Deleveraging-Confirmed Idiosyncratic Down-Shock Rebound on Hyperliquid, BTC-Beta-Hedged: long an HL alt perp + short beta-weighted BTCUSD.HYPERLIQUID at the close of a UTC day whose BTC-residual return is <=-2 sigma AND whose Binance USD-M open interest (coin units) fell >=5% the same day, fixed 3-day clock, 10% per leg, max 6 concurrent (55-name point-in-time HL panel, 1-DAY)

Original Research Lead hypothesis (not seeded). It combines two pieces of the factory's own evidence that have each been promoted separately but never together, on a venue below its quota (HYPERLIQUID 14% vs >=20%). (1) The BTC-residual, BTC-beta-hedged long down-shock on Hyperliquid was promoted, but it was weak: 547 trades, PF 1.08, 0.16% per trade, 12h clock, no positioning filter. (2) Lesson L254: open interest works as a CONFIRMATION filter on a one-direction event. The Binance SPOT capitulation buy, taken only when same-day USD-M OI fell >=8%, was promoted with PF 1.65 and 1.86% per trade. MECHANISM: a coin that drops far more than BTC explains on a day its derivative open interest collapses is being sold by forced or panicked closing of leveraged longs (liquidations, margin calls), not by new informed shorts. Such flow is price-insensitive and finishes when the positions are gone, so the residual overshoot partly reverts over the next few days. When OI does NOT fall, the drop is more often new shorting on information, which does not revert; this is why the filter should separate the two. The trade is hedged with BTC so it earns the residual rebound rather than crypto beta (L256: a residual-defined trigger with a BTC hedge leg). SIGNAL (computed at each 1-DAY bar close, 00:00 UTC). For each panel alt i: r_i is the close-to-close log return and r_btc is BTCUSD.HYPERLIQUID's log return. beta_i is the 90-day rolling OLS slope of r_i on r_btc, lagged one bar and clipped to [0.5, 2.0]. The residual is e_i = r_i - beta_i x r_btc, and z_i = e_i / (60-day rolling std of e_i, lagged one bar). The OI change is dOI_i = ln(OI_i at the bar close / OI_i 24h earlier), using the Binance USD-M open interest in coin units for the same coin (supplementary key open_interest / metrics, symbol <COIN>USDT). Read it ONLY through supp_as_of(): the latest 5-minute observation at or before the bar close, never a future one. A shock is z_i <= shock_z (baseline -2.0) AND r_i < 0 AND dOI_i <= -oi_drop (baseline 0.05). If OI data is missing at the bar, there is no trade. EXECUTION: Hyperliquid trades 24/7, so the capability next_bar_open_fill is use_equivalent. Entry is a market order at the signal bar's close, which on a continuous market is the next bar's open; this substitution is named here. Long the alt perp at 10% of equity notional, and in the same step short BTCUSD.HYPERLIQUID with notional beta_i x the alt notional. Exit both after exactly 3 daily bars by market order at the close of the 3rd bar after entry. There is no stop and no target: risk-control route (b), a time exit plus capped sizing. The BTC leg is one NETTED short (hedge_mode is unsupported): its target size is the sum of the hedges of the alts currently open, re-sized on every alt entry and exit, and the strategy must check legs_balanced(). CONCURRENCY CAP (L267): at most 6 alt legs open at once. When more names qualify on one date than there are free slots, take the most negative z first. Maximum same-date gross is therefore 6 x 10% long alts plus about 6 x 10% x beta of BTC short, roughly 60% long / 60-80% short, about 1.4x gross and near zero net beta. That stays inside the 5x cap and the 8-daily-move liquidation buffer. PANEL (pre-registered, point in time): 55 Hyperliquid perps that also have a Binance USD-M OI series. A name is tradable only while is_listed() and once it has 120 days of its own HL bars (listed_since()). The panel: SOL AVAX LINK DOGE ARB OP SUI APT INJ TIA SEI WLD NEAR LDO AAVE UNI LTC BNB XRP ADA DOT ATOM FIL ORDI WIF ENA JUP CRV ETC BCH TRX STX RUNE PENDLE ONDO TAO FET MKR GALA APE DYDX IMX BLUR MATIC HBAR TON EIGEN JTO PYTH W STRK ZRO NOT POPCAT ETHFI. MATIC, delisted on HL, stays in up to its delisting. HL 1-DAY bars exist before the 2023-05 HL listings, as noted in the HL weekend-pump failure: NO entry may be taken before a name's HL listing plus 120 days. The usable span is therefore about 2023-09 to 2026-10. PRE-STUDY (frozen baseline rule, measured on these exact names after each name's HL listing + 30d, 2023-06 to 2026-10, close-to-close, gross of costs, NO concurrency cap; used only to register sample size and a prior, not to choose parameters). z<=-2 with no OI filter: 1,136 events on 350 dates, beta-hedged 3-day return +0.95%. z<=-2 with dOI<=-5%: 363 events on 141 DISTINCT dates, hedged 3-day mean +2.08%, median +1.35%, raw unhedged +3.21%, naive t 4.5. By year (hedged mean): 2023 -0.6% (32 events), 2024 +1.8% (165), 2025 +4.4% (101), 2026 +0.5% (65). The OI gate roughly doubles the hedged per-event return over the ungated trigger at every threshold checked (z -2 / -2.5, OI -3/-5/-8%). This is the testable claim. Distinct event dates (141, about 41 a year) clear lessons L237/L259's ~30 floor. The events cluster on market-wide deleveraging days, which is why the concurrency cap and the BTC hedge are part of the design. COST ARITHMETIC: each round trip pays the HL taker fee on two