meta-learninglisted

# Meta-Learning > The learning layer is write-only by default. This skill is the > read-and-grade pass that keeps it honest. The wiki accumulates two kinds of learning: - **`wiki/lessons-learned.md`** — calibration events: "I claimed X, evidence overturned it." One entry per overturned claim. - **`wiki/pattern-index.md`** — recurring issues promoted to **named durable defaults** ("New default: ..."). One entry per pattern that hit ≥ 2 times. Without a check on whether those entries actually fire when they should, they become decorative. This skill is that check. --- ## When to invoke - **Mandatory: ON SESSION END** per `CLAUDE.md` lifecycle hooks. - When the user asks to audit recent learning ("did we apply X?", "was the new default load-bearing this week?"). - When `wiki/log.md` shows ≥ 3 sessions since the last L2 entry in `wiki/meta-learning.md` (catch-up audit). - **Do NOT** invoke during the same session that wrote a fresh lesson unless explicitly asked — see Same-session exclusion below. --- ## Levels - **L1** — the raw learning layer: `lessons-learned.md` + `pattern-index.md`. Authored at session end by the operator. - **L2** — this skill, auditing L1. - **L3** — this skill recursing on its own past L2 entries. - **L4+** — same procedure, one level higher. Permitted only while the termination rule keeps allowing it. --- ## Inputs 1. `wiki/lessons-learned.md` — current state. **HEAD+archive since 2026-06-10:** the head keeps the newest ~5