learning-layerlisted

# Learning Layer The rest of this package closes **one** gap: the codebase's (`gap = diff(to_be, as_is)`). This skill closes a **second** gap the others ignore — the **operator's**: the distance between this user's current judgment and a senior's. An agent that only produces good output leaves the operator where they were (or deskills them); this layer raises the operator *while* producing the same output. It is a **layer**, not a workflow: it wraps the surfaces that already exist (the interview, the plan/roadmap, test-first, the review) and adds a coaching hook to each. It authors no new phase and blocks no delivery. ## The one design decision that makes it work: one mode, not two The obvious design — a "learn mode" you toggle on — fails, because a mode is a switch and switches get set to "fast" under deadline, permanently. So there is **one mode**: - **Delivery always ships senior-grade output** and **always emits a passive why-trace** (the reasoning, the assumptions surfaced, the class each review flag belongs to). Free to ignore, always there. Delivery is never held hostage to teaching. - The **only** opt-in is a per-artifact **micro-retrieval**: *commit a cheap prediction before the answer is revealed*. It is **default-on with a one-key skip**, never opt-in. Defaults beat willpower: under deadline you skip and still ship; on the median task you don't bother to skip, and you learn. Flipping this default from opt-in to opt-out is the difference between a too