wolverine-strategy

What this skill does

# 🦡 WOLVERINE v6.1.0 — HYPE Patient-Conviction

## v6.1.0 (2026-06-01) — Phase 1 retrace leash widened for HYPE volatility

**One isolated DSL change. NO scoring, gate, leverage, or Phase 2 change.**
`exit.dsl_preset.phase1.retrace_threshold` **8 → 15**.

**Why.** During HYPE's epic run Wolverine caught a +16.4% ROE winner but
left the bulk of the move on the table, and its recent window ran a 25% win
rate with losses clustered at −5% to −9% ROE. The provable structural cause:
at 5x leverage an 8-ROE-point retrace is a **1.6% HYPE price leash** (8 ÷ 5).
HYPE routinely wicks 3-5% inside a single 15m candle, so this single-breach
(`consecutive_breaches_required: 1`) retrace fired on ordinary noise — exiting
positions at small losses before they could become the BIG trend the wide
Phase 2 "let winners run" ladder exists to ride. The −5% to −9% loss cluster
is consistent with a trade peaking +2-3% ROE, pulling back 1.6% price, and
tripping the 8-point retrace.

**The fix.** 15 ROE pts = **~3.0% price leash at 5x** — roughly double the
buffer, still well below the 20% `max_loss_pct` hard stop so it remains a
distinct trailing-from-peak rule, not a collapse into the hard stop (which is
why 15 and not the 20 that would equal max_loss).

**Why ONLY this.** The recent window is n=12 (score buckets n=2-4) —
statistically indistinguishable from a 50% win rate, far too thin to justify
retuning Phase 2 locks, the entry score floor, or `MIN_4H_STRUCTURE`. This
retrace fix is the one change defensible on **stop-width-vs-volatility math
alone**, independent of sample size. Everything else waits for data.

**Decision rule.** Re-measure expectancy after 40-50 trades. A let-winners-run
trend agent is *supposed* to have a low win rate; it only fails if the wins
can't clear the loss size. Break-even at a 25% win rate needs a 3:1 avg-win/
avg-loss ratio (recent window was 1.56:1 — wins truncated). If, after this
fix and a real sample, expectancy is still negative, the trend-follower
archetype is wrong for HYPE and a rewrite is earned. Until then it's a
mis-calibrated strategy, not a broken one.

## v6.0.0 (2026-05-18) — Patient-Conviction adoption (Bison v3.0.1 port)

NO scoring change. NO thesis change. NO gate change. Behavioral
profile shifts from "4 entries/day on score ≥9 with tight T0 lock=35"
to "2 entries/day on score ≥10 with wide T0 lock=0 multi-day holds."

**Why.** Bison v1.0 (BTC/ETH/SOL conviction holder) shipped this
exact pattern on 2026-05-12 and 4 days later was sitting on a
**SOL SHORT at +93.5% ROE** that had been held 4 days through
chop, V-recoveries, and an intraday paper drawdown to -8%. The
SOL position rode SOL $92.30 → $83.67 (-9.4% spot) — a move that
every fast-rotation agent in the fleet (Cheetah, Scorpion,
Turbine) flipped out of within minutes of entry. Bison held
through the noise because its DSL gave the position room to
breathe past +10% retrace before locking anything. That T0
lock=0 first tier is the single highest-leverage knob in the
Bison ladder.

Wolverine has been red since the +12.8% HYPE-day post-mortem
(PR #281, 2026-05-15) tightened gates to reject the obvious
chase. The gates now correctly identify clean trends. What was
missing was the back-end discipline to let those trends run.
v6.0.0 ports the Bison back-end without touching the Wolverine
front-end.

**What changed (5 small edits):**

| File | Change |
|---|---|
| `runtime.yaml` `risk.guard_rails.max_entries_per_day` | 4 → 2 |
| `runtime.yaml` `risk.guard_rails.per_asset_cooldown_minutes` | 120 → 240 |
| `runtime.yaml` `exit.interval_seconds` | 30 → 60 (REDUCE_ONLY spam mitigation) |
| `runtime.yaml` `exit.dsl_preset.phase2.tiers` | wide Bison ladder (T0 lock 0 through T5 lock 85) |
| `config/wolverine-config.json` `minScore` | 9 → 10 |
| `scripts/wolverine-producer.py` + `wolverine_config.py` | recent-signals race-window dedup (Bison v3.0.1 pattern) |

**What did NOT change.**
- Six-gate entry validation (4h trend, structure ≥0.65, 1h-4h
  alignment, 15m momentum ≥0.15, base-tech floor, 4h magnitude
  ≥1.0%) preserved verbatim.
- SM hard block on opposing direction preserved.
- RSI hard gates (74 LONG / 26 SHORT) preserved.
- Multi-factor scoring (~17 max) preserved.
- Conviction-tiered leverage (3x standard / 5x apex) preserved.
  Bison uses 7-10x on majors; HYPE's volatility doesn't yet
  warrant that — we get the patient-conviction benefits from the
  DSL ladder, not from leverage.
- FP-001 quiet hours (00-04 UTC unless apex score 11+) preserved.
- All time-cuts remain DISABLED (hard_timeout, weak_peak_cut,
  dead_weight_cut from v3.0.4 sweep).

