Lion Strategy

🦁 LION v1.0 — Two-Speed-Market (K-Shaped) Cross-Asset Long/Short

Lion bets on a K-shaped divergence: the AI complex and HYPE/SOL keep booming while the broad U.S. economy and the rest of crypto struggle. It expresses that as a single thematic long/short book:

• LONG the "haves" — the AI complex (semis / AI hardware / AI infra / AI

software) plus crypto's winners (HYPE large, SOL modest).

• SHORT the "have-nots" — the broad U.S. market (the SP500 index product)

plus a curated basket of laggard crypto alts.

The P&L is the dispersion between the two speeds, not the market's direction. One producer script (lion-producer.py) serves both books; the LION_LEG env var selects which.

| Book | Holds | Wallet env | Runtime | Scanner | |---|---|---|---|---| | long | AI complex + HYPE/SOL (LONG) | LION_LONG_WALLET | runtime-long.yaml | lion_long_signals | | short | SP500 + laggard alts (SHORT) | LION_SHORT_WALLET | runtime-short.yaml | lion_short_signals |

Why this is a cross-asset fund

Unlike the other long/short funds, Lion's two universes span both tokenized U.S. equities (Hyperliquid XYZ) and main-DEX crypto on one cross-margined wallet. _dex_for() routes each asset to its DEX automatically. The get_positions() helper reads both the main and xyz clearinghouse sections (two views of one wallet) and takes account value via max(), never the sum.

The thesis, asset by asset

| Sleeve | Direction | Default names | Rationale | |---|---|---|---| | AI chips | LONG | NVDA, AMD, MRVL, ARM, AVGO, INTC, TSM, ASML, CBRS | The compute layer of the AI boom | | AI memory | LONG | MU, SMSN, SKHX, SNDK | HBM/DRAM/NAND — the AI-capex supply squeeze | | AI infra / cloud | LONG | CRWV, NBIS, DELL, LITE | GPU cloud, AI servers, datacenter optics | | AI software / hyperscalers | LONG | GOOGL, MSFT, META, AMZN, ORCL, PLTR, NOW, IBM | The model + cloud layer | | Frontier compute | LONG | SPCX (owns xAI), QNT (quantum+ML) | Speculative AI optionality — sized smaller | | Crypto winners | LONG | HYPE (1.5× size), SOL (0.6× size) | HYPE keeps booming; SOL modest growth | | Broad U.S. market | SHORT | SP500 (1.2× size) | The rest of the economy suffers | | Laggard crypto | SHORT | ETH, XRP, DOGE, AVAX, LINK, ADA, LTC, NEAR, APT (0.7× size) | The rest of crypto struggles vs HYPE/SOL |

The AI long basket is the full live-board AI complex (27 names), resolved against the trade.xyz instrument board on 2026-06-17. Speculative frontier names size smaller (SPCX 0.6×, QNT 0.5×, CBRS/NBIS 0.7×); the core megacap chips and hyperscalers run full slots. The universe auto-extends as new AI names list.

The long-AI / short-SP500 overlap is intentional. SP500 contains the AI names; longing NVDA while shorting the index isolates the pure AI-outperforms-the-broad-market spread — exactly the bet.

How it picks (producer-side)

1. Curated thematic universe — config.universe (the haves for the long

leg, the have-nots for the short leg), intersected with the live instrument board + a relative-to-market liquidity gate (an instrument's 24h volume must be ≥ volFloorPctOfMedian × the universe's median volume — no hardcoded dollar floor; it adapts to the market and drops anomalously-thin names). New listings auto-join once added.

1. ABSOLUTE trend is the gate — long a have only while its 4h structure is

not bearish; short a have-not only while its 4h is not bullish. This is the hard gate: never long a downtrend, never short an uptrend, even for a thesis name.

1. Relative strength is a tiebreaker — cross-sectional excess vs the

leg-universe mean tilts ranking and score, but does not bench a genuinely-trending winner on a day its peers ran harder.

1. Conviction sizing — `margin = account_value × marginPct ×

sizingWeights[name]`. HYPE big, SOL small, SP500 core. No hardcoded dollars; every size scales with the budget.

1. Guards — RSI blow-off guard (long) / capitulation guard (short); held-asset

dedup; per-candidate affordability cap (never emit an order the wallet can't fund).

Net exposure is YOUR dial

Net exposure (the long/short balance) is an operator decision, not a hardcoded constant — see each config's _net_exposure_note. The two books run on separate wallets, so you set the posture by:

• how much capital each wallet holds (the primary lever), and
• the per-leg knobs: maxSlots, marginPct, sizingWeights.

The shipped default is a modest net-long tilt (~60/40): the long book carries more slots (5 vs 4), higher margin (18% vs 15%), and a higher leverage cap (5x vs 4x) than the short book. To run market-neutral, fund the wallets equally and equalize slots × marginPct. To run net-short, tilt capital toward the short book.

Fleet-standard rules (enforced)

• Max leverage 5x long / 4x short (strict clamp + runtime gate; venue caps below).
• Per-position margin ≤ 18% long / 15% short (× conviction weight, ≤ 25% cap).
• Drawdown halt 20% long / 18% short from rolling peak, reset_on_day_rollover.
• Mandatory DSL; entries + exits use FEE_OPTIMIZED_LIMIT with taker fallback.
• Verbose per-tick JSON (never silent): `scanned / ranked_pool / candidates /

signalspushed / emitted / meanrs_24h`.

• Sizes off max(main, xyz) account value — never the sum (cross-margin

two-views double-count fix).

• Short squeezes respected: short book runs tighter (lower leverage, tighter

max-loss, faster stall-cuts, smaller alt weights, BTC omitted from the default short basket).

Hard rule — user-conversation sessions are READ-ONLY

A Claude session conversing with a user MUST NOT call create_position, close_position, edit_position, ratchet_stop_*, cancel_order, or any strategy_close* tool against Lion's wallets. Entries are emitted only by the producer daemon; exits are owned only by the runtime DSL.

v1.0 — initial build

Cloned from the Cougar helpers-native two-book pattern, generalized from a single-universe equity cross-section to a cross-asset thematic book (xyz equities + main-DEX crypto), with absolute-trend gating (vs Cougar's pure cross-sectional gate) so a thesis winner is held while it trends, and per-group conviction sizing weights so HYPE is sized larger than SOL without any hardcoded dollar amount.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: Senpi-ai
• Source: Senpi-ai/senpi-skills
• License: MIT

Install and usage instructions live in the source repository linked above.