# LOOP — {{PROJECT_NAME}} autonomous improvement loop

> The **dynamic** ATLAS surface: how an agent continuously improves this repo
> without a human in the loop. Pairs with the static quartet — [`ATLAS.md`](ATLAS.md)
> (*where*), [`SCARS.md`](SCARS.md) (*what breaks*), the SKILL playbook (*how*),
> and [`CLAUDE.md`](CLAUDE.md) (*how to act*). The task queue is [`ROADMAP.md`](ROADMAP.md).
>
> OPTIONAL: the static quartet is the core; this loop is an opt-in layer. Remove
> `LOOP.md` + `ROADMAP.md` if you don't run an autonomous loop.

## One iteration

```
# 0. orient — rely on the SessionStart hook, or run: atlas orient
# 1. pick the top ROADMAP.md item by expected value (see rule 2), not queue order
# 2. implement the smallest change that satisfies its test
# 3. atlas check --strict      # conformance gate — must pass before commit
# 4. commit (cite SCARS §ANCHORS); tick the item + move it to the Done log
```

Drive it with your runtime's recurring loop (e.g. `/loop LOOP.md`). Watch
iteration history with `atlas loop log`, which reads a local `LOOP_HISTORY.md`
append ledger plus tracked `LOOP_REPORT.md` history.

## Rules (battle-tested — keep them)

1. **Anti-churn pre-flight.** Grep before building; verify + tick what already
   ships. Never rebuild working code.
2. **EV-ranked selection.** Pick by `edge × P(real) × leverage ÷ cost`, not
   position in the queue. When the repo has an `AKIGI.md`, purpose-fit gates
   the EV: work (and incoming `FRQ.md` requests) that doesn't serve the
   AKIGI's Purpose is declined, however attractive it looks.
3. **Novelty mandate.** Each iteration adds ≥1 new falsifiable hypothesis to the
   idea ledger and clears the novelty bar (no near-duplicates).
4. **Self red-team before commit.** Check for overfit, data leakage, and honesty.
5. **Measure-then-gate.** Ship descriptive-only first; wire to behavior only
   after out-of-sample validation.
6. **Grow SCARS.** Every new failure mode becomes a `§ANCHOR` in `SCARS.md`.
7. **Difficulty routing.** Escalate `difficulty: hard` items to a stronger
   model / sub-agent.
8. **Conformance gate.** `atlas check --strict` must pass on every commit.

## Sub-agent brief

Every sub-agent you dispatch must be told: read `ATLAS.md`, `SCARS.md`, and the
SKILL playbook first; follow `CLAUDE.md`; run `atlas check --strict` before
returning. (Sub-agents don't inherit the SessionStart hook — say it in the prompt.)