confidence-check

# Confidence Check

## Overview

**Purpose**: Shannon's quantitative 5-check validation algorithm prevents wrong-direction work by ensuring ≥90% confidence before implementation. Each check contributes weighted points (total 100%) across duplicate verification, architecture compliance, official documentation, working OSS references, and root cause identification.

**Critical Role**: This skill prevents the most expensive failure mode in software development - building the right thing wrong, or building the wrong thing right. Proven 25-250x token ROI in SuperClaude production use.

---

## Inputs

**Required:**
- `specification` (string): Implementation request or feature description from user
- `context` (object): Optional context from spec-analysis skill (8D complexity scores, phase plan)

**Optional:**
- `skip_checks` (array): List of checks to skip (e.g., ["oss", "root_cause"] for simple tasks)
- `confidence_threshold` (float): Override default 0.90 threshold (e.g., 0.85 for fast iterations)

---

## Anti-Rationalization (Critical - Read First)

**WARNING**: Agents systematically rationalize skipping confidence checks. Below are the 6 most dangerous rationalizations detected in production, with mandatory counters.

### Rationalization 1: "User seems confident, proceed"
**Example**: User says "I'm 75% sure this is right" → Agent responds "Let's proceed..."

**COUNTER**:
- ❌ **NEVER** accept user's confidence level without running 5-check algorithm
- ✅ User intuition is ONE data point (cognitive), not the total score
- ✅ Run all 5 checks objectively: duplicate (25%), architecture (25%), docs (20%), OSS (15%), root cause (15%)
- ✅ If calculated score <90%, STOP regardless of user's stated confidence
- ✅ Example: User "75% confident" + passed all 5 checks = 100% confidence → PROCEED
- ✅ Example: User "95% confident" + failed duplicate check (0/25) = 75% → STOP

**Rule**: Algorithm score overrides stated confidence. Always calculate objectively.

### Rationalization 2