

Threshold Degradation

• "88% is close to 90%"
• "95% mandatory validation is outstanding"
• "Close enough to pass"
• "The gap isn't material"

Skip Validation Steps

• "Validation dashboard is redundant"
• "Mappings look correct visually"
• "Testing is busywork"
• "We'll catch issues in Gate 3"

Workaround Thinking

• "Use workarounds for rejected fields"
• "Patch it in Gate 3"
• "Fix in next sprint"
• "This is an edge case"

Justification Language

• "Being pragmatic"
• "Following spirit not letter"
• "Outstanding is good enough"
• "Rarely finds issues anyway"
• "Experience says this is fine"

If You See These Red Flags

1. Acknowledge the rationalization ("I'm trying to skip LOW uncertainties")
2. Read the NO EXCEPTIONS section (understand why ALL means ALL)
3. Read the Rationalization Table (see your exact excuse refuted)
4. Meet the threshold completely (100%, >90%, ALL)

Validation thresholds are binary gates, not aspirational goals.

---

Pass/Fail Criteria

PASS Criteria

• ✅ All Gate 1 uncertainties resolved (confirmed or alternatives found)
• ✅ Test data validates successfully (>90% pass rate)
• ✅ No new Critical/High issues
• ✅ All mandatory fields have confirmed mappings
• ✅ Validation rules defined for all critical fields

FAIL Criteria

• ❌ Uncertainties remain unresolved
• ❌ Test failures on mandatory fields
• ❌ Cannot find alternative mappings for rejected fields
• ❌ Data type mismatches that can't be transformed
• ❌ Mandatory fields validation < 100%

---

Mandatory Fields Final Validation

CRITICAL: Execute before Gate 2 completion

Per mandatory field, check:

• mapped (in gate1.field_mappings)
• confidence_ok (≥80%)
• validated (in gate2.validated_mappings)
• tested (in gate2.test_results)
• transformation_ok (works correctly)

Status: All checks PASS → field PASS; any FAIL → field FAIL

Gate 2 Pass Condition: all_mandatory_fields_valid == true required. Coverage must be 100%.

---

State Tracking

PASS: SKILL: regulatory-templates-gate2 | GATE: 2 | STATUS: PASSED | RESOLVED: {n} uncertainties | RULES: {n} defined | TESTS: {passed}/{total} | NEXT: → Gate 3

FAIL: SKILL: regulatory-templates-gate2 | GATE: 2 | STATUS: FAILED | UNRESOLVED: {n} | TEST_FAILURES: {n} | BLOCKERS: {description}

---

Technical Validation Checklist

Category	Validations
Field Naming	snake_case (not camelCase), check MCP API Dog naming
Data Types	String (length, UTF-8), Number (precision), Date (format), Boolean, Enum
Transformations	CNPJ/CPF slice, date timezone, decimal format, string trim/uppercase/padding, null defaults
Cross-Field	Dependent consistency, date ranges, calculated fields, conditional logic

Common Validation Patterns

Pattern	Input → Transformation → Output
CNPJ extraction	"12345678000190" → slice:':8' → "12345678"
Date format	"2025-01-15T10:30:00Z" → date_format:'%Y/%m' → "2025/01"
Decimal precision	1234.5678 → floatformat:2 → "1234.57"
Conditional	tipoRemessa == "I" → include all; "S" → approved only

---

Output to Parent Skill

Return: gate2_passed, gate2_context (merged), all_uncertainties_resolved, validation_rules_count, test_success_rate, next_action (proceed_to_gate3 | fix_validations_and_retry)