bison-strategy

name: bison-strategy description: >- BISON v3.0.1 — Conviction Holder. Patient multi-asset (BTC/ETH/SOL) conviction trader. MIN_SCORE 11 across 9 components (4H/1H structure, momentum, SM alignment, funding, volume, OI proxy, RSI). Conviction- scaled margin tiers (25%/31%/37%) and 7-10× leverage. DSL is wide by design — time-cuts DISABLED — so a real directional thesis can run multiple days into the move. Few trades. Long holds. Big moves. v3.0.1 ships a race-window dedup cache that eliminates the ENGINE_FAILURE retry noise on already-held assets, plus a doubled DSL exit interval to throttle REDUCE_ONLY spam when runtime position-state lags HL. license: Apache-2.0 metadata: author: jason-goldberg version: "3.0.1" platform: senpi exchange: hyperliquid requires: - senpi-trading-runtime>=1.1.0 - senpi_runtime_helpers

🦬 BISON v3.0.1 — Conviction Holder

Asset whitelist (BTC/ETH/SOL). Conviction floor (minScore 11). No time-cuts — Phase 1 + Phase 2 ratchet ladder own all exits.

v3.0.1 (2026-05-18) — held-asset dedup race-fix + DSL throttle

Operational reliability patch. NO thesis change. NO scoring change.

Bug observed. Audit on Bison12 (M192943) 2026-05-13 to 2026-05-17 showed 16 create_position ENGINE_FAILUREs carrying BISON_CONVICTION reasoning text, all for assets the wallet already held. Each failure was a runtime LLM-gate fire on a producer-emitted signal that should have been deduped — the producer's on-chain heldAssets check leaked during the race window between push_signal() returning OK and the resulting position appearing in the next-tick clearinghouseState pull.

Fix 1 — recent-signals cache (scripts/bison_config.py). record_signal(coin) writes {coin: epoch_seconds} to state/recent-signals.json after every successful push. main() calls was_recently_signaled(coin) BEFORE build_thesis() and skips any coin seen within RECENT_SIGNAL_TTL_SEC (default 180s ≈ 3× the typical ALO open-fill window). Skipped coins are echoed in recently_signaled_skipped in the per-tick output for audit.

Fix 2 — DSL exit interval 30s → 60s (runtime.yaml). Audit also showed clusters of CLOSE_FEE_OPTIMIZED_FAILED "Reduce only order would increase position" + CLOSE_NO_POSITION errors firing 1-2 ticks after a successful close. The runtime's view of the closed position takes longer to refresh than the 30s DSL interval, so DSL re-fires close on a flat HL position. Doubling the interval gives the runtime position-state pull room to catch up and halves the wasted close attempts. Root cause (runtime- side position-state sync) is escalated to the runtime team — this YAML change is downstream mitigation only.

What this does NOT fix. Phantom orphan positions where DSL state and HL state are durably out of sync (i.e. runtime thinks a position is open after a complete external close, or vice versa). That class of bug needs runtime-side reconciliation; the producer has no authority over DSL state.

v3.0.0 (2026-05-12) — plumbing-only migration

NO thesis change. v2.1 asset whitelist, minScore 11, conviction-scaled margin tiers, direction waterfall, 9-component scoring, and DSL preset all preserved verbatim. Six-layer plumbing flip:

1. MCP transport: mcporter subprocess (2.5-5s cold-start per call) → senpi_runtime_helpers.SenpiClient.mcp_call() in-process HTTPS (~280ms).
2. Signal emit: scanner called create_position directly (Wolverine pattern); producer now emits via push_signal() to runtime /signals. Runtime LLM-gated bison_entry action opens via FEE_OPTIMIZED_LIMIT, passing marginUsd + leverage through from signal data for tier-based sizing.
3. Reentrancy: no v2 lockfile to replace (Bison v2.x ran on openclaw cron with no producer-side lock); producer_daemon owns the per-tick scanner_lock with stale-PID auto-recovery.
4. Scheduler: openclaw cron (5 min) → producer_daemon(interval_seconds=300). Long-lived process; zero per-tick cold-start cost.
5. Risk gates: Python MAX_DAILY_LOSS_PCT, dynamic-slots PnL-based cap, per-asset cooldown state files → declarative risk.guard_rails block. state/trade-counter.json and state/asset-cooldowns.json are vestigial in v3.0 and can be deleted.
6. Exit fee: DSL exits switched from MARKET (taker, 0.045%) to FEE_OPTIMIZED_LIMIT (maker-first, 0.015%, 60s ALO timeout, taker fallback). Entry exits keep ensure_execution_as_taker: false (v2.1 ALO patience rule).

