ails check reports a Level: L# <Name> line in the scorecard between Agent: and Scope:. The level is a read-out, not a gate — every rule fires when its match conditions apply regardless of which level your project is at. Use the level to self-locate; use the symptom table at the bottom to decide when to climb.

The ladder sorts by the channel each rung runs on: soft attention (L0–L5), hard enforcement (L6), self-writing memory (L7). Each rung adds a new diagnostic concern — scope leakage at L3, skill-instruction coherence at L4, governance-instruction alignment at L6, drift detection at L7.

• Attention — text the model reads and weights against everything else loaded. Fails probabilistically; competes for budget; decays with load. Fixes are content and ordering.
• Enforcement — hooks, deny-permissions. Acts outside the model's context. Fails deterministically when configured wrong, never silently. Fixes are scripts, schemas, permission rules. An MCP config is not enforcement: its tool definitions load into the model's context.
• Self-writing — agent-authored instructions written between sessions. At read time these land in attention like anything else; at write time the user never saw the prompt that produced them. Fixes are review cadence and explicit auto-memory boundaries.

The displayed level is the highest level whose own capability is present. A lower level does not need to be present: skills raise the level to L4 on their own, without rule files.

Each level is read on its own, so a project can skip a rung. Observed with ails check:

• CLAUDE.md plus one skill reads L4 (Delegated), with no rule files.
• CLAUDE.md plus hooks in .claude/settings.json reads L6 (Governed), with no skills.
• CLAUDE.md plus a .mcp.json reads L1 (Primer): an MCP config on its own adds no level.
• CLAUDE.md plus a .githooks folder of plain git hooks reads L1 (Primer): git hooks add no level.
• Claude's auto-memory for the project (kept in your home folder, not in the repository) counts for L7 (Adaptive) when Claude is the agent being checked, so the same project can read L7 on a machine that has it and lower in CI.
• A lone skill with no CLAUDE.md reads L4, both with a plain ails check . and with --agent claude.
• Hooks with no instruction file read L0: ails check . prints No instruction files found..

Each rung exists because the rung below it fails in a specific way. The trigger is the failure, not a feature wishlist.

Climbing without a symptom adds structure the model has to navigate without solving a problem you had. Under-climbing is more common: "agent didn't run tests before pushing" reads like a prompt-engineering problem but is usually a missing L6 hook; "agent forgot we use pnpm, not npm" reads like context drift but is usually a missing L7 memory entry.

← Rules CLI · Capability Levels · FAQ →