Anti-Sycophancy in Learning

Sycophantic feedback — telling a learner they understood something when they didn't — produces measurably worse learning outcomes. Prometheus blocks this architecturally.

The problem

LLMs have a natural tendency to:

• Praise mediocre explanations ("Great job! You almost got it!")
• Understate gaps ("You just need to review the details")
• Affirm self-reported confidence ("If you feel confident, you probably are")
• Avoid saying "No" or "Wrong" directly

In a learning context, this sycophancy is actively harmful — it closes loops that should stay open, and creates false confidence in unmastered material.

The architectural solution

learn-grade is on the critical path of the sycophancy correction gate:

learner explanation
  → learn-grade (draft score)
    → sycophancy-correction S-02 check
      → rewrite if sycophantic
    → final score delivered

A grade that says "no gaps" when gaps are present is rewritten before delivery. This is not optional guidance — it is enforced in the learn-grade skill script.

Sycophancy patterns blocked

Pattern	Example
False positive	"Score: 0.9 — only minor gaps!" when multiple misconceptions present
Gap minimization	"The explanation was clear" when key concept was omitted
Self-report affirmation	"Since you feel confident, you've mastered this"
Encouragement inflation	"Excellent explanation! Just a few edge cases missing..."

Behavioral rules

• "Never tell the learner they did well when they did not" — this is a hard operator invariant
• Self-reported fluency NEVER closes a Feynman loop — the 3-condition gate is the only path
• learn-grade MUST include concrete gaps if any exist

For the Reflector gate

The same sycophancy-correction system that guards learn-grade also guards agent reflections via the reflector SubagentStop hook. See Sycophancy Correction for full details.