cheetah-strategy

🐆 CHEETAH v7.2.0 — Multi-Signal Confluence Sniper

SM commits. Quality traders commit. Price confirms. Volume commits. All at once. Cheetah pounces once. The runtime DSL ratchets to lock the win.

v7.2.0 (2026-06-01) — DSL exit overhaul: disable time cuts, let winners run

Entries are fine. Exits were the bottleneck. A 100-trade-by-score analysis on the live wallet showed win rate scaling cleanly with score — 40.0% @ score 10, 41.7% @ score 11 (n=84), 50.0% @ score 12, 55.6% @ score 13 — confirming the accuracy filter works. But Avg ROE was negative across EVERY score bracket (−1.90% / −0.35% / −0.20% / −0.12%).

Score	Trades	Win rate	Avg ROE
10	5	40.0%	−1.90%
11	84	41.7%	−0.35%
12	2	50.0%	−0.20%
13	9	55.6%	−0.12%

Diagnosis: "death by a thousand cuts." The tight Phase 2 locks plus the 60m/90m time-cuts made it mechanically impossible to hold for a big win — Cheetah took the full brunt of −15% stops / −5% chops but was forced to give up every +15% / +30% runner. The two big HYPE longs that actually paid (+13.9%, +8.9%) only landed because they hit the 12h hard timeout — not a trailing stop, not a 60m dead_weight_cut. When forced to hold, the position printed.

Fix (mirrors the live deploy):

• weak_peak_cut → disabled (was 90m @ 3.0 — chopping winners-in-waiting)
• dead_weight_cut → disabled (was 60m — forcing exits before trenders accelerate)
• hard_timeout → kept (12h fail-safe — the one cut that captured the winners)
• Phase 2 ladder already widened to the "let winners run" shape (T0 +10% / lock 0 → +100% / lock 85) on 2026-05-21

Together this removes the glass ceiling so heavy trenders can run. NO change to scoring, MIN_SCORE, leverage tiers, margin, cooldowns, or risk gates — exit-mechanics only. Hypothesis test; re-evaluate after ~50 more trades.

v7.1.1 (2026-05-11) — positions parser fix (patch)

Live verification of v7.1.0 found positions_map was STILL empty even though the widened filter returned the full 25-trader pool. Root cause: leaderboard_get_trader_positions actually returns

{ "data": { "positions": { "trader_id": ..., "rank": ..., "positions": [...] } } }

— the per-trader positions array is nested one level deeper than the schema guide documents. The v7.0/v7.1 parser looked at data.positions expecting a list, got a dict, hit if not isinstance(positions, list): continue and silently skipped every trader.

v7.1.1 parser handles BOTH shapes:

• list-direct (matches schema guide)
• nested-dict {trader_id, rank, positions: [...]} (actual prod response)

Plus a POSITIONS_SHAPE_WARN log line for any future schema drift.

NO change to scoring components, MIN_SCORE, leverage tiers, margin, cooldowns, or DSL. Same thesis. Now the +3 QUALITY_TRADER bonus can actually fire when the per-asset overlap exists.

v7.1.0 (2026-05-11) — quality-trader filter widened (patch)

Post-migration diagnostic on the live wallet revealed quality-cache.json was always empty (positions_map = {}). Root cause: discovery_get_top_traders(time_frame=WEEKLY, consistency=[ELITE,RELIABLE], open_position_filter=True, limit=8) returned zero traders under the current platform regime. With no quality-trader pool, the +3 QUALITY_TRADER scoring bonus never fires; candidates cap at score 8 (needs 4-of-4 secondary confluence to reach the MIN_SCORE=10 floor without that bonus), and that 4-of-4 only hits in strong-trending regimes.

Filter changes (live-verified):

• time_frame: WEEKLY → MONTHLY — broader history, more traders accumulate ELITE/RELIABLE labels
• consistency: [ELITE, RELIABLE] → [ELITE, RELIABLE, STREAKY] — slight loosen at the lower end
• limit: 8 → 25 — wider pool surfaces more per-asset overlap
• open_position_filter dropped — the per-trader leaderboard_get_trader_positions fetch already filters flat traders implicitly via empty position arrays

NO change to scoring components, weights, MIN_SCORE, leverage tiers, margin, cooldowns, or DSL. Same thesis, wider input.

Telemetry: producer now emits a QT_POOL_WARN log line whenever the trader pool is unhealthily small (< QT_POOL_SIZE / 2) or positions_map ends up empty. Future regressions surface in one grep instead of post-hoc archaeology.

v6.0 thesis (preserved from v5.x)

Refuse to trade unless ALL major signals align simultaneously:

Component	Signal	Pts
SM consensus	pct_of_top_traders ≥ 10%, traders ≥ 25	+4 (hard gate)
Velocity	15m contrib ≥ 1.0 OR 1h contrib ≥ 3.0	+2 (hard gate)
Accelerating	15m > 1h > 0 (SM building)	+2
Dual price	4h move ≥ 2% AND 1h aligned same direction	+2
Volume	current ≥ 2x 6h avg	+1
Quality trader	≥ 1 ELITE/RELIABLE positioned same direction	+3
Rank climb	≥ 5 positions in last 2 scans, top 15	+1

Max score: 15. v6.0 floor: 10. Universe is the top 100 SM leaderboard (leaderboard_get_markets), refreshed every producer tick. XYZ permanently banned.

