CARD: scaffold-e-fast-loop — Per-Cell Fast Verification Loop

**Discipline v0.2 · BETA**

Classification: layer=E (verification) + H (weak-reader); mechanism=scaffold E.

Intent: Guarantee every cell is independently compilable and testable in seconds, so an agent's edit→check→read-error→edit loop gets a first signal fast enough to steer — the substrate that makes every other scaffold's pass/fail usable.

Also Known As: tight feedback loop; incremental check; per-package test; smoke loop; agent-computer interface.

Applicability / Recognition: Apply when — a cell cannot be checked without building the whole repo; the only verification is a multi-minute CI run; an agent must wait minutes for a signal. *Detector seed:* `cargo test -p <cell>` fails to run in isolation, OR cell has no fast local check → recognition fires (verification locality beats capability, R3-007).

Motivation: A weak agent in a bounded loop gets one signal per 30-min CI run — useless. With `rust-ai-native fast-loop --cell <crate>` + `cargo test -p <cell>` returning in seconds, the same agent iterates ten times in the budget that previously bought zero feedback. The interpreter-budget result confirms: more local runs help agents that can use feedback (Opus, Sonnet); the loop is the amplifier.

Structure & Participants: *Cell isolation* (independently buildable package) · *Fast checker* (conform tiers + per-cell tests) · *Structured error* (Class F) · *Budget guard* (first signal < ~60s).

Collaborations: Runs Classes C/D/G checks; consumes Class B's compiler checks; emits Class F diagnostics. The raid executor runs this per batch.

Goals / Non-Goals: *Goals:* sub-minute first signal per cell; make the loop the standard agent workflow. *Non-Goals:* NOT replacing full CI (still runs at merge); does NOT create capability — Haiku didn't improve with more runs (it amplifies, not creates).

Consequences: (+) every other scaffold becomes loop-usable; (+) iteration count rises within budget. (−) requires cells to be genuinely isolable (drives §1 cell design); (−) fast checks may miss what full CI catches — gate-mode checks backstop.

Alternatives: whole-repo CI only (too slow for an agent loop); manual testing (not mechanical). Neither is a substitute.

Risks & Assumptions: assumes cells are isolable; god-files break this (anti-pattern). *Sunset:* none — the loop is foundational.

Evidence & Transfer-strength: R3-007 (verification locality, theory), R2C interpreter-budget result (more runs amplify capable agents, benchmark), BLD-010 (30–60s practical threshold). Class: benchmark + theory. Tag: **[E-strong]**.

trigger: WHEN a cell lacks a sub-minute isolated check, OR verification requires a full-repo build THEN apply mode: gate # a structural precondition, verified at merge routine: 1. Make the cell an independently buildable unit (own package or test target). 2. Provide `rust-ai-native conform check --scope <crate>` + `rust-ai-native fast-loop --cell <crate>` for it. 3. Ensure `cargo test -p <cell>` runs in isolation, < ~60s. 4. Wire the cell's Class C/D/G checks into that command. 5. Confirm first-signal latency is within budget; if not, reduce test case counts. checker: harness assertion `cell-fast-loop-present` (cell builds+tests in isolation under budget) raid_role: layer=infrastructure; order=after:cell-boundaries; batch=cell budget: active_rules=1; first_signal=<60s by definition