Leverage & Multipliers
Status: plan.ai-labs mandate extension, 2026-05-31, council-vetted (Claude Opus 4.8 · OpenAI codex · Cursor Gemini 3.1 Pro · Cursor GPT-5.5, unanimous). This supersedes playbook Teil E.1 (no leverage / no options / no leveraged ETFs) for the lab — the binary ban becomes a deterministic bound.
The principle (reinforced, not relaxed)
Section titled “The principle (reinforced, not relaxed)”Leverage is a risk input, not a strategy. “Max broker leverage” is a ruin setting, not a policy. Adding multipliers raises the stakes on the deterministic safety layer — it does not replace its thesis. Every multiplier enters through hard, testable caps enforced in code, right of the place_order boundary. The owner’s choice (margin + leveraged ETFs + long options, up to broker capability — bounded by deterministic caps below it — paper-first/live-ready) is therefore implemented as a Risk/Exposure Engine, not as a loosened gate.
Risk/Exposure Engine (extends Role 6)
Section titled “Risk/Exposure Engine (extends Role 6)”A single deterministic RiskState, computed in Python every cycle, is the only thing place_order consults for sizing/approval:
- Exposure: gross & net, per-instrument and portfolio leverage.
- Margin: buying power, margin used, distance to maintenance margin.
- Options: aggregate Greeks (delta/gamma/theta/vega), premium-at-risk (= max loss), expiry/assignment calendar.
- Caps (deterministic, portfolio-level — YOURS, below the broker max): max gross exposure, max margin %, max option premium-at-risk, max single-position, max correlated-cluster exposure, and a portfolio max-drawdown HALT.
Per-instrument handling
Section titled “Per-instrument handling”| Instrument | Deterministic treatment |
|---|---|
| Margin (equities) | Track buying power + maintenance margin; auto-deleverage before a maintenance breach; never rely on the broker margin call. |
| Leveraged ETFs (2x/3x) | Count 3x as 3× exposure in sizing; decay-aware MAX-HOLD (short holds only); flag high-decay; no overnight unless cap allows. |
| Long options (calls / debit spreads) | Defined-risk only. Premium = position size (max loss known). Track Greeks/expiry/assignment/early-exercise; no naked / undefined-risk; auto-handle expiry & assignment in the accounting layer. |
The critical amendment — risk-reduction is automatic (Gemini/Critic)
Section titled “The critical amendment — risk-reduction is automatic (Gemini/Critic)”At max leverage the HITL human is a single point of failure; approval latency is risk (a 15-min token is an eternity for a 3× position). Therefore:
- HITL gates ENTRIES only.
- Exits, stop-outs, de-leveraging, margin-call avoidance, and the portfolio-drawdown HALT fire deterministically and immediately — WITHOUT waiting for a human tap. The human/cockpit is notified, not awaited.
This resolves playbook open question F.2 #10 and is required once leverage is on.
Fail-safe, not just fail-closed — broker-side protection (Gemini/Critic, top gap)
Section titled “Fail-safe, not just fail-closed — broker-side protection (Gemini/Critic, top gap)”Under leverage, “fail-closed = HOLD” is no longer sufficient. Fail-closed only stops new orders. If the whole stack (core / Hermes / cockpit / network) is down during a margin event, leveraged positions are left unmanaged and the account can be liquidated — the single worst failure mode Gemini surfaced.
Therefore, once leverage is on, the system must be fail-safe, not merely fail-closed:
- No leveraged position without a broker-side resting protective order. The deterministic engine places a GTC stop / bracket protective leg at the broker, atomically with entry — so the position is protected even if your entire system is offline. The broker is the last line of defense.
- Options carry their protection structurally — defined-risk only (long call / debit spread), so max loss is fixed without active management.
- Reconcile resting protection every cycle. If a protective order is missing, rejected, or stale, auto-flatten or block — it must not run naked leveraged exposure.
- Two distinct guarantees, both required under leverage: fail-closed (no new risk on failure) protects against bad entries; fail-safe (resting broker-side stops) protects existing leveraged risk during a total outage.
Accounting & eval rework (do this BEFORE UI or live)
Section titled “Accounting & eval rework (do this BEFORE UI or live)”Options/margin break the accounting model, not just sizing (codex, Gemini, GPT-5.5 all flagged it). Required before any multiplier is enabled:
- Position-level mark-to-market; margin/buying-power tracking; expiry/assignment & corporate-action handling.
- Risk-adjusted eval metrics (max drawdown, return/drawdown, not just P&L), measured with leverage ON.
- Otherwise the cockpit will show confident, wrong P&L and the paper eval won’t transfer to live.
Paper-first, live-ready
Section titled “Paper-first, live-ready”One broker-interface abstraction; the same RiskState engine + caps run in paper and live. “Easy switch to live” = credentials/flag swap + stricter caps + the live Go/No-Go. The double-gate stays.
Biggest risk
Section titled “Biggest risk”Max leverage + options + 3x ETFs on a system with no proven edge = a fast, correlated path to ruin (overnight gap · gamma · assignment · L-ETF decay) — possibly faster than HITL or the 2-week eval can react. Mitigations: deterministic auto-deleverage (above); eval gates on drawdown; enable multipliers one at a time (margin → L-ETF → options), each behind its caps + tests; start with tiny capital.
Where it lands in the plan
Section titled “Where it lands in the plan”The RiskState engine + accounting rework slot into Phase 2 (the deterministic right side) and must be green before any multiplier is enabled. Multipliers are added incrementally, each behind its caps and adversarial tests. See Build Plan and Cockpit & Frontend.