System architecture alignment

Purpose

Keep coding agents, skills, and PRs consistent with:

1. docs/FINAL_SYSTEM_VISION.md — canonical product and architecture vision (10 layers).
2. AGENTS.md — enforceable module boundaries, safety defaults, and current code mapping.
3. docs/IMPLEMENTATION_GAP.md — short vision vs implementation honesty (optional read when scoping work).

This skill does not replace reading those files.

Layer anchors (must not swap)

• Layer 5 = AI support — LLM confirmation / recommendation; never the only authority for execution.
• Layer 6 = Learning and optimization — outcome-driven improvement, shadow/diagnostic first in this repo unless explicitly extended.

Cross-cutting principles (binding with AGENTS.md)

• Strong operator observability — L9–L10 surfaces and journals should reflect what the system did and why.
• Continuous research / testnet data — prefer runs that feed journals and metrics for L6; avoid untracked “fire and forget” experiments.
• Bounded autonomy — autonomous loops stay within config, env, and operator controls; no silent background traders.
• No black-box runtime — no untraced changes to effective signals, weights, or risk; decision paths remain structured and logged where the repo already does so.

Use when

• Designing or reviewing features that span multiple packages (market, strategies, ai, learning, risk, execution, operator_panel).
• Documenting or refactoring “layers” in skills or README.
• Clarifying whether a change belongs to advisory AI (L5) vs learning (L6) vs risk (L7).

Do not use when

• Single-file trivial edits with no architectural implication.
• Replacing a full read of FINAL_SYSTEM_VISION.md for detailed formulas or future data sources.

Workflow

1. Read AGENTS.md (boundaries + safety).
2. Skim relevant sections of docs/FINAL_SYSTEM_VISION.md for the layers you touch.
3. Ensure the change advances the vision without violating AGENTS (paper default, risk not bypassed, AI advisory, learning not secretly trading).
4. Prefer minimal diff; update skills/docs only when guidance would otherwise stay wrong.

Output

• State which layers are affected (numbers + names).
• Note vision vs current repo gap if the implementation is partial.
• Explicit safety confirmation: risk ordering, paper defaults, no live enablement.