Add anti-sycophancy audit dimension

[mickeylorenzini]

Problem

LLM outputs can exhibit sycophancy — agreeing with the user, giving overly positive assessments, and avoiding pushback, driven by training optimization for user satisfaction. This degrades deliverable quality by:

• Affirming incorrect assumptions instead of challenging them
• Providing false positive assessments ("great approach!") without critical analysis
• Avoiding necessary pushback on flawed designs or specs
• Expressing confidence without supporting evidence

This is distinct from anti-rationalization (agent deceiving itself about skipped work). Sycophancy is agent-to-user deception via excessive agreeableness.

Proposed Solution

Investigate and implement detection of sycophantic patterns in agent-generated deliverables:

1. Research existing AI alignment literature on sycophancy detection
2. Define detectable patterns: unqualified agreement, false positives, avoidance of tradeoff discussion, confidence without evidence
3. Design as harden audit pass or detection dimension
4. Consider: can harden detect sycophancy in its own output? (meta-concern)

Likely Affected Files/Modules

• SKILL.md — new detection dimension or dedicated pass

Acceptance Criteria

• Literature review summary on sycophancy detection approaches
• At least 5 documented sycophantic patterns with examples
• Detection heuristics for each pattern
• Recommendation on integration approach (new pass vs. existing pass dimension)
• Address meta-concern: how to prevent harden itself from being sycophantic during audit