Feynman Loop

The /feynman-loop skill is the core of the Learn domain. It implements a structured Explain → Gap-Find → Re-Study cycle for any concept.

The 3-step loop

1. Explain

The learner explains the concept in their own words (to "a 12-year-old" or to a non-expert audience). No notes, no lookup.

2. Gap-Find

learn-grade examines the explanation and identifies:

• Concepts that were omitted
• Misconceptions that were present
• Transfer problems not correctly solved

The grade routes through sycophancy-correction to prevent false "you got it!" feedback.

3. Re-Study

For each gap identified, learn-plan creates a targeted micro-study session. The cycle then repeats from step 1.

Mastery closure

A Feynman loop is closed when all three conditions hold:

1. learn-grade score ≥ 0.7 AND misconceptions_absent == 1.0
2. Two novel transfer problems solved at ≥ 0.7
3. Retention check via learn-retain at ≥ 24h interval after initial mastery

Self-reported fluency ("I feel like I understand this") never satisfies any condition.

The artifact a closed loop writes

Closing a loop writes one artifact, and every learn skill reads it from the same place:

<learn-home>/goals/<goal-id>/artifacts/<concept-id>/<artifact-id>.json

<learn-home> is ${PROMETHEUS_LEARN_HOME:-~/.prometheus/learn}. learn-retain globs artifacts/<concept-id>/*.json for the most recent artifact (latest closed_at); learn-certify reads the same concept directory for its evidence entries.

write-artifact.sh refuses an artifact that cannot say how it was checked. Two blocks are required:

• verification — exactly one entry per transfer score, each with a label from the four-value vocabulary shared with deep-research (verified | unverified | blocked | inferred) and non-empty evidence.
• provenance — the grade file and corpus path the scores came from.

A self-reported score is unverified and never satisfies mastery criterion 2.

What the learner model records

The learner-model binary carries the write paths the loop needs, so nothing a later skill reads is a value no earlier skill wrote:

Method	Written by	Read by
add_observation	learn-grade, learn-practice	mastery estimate
add_gap / resolve_gap	learn-grade	the next iteration's targets
add_session	learn-practice (open and close)	learn-certify practice-breadth gate
set_certified	learn-certify	learn-certify concept gate

Mastery holds at the seeded prior until the fifth observation, then follows mastery_new = mastery_old + 0.3 × (score − mastery_old). Spaced-repetition scheduling runs through FSRS, so difficulty is read and updated on every review.

Trigger phrases

• "teach me X"
• "explain this concept to me"
• "I want to really learn X"
• "feynman loop on X"