CRITICAL RULES

RULE 1: Producer enters. DSL exits. Producer NEVER exits positions.

v1.x had thesis re-evaluation that chopped winners. v2.0 removed it. v3.0 makes it structurally impossible — the producer has no close_position call site.

RULE 2: Producer emits signals via push_signal(); runtime opens

The producer scores conviction theses and emits via client.push_signal(). The runtime's LLM-gated bison_entry action opens the position via FEE_OPTIMIZED_LIMIT with ensure_execution_as_taker: false (v2.1 patience rule preserved — entry ALO waits 30s for maker fill).

RULE 3: All signals are score contributors, not hard gates

4H trend, 1H trend, 1H momentum, SM direction, funding — all add/subtract points. The minScore threshold (v2.1+: 11) is the ONLY gate. Nothing kills a signal before scoring.

RULE 4: Asset whitelist enforced (v2.1+)

Only BTC/ETH/SOL by default (configurable via allowedAssets in bison-config.json). Pre-v2.1 fetched top 10 by 24h notional volume — let small-caps that volume-spiked into the universe (e.g. ZEC on May 4) consume the single daily slot.

RULE 5: ensureExecutionAsTaker = false on ENTRIES

Entries use FEE_OPTIMIZED_LIMIT with ensure_execution_as_taker: false (v2.1 patience rule). Exits use ensure_execution_as_taker: true (closes MUST happen — taker fallback is the safety floor).

RULE 6: Use the DSL preset in runtime.yaml — not a custom DSL state file

v2.x shipped a hardcoded DSL state template in scanner output. v3.0 uses the runtime YAML's dsl_preset block as the single source of truth. No scanner-side DSL state.

How BISON v3.0 trades

Entry (all score contributors — preserved from v2.1)

Min score: 11. Max possible: ~16.

Direction determination (waterfall)

1. 4H trend structure → LONG if BULLISH, SHORT if BEARISH
2. If 4H NEUTRAL → follow SM direction
3. If SM NEUTRAL → follow 1H momentum (>0.5%)
4. If all neutral → no signal

Conviction-scaled margin

Producer computes marginUsd from tier; signal data carries it; LLM gate passes through to runtime payload. Same pattern for leverage (default 10x, MIN 7x, MAX 10x).

Exit — RatchetStop only

DSL preset wide-by-design for conviction breathing:

• +10% ROE: no lock (confirms working)
• +20% / +25% lock
• +30% / +40% lock
• +50% / +60% lock
• +75% / +75% lock
• +100% ROE: lock 85% — infinite trail

Phase 1 max_loss: 30%. Time-cuts: hard_timeout + dead_weight_cut DISABLED. weak_peak_cut KEPT (self-limiting at 60min/3.0% peak).

Risk gates (runtime.yaml risk.guard_rails)

v2.1's PnL-aware dynamic daily cap (3 base / 6 hard cap, tiered by realized PnL) is dropped in favor of static max_entries_per_day=3 + drawdown_halt_pct=25. The conservative ceiling and -25% hard stop are preserved; the "ride the hot hand" tier is sacrificed for state-file-free reliability.