**Tradeoffs to monitor.**
- Wider T0 lock means more giveback on weak winners (a +12%
  mover that retraces to +0% will close at breakeven instead of
  v5.x's locked +4.2%). Net positive ONLY if the captured BIG
  trends exceed the lost small winners. Bison's track record
  validates this on majors; HYPE-specific data needed for 30+
  days post-deploy.
- Cutting entries 4 → 2 means more idle time when HYPE chops.
  Acceptable — Wolverine's red period under v5.x was driven by
  too-frequent entries on borderline setups, not too-rare ones.
- The 4h cooldown stacks with the recent-signals dedup —
  Wolverine should fire HYPE at most every 4h regardless of
  signal cadence.

**Rollback condition.** If v6.0.0 produces zero or one entry per
week for 2 consecutive weeks, the floor is too high. Drop
`minScore` back to 9 in `wolverine-config.json` (the producer
floor remains 9; only the operator's selectivity is config-
driven). Other deltas should be preserved.

## v5.0.0 (2026-05-08) → v5.0.1 (preserved)

**v3 → v4 architectural rewrite.** v3.x was a full-agency scanner. v4.0 flips to the standard senpi-trading-runtime plugin pattern: producer emits signals, runtime owns execution + state.

**What changed structurally:**
- `wolverine-producer.py` (NEW) replaces `wolverine-scanner.py` (DELETED)
- v2-runtime-native: external_scanner + LLM-pass-through gate + native `risk.guard_rails`
- DSL exits via `FEE_OPTIMIZED_LIMIT` (saves ~0.020-0.030% per maker-filled close)
- Trade chain DB emits per-trade telemetry — chain DB visibility on Wolverine for the first time
- Python-state-crash class of bug (load_tc / set_cooldown / has_resting_orders) is structurally impossible in v4.0

**What's preserved from v3.0.3/v3.0.4 EXACTLY:**
- HYPE single-asset thesis
- **Six-gate entry validation:**
  1. 4h trend != NEUTRAL
  2. 4h structural strength ≥ 0.75
  3. 1h direction matches 4h
  4. 15m momentum ≥ MIN_MOM_15M (0.15) in direction
  5. Base-tech floor (strong_15m OR aligned_5m)
  6. **4h MAGNITUDE ≥ 1.5%** — the v3.0.3 fix that rejects dead-flat chop. Wolverine's own 2026-04-23 self-diagnostic: all 6 Week 5 trades died because trend_strength_4h is structural (lower-highs count) and passes even when 4h price magnitude is nearly flat.
- SM HARD BLOCK if direction opposes
- RSI hard gates (74 LONG / 26 SHORT)
- Multi-factor scoring (~17 max points): base-tech + SM concentration + SM velocity + funding alignment + funding regime + funding persistence + volume + OI velocity + BTC correlation + RSI room + 4h momentum bonus + move-exhaustion penalty
- MIN_SCORE = 9 (config-overridable)
- Conviction-tiered leverage: 5x apex (score ≥11) / 3x standard (≥9)
- DSL preset preserved: time-cuts ALL DISABLED, Phase 1 max_loss 20% / retrace 8 (v6.1.0: → 15) / 3 breaches, Phase 2 ladder (10/15, 20/35, 35/55, 55/70, 80/85)

**v3.0.1/3.0.2/3.0.4 v1-DSL fixes preserved:**
- `dead_weight_cut`: DISABLED (single-asset has no rotation cost)
- `hard_timeout`: DISABLED (v1 DSL fired this in Phase 2 incorrectly)
- `weak_peak_cut`: DISABLED (v3.0.4 — completes time-cuts sweep)
- All exits now 100% price-action

---

## ⛔ Hard Rules (Fleet Patches)

### RULE FP-002: User-conversation Claude sessions MUST NOT trade

**Hard rule, not a heuristic.** When responding to a user message, the Claude Code session MUST NOT call any of:

- `create_position`
- `close_position`
- `edit