vcsdd-adversarial-refinement

VCSDD Adversarial Refinement

When to Activate

• Phase 3 (adversarial review)
• Interpreting adversary findings
• Calibrating review intensity

Adversary Calibration Principles

The Adversary MUST:

1. Never say "overall looks good" or equivalent
2. Always cite real file paths and line numbers
3. Always produce a binary PASS or FAIL per dimension
4. Never soften findings with qualifiers like "minor" or "nitpick"
5. Always specify routeToPhase for actionable routing

Severity Classification

Severity	Definition	Examples
critical	Pipeline cannot continue	Missing spec requirement entirely, security vulnerability
high	Significant defect requiring rework	Wrong return type, missing error handling, untested edge case
medium	Defect that reduces quality	Misleading name, unnecessary abstraction, dead code
low	Improvement opportunity	Style inconsistency, minor duplication

Common Leniency Traps

Watch for these patterns and reject them:

• "The code handles most cases" → HOW MANY? WHICH ONES MISSING?
• "Error handling is adequate" → ADEQUATE BY WHAT STANDARD?
• "Generally follows the spec" → WHICH REQUIREMENTS ARE MISSING?
• "Minor refactoring could help" → IS THIS STRUCTURAL_INTEGRITY FAIL OR NOT?

Finding Routing Guide

Category	Meaning	Route To
spec_ambiguity	Requirement or contract wording is ambiguous	1a
spec_gap	Behavior exists but spec/contract does not cover it	1a
requirement_mismatch	Spec is clear, but tests or implementation do not satisfy it	2b
missing_edge_case	Edge case absent from spec or tests	1a or 2a
test_coverage	Required behavior has no meaningful test	2a
test_quality	Tautological or over-mocked test, or implementation-detail assertion	2a
implementation_bug	Logic defect	2b
error_handling	Incorrect or missing error path	2b
security_surface	Validation/auth/injection issue	2b
code_structure	Cohesion/abstraction/organization issue	2c
naming	Misleading or imprecise names	2c
duplication	Repeated logic that should be unified	2c
proof_gap	Missing proof harness or unverifiable claim	5
invariant_violation	Proof/testing evidence contradicts required invariant	5
purity_boundary	Supposedly pure logic still depends on effects	2c or 5