Hardcoded constants (not configurable)

• MAX_LEVERAGE: 10
• MIN_LEVERAGE: 7
• XYZ_BANNED: true (banned at producer scan level)

Operator deployment topology

Bison's runtime + producer pair runs per strategy wallet, not per operator. Operators wanting more capital exposure deploy multiple strategy wallets, each running its own runtime instance and its own bison-producer daemon. Each instance:

• Reads its own ${WALLET_ADDRESS} (typically via env var) — the runtime YAML's ${WALLET_ADDRESS} placeholder is resolved at runtime create time, and the producer falls back to bison-config.json's wallet field.
• Computes held_assets from its own wallet's on-chain clearinghouseState — instances do NOT see each other's positions.
• Writes its own state/recent-signals.json (per-instance cache).
• Reuses the same SENPI_AUTH_TOKEN (operator-scoped) — a single operator can run N strategy wallets under one auth token.

Implication: two-wallet deployment doubles capital exposure and parallelizes asset selection. Wallet A may take SOL while Wallet B simultaneously takes BTC, capturing uncorrelated entries. Each wallet's DSL ratchet is independent.

This topology is the production reality on Bison's live deployment (see Predator MCP bison-v1-0 slug, strategyWalletCount: 2). It is NOT a Bison-specific feature — it's how the senpi-trading- runtime plugin natively scales: one runtime per wallet, multiple wallets per operator.

Signal cadence (what to expect)

• Producer tick: every 300s (5 min). Long-lived daemon — no per-tick cold-start cost.
• Universe scanned per tick: ≤ 10 assets, filtered to the whitelist (default BTC/ETH/SOL = 3 assets).
• MCP cost per tick: ~4-7 calls (market_list_instruments, leaderboard_get_markets, market_get_asset_data per candidate, strategy_get_clearinghouse_state for held check).
• Emission rate: 0-5 signals per day in typical regimes. The MIN_SCORE 11 floor + 2h per_asset_cooldown keeps emission selective.
• Quiet day signature: every tick output ends with "note": "WAITING — no conviction thesis (min score 11)". This is the most common state — the producer is alive, the thesis is just not firing.
• Suspected silent producer: check audit_query({tool_name: "market_get_asset_data", user_ids: [<MID>]}) — if entries fall to zero for >15 min during market hours, the daemon has died and needs disown re-launch (fleet pattern from 2026-05-13).

Reading the decision log

Bison emits per-trade decision telemetry via the runtime LLM gate. To inspect the reasoning chain for a specific window:

mcp__senpi-prod__audit_query({
  user_ids: ["<your Senpi MID>"],     # e.g. "M192943" for Bison12
  tool_name: "create_position",
  action_type: "create",
  limit: 50
})

Each entry's ai_reasoning field is the LLM gate's reasoning output — typically "BISON_CONVICTION <ASSET> <DIRECTION> - <top_score_reason>". Successful entries have success: true; failures carry an error_code and error_message.

Pair with close_position audit queries to see the DSL exit reasons (weak_peak_cut, dsl_breach, etc).

The audit-query path is the canonical post-hoc validation tool for Bison — the Predator MCP shows aggregates (PnL, ROI, volume) but does not expose per-trade reasoning. For "why did this trade fire" questions, always go to audit_query.

v3.0.0 (2026-05-12) — plumbing-only migration

Hard rule for user-conversation Claude sessions

User-conversation Claude sessions MUST NOT call: 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.

Operator install

See README.md for fresh-install + migration commands from v2.1.

Changelog

v3.0.0 (2026-05-12) — plumbing-only migration

Migrated to helpers-native (senpi_runtime_helpers). Thesis preserved verbatim. See six-layer flip above.

v2.1 (2026-05-07)

• Asset whitelist (BTC/ETH/SOL); pre-v2.1 fetched top-10-by-volume which let small-caps like ZEC consume the daily slot
• minScore 8 → 11; was firing on first mediocre setup post-midnight, now demands real conviction
• DSL time-cuts disabled in runtime.yaml per single-asset pattern

v2.0 (2026-04-06)

• ALL hard gates converted to score contributors
• Thesis exit (evaluate_held_position) REMOVED
• Scanner calls create_position internally (Wolverine pattern; superseded in v3.0 by helpers-native push_signal)
• Direction waterfall: 4H → SM → 1H momentum

License

Apache-2.0 — attribution required for derivative works. Copyright 2026 Senpi (https://senpi.ai).

Skill Attribution

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