Position with the smartest money, against the crowd. WhaleHunterHedge measures what the cohort with the largest lifetime realized gains is actually doing — net long or net short, per asset, and whether they're adding — and fires when the smart money diverges hard from the crowd. The signal it's built to catch:
This week the most profitable wallets on Hyperliquid went heavily short while the crowd piled into longs — and the winners added to that short every single day. That's the setup: smart money net-short + adding, crowd net-long → short it with them.
| Sleeve | What it positions | Direction | Wallet |
|---|---|---|---|
| long | assets the smart cohort is net-long + adding | LONG only | one |
| short | assets the smart cohort is net-short + adding | SHORT only | one |
Two wallets, fully independent — so the book can hold conflicting positions on the same asset (smart money net-long ETH while net-short HYPE → long sleeve holds ETH, short sleeve holds HYPE, no netting). Funding default 50/50 — no built-in directional bias.
Why cohort divergence is the edge
A single whale's trade is noise. The weight of the entire >$1M-realized cohort is signal — it's the closest thing Hyperliquid has to a "commitment of the smartest traders." When that cohort is lopsided on an asset and growing the position day over day, while the crowd (the $10k–$100k wallets that are net buyers of every rally) is on the other side, you have a positioning divergence — the most durable read in the market. You're not chasing one bet; you're standing where the people who've actually made money are standing, and fading the people who haven't. Because it's a positioning signal (not a momentum one), a wide DSL fits: you hold while the smart cohort holds.
The engine — four steps, all from Senpi Discover
- Cohorts by lifetime realized gains. The ALL-TIME realized-PnL ranking, paged by
offset, bucketed by
$: smart = ≥smartMinRealizedUsd($1M), crowd =crowdMinRealizedUsd..crowdMaxRealizedUsd($10k–$100k). Paging matters: the ranking is realized-descending, so the smart cohort sits at the top but the crowd lives thousands of ranks deeper — a single top-N pull catches only a weak tail of the crowd (the inverted "crowd ≪ smart" symptom). The walk pages down (cohortMaxPages) until the crowd is representatively sampled or the ranking drops below the crowd floor, capping each cohort (cohortSampleCap) to bound per-tick load. Membership cached daily. - Net positioning per cohort, per asset.
discovery_get_trader_stateacross each cohort → sum signed notional per coin → bias = net / gross in [-1,+1] (+1 = all long, -1 = all short), plus member counts. - The "adding daily" trend. A daily cohort ledger snapshots the smart cohort's net
per coin; today's net minus the earliest snapshot in the window = growth. The smart
cohort must be growing its position in the signal direction (
requireGrowing). - The divergence strike. For this sleeve's direction: the smart cohort is net-directional
past
biasThreshold(default 0.50) and growing, scored higher when the bias is strong (≥0.7), when it's adding, and when the crowd diverges (net-opposite by ≥crowdDivergenceMin). Score ≥cohortMinScoreemits.
~1-day warmup: the growth gate needs ≥2 daily ledger snapshots, so a fresh deploy emits no divergence strikes on day 1 — it's building the baseline. This is by design.
Mirror + ride + surface
- Size to YOUR budget, score-scaled —
margin = equity × marginPct, scaled up +25% per point above the score floor, capped atmaxMarginPct. No hardcoded$. - Leverage capped — conviction shows in size, not leverage (clamp to
maxLeveragethen venue max). - Surfaces the insight — every tick the producer emits a human-readable
insightline per top divergence ("HYPE: smart -0.85 vs crowd +0.65, Δ-$2.1M → SHORT"), so the agent can report what the smart money is doing, not just trade it. - Wide DSL — wide disaster stop,
weak_peakOFF, time-cuts OFF, Phase 2 does nothing until a big run. Ride the divergence. Planned invalidation exit: close when the smart cohort flips or unwinds (the mirror of the entry).
Deploy — two wallets
WHALEHUNTER_LEG=long WHALEHUNTER_LONG_WALLET=<wallet A>
WHALEHUNTER_LEG=short WHALEHUNTER_SHORT_WALLET=<wallet B>
Fund 50/50 for a balanced smart-money long/short book. Requires a USER-scoped
SENPI_AUTH_TOKEN — the discovery_* tools need a valid user id (no user scope → the
smart cohort comes back empty and the producer reports "cohort too small").
Fleet-standard rules (enforced)
- Max leverage 5x (clamp + venue). Score-scaled margin (no hardcoded $); per-position cap
maxMarginPct; up tomaxSlots(6) per sleeve. - Drawdown halt 25%, baseline-seed guard, per-asset cooldown.
daily_loss_limit_pctdisabled on multi-wallet funds (perpDay base reads ~$0 → $0 limit; DSL + drawdown_halt protect). - Mandatory DSL; entries + exits
FEE_OPTIMIZED_LIMITwith taker fallback (a confirmed divergence must fill). - Sizes off
max(main, xyz)account value — never the sum (cross-margin). - Signature-adaptive daemon launch; per-(leg,wallet) lock; daily-cached cohorts + daily ledger.
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
WhaleHunter's wallets. Entries are emitted only by the producer daemons; exits are owned only
by the runtime DSL.
Versions
- v2.0 (current) — the cohort-divergence engine. Pivots from per-whale copying to
measuring the NET positioning of the >$1M-realized cohort vs the $10k–$100k crowd, per asset,
with an "adding daily" growth gate and a crowd-divergence booster — and surfacing it as a
readable insight. Positions WITH the smart money against the crowd. The v1.x per-whale
conviction copier is retained behind
enableIndividualCopy(OFF by default). - v1.2 — activity tune + near-miss observability for the per-whale copier (convictionPct 0.25→0.18, poolSize 30→50) after it sat 0 trades in 3 days (patient whales rarely open a ≥25%-of-book new position — the rarity that v2.0's cohort approach sidesteps).
- v1.1 — tiered sizing across four consistency×style tiers (
tagWeights). - v1.0 — patient-whale conviction copier (ELITE+PATIENT, conviction gate, wide-DSL ride).