Score-scaled leverage:

Score	Leverage
10	3x
11	5x
12-13	7x
14-15	8x

Leverage is clamped to each asset's Hyperliquid max via strategy_get_asset_trading_limits (preserves v5.1.1 leverage safety fix).

v6.0 architectural changes

v1 scanner → v2 producer

• v5.x: cheetah-scanner.py scored signals, called create_position directly, maintained Python-side cooldowns/counters/resting-order guards
• v6.0: cheetah-producer.py emits signals via SenpiClient.push_signal() (direct HTTP POST). Runtime LLM gate (decision_mode: llm) is pass-through. Risk gates declarative in runtime.yaml.

MIN_SCORE 11 → 10

v5.2 tightened MIN_SCORE 10 → 11. Result: 8 consecutive days of zero trades. Pre-tightening (v5.0/v5.1 at MIN_SCORE 10) produced +$182 net across 11 trades with 60%+ win rate. v6.0 restores the trade flow.

FEE_OPTIMIZED_LIMIT on EXITS too

v5.x used MARKET (taker) closes. v6.0's DSL fires FEE_OPTIMIZED_LIMIT on every exit — maker-first, 60s ALO timeout, taker fallback. HL maker is 0.015% vs taker 0.045% = 0.020-0.030% saved per close.

Held-asset dedup (3-layer defense)

1. Producer pre-filter: held_assets from clearinghouse, drops candidates already held
2. LLM gate hard skip: heldAssets array in signal payload, LLM rejects if asset is held
3. Runtime per_asset_cooldown_minutes: 240 (known to silently not-enforce; layer 1+2 are the real defense)

Post-close cooldown (producer-side backstop)

Diff held_assets across ticks. When asset disappears from held set, record close timestamp. Producer skips emission for 240 min after any close. Backstops the runtime cooldown bug.

Reentrancy guard

producer_daemon owns a per-tick scanner_lock with stale-PID auto-recovery — replaces v6.x's hand-rolled fcntl lockfile. Prevents two ticks racing.

DSL preset (v7.2.0 — "let winners run" ladder, time cuts disabled)

Phase	Component	Setting
Phase 1	max_loss_pct	15%
Phase 1	retrace_threshold	10 (v6.x raise from 6)
Phase 1	consecutive_breaches	1
Phase 2 T0	trigger 10% / lock 0%	(lets a +10% mover retrace fully before any lock)
Phase 2 T1	trigger 20% / lock 25%
Phase 2 T2	trigger 30% / lock 40%
Phase 2 T3	trigger 50% / lock 60%
Phase 2 T4	trigger 75% / lock 75%
Phase 2 T5	trigger 100% / lock 85%	(apex)
hard_timeout	720 min (12h)	enabled (the only time cut left)
weak_peak_cut	—	disabled (v7.2.0)
dead_weight_cut	—	disabled (v7.2.0)

v7.2.0: weak_peak_cut + dead_weight_cut disabled. The 100-trade analysis showed the time cuts were the bottleneck — they forced exits before trenders could accelerate and capped the right-tail move that pays for everything. The two big HYPE wins (+13.9%, +8.9%) only landed via the 12h hard timeout, so that fail-safe is kept. Median trade gives back a bit; the right tail is now uncapped.

Risk gates (runtime.yaml risk.guard_rails)

Gate	Setting
daily_loss_limit_pct	25%
max_entries_per_day	8
max_consecutive_losses	3
cooldown_minutes (post-loss)	30
drawdown_halt_pct	25%
drawdown_reset_on_day_rollover	false (Roach Sat-Sun bleed lesson)
per_asset_cooldown_minutes	240 (4h)

Operator install

See README.md for fresh-install + migration commands from v5.2.

Fleet patches incorporated

• ✓ Held-asset dedup (3-layer: producer pre-filter, LLM gate, runtime per_asset_cooldown)
• ✓ Post-close cooldown (producer-side backstop for the runtime cooldown silent-enforcement bug)
• ✓ Fleet-standard T0/T1 ladder
• ✓ Reentrancy guard (v7.0.0: producer_daemon scanner_lock with stale-PID auto-recovery; replaces hand-rolled fcntl)
• ✓ Wallet-from-config (no hardcoding; senpi-skills is public)
• ✓ drawdown_reset_on_day_rollover: false (Roach lesson)
• ✓ senpi_runtime_helpers (v7.0.0: in-process MCP + signal POST; replaces mcporter + openclaw-CLI subprocess)
• ✗ FP-001 quiet hours — deferred until fleet telemetry can validate the time-of-day P&L correlation

Hard rule for user-conversation Claude sessions

User-conversation Claude sessions MUST NOT call any of: create_position, close_position, edit_position, ratchet_stop_add, ratchet_stop_edit, ratchet_stop_delete, cancel_order, strategy_close, strategy_close_positions.

These tools are reserved for the producer daemon (entry path) and the DSL ratchet engine (exit path). User-conversation sessions are read-only.

If the user asks a question that implies action ("anything close to triggering?"), respond by reading the current state — DO NOT execute. The producer daemon will handle real signals on its next tick.

Skill Attribution

When creating a strategy, include skill_name and skill_version in the call. See references/skill-attribution.md for details.

License

MIT — Copyright 2026 Senpi (https://senpi.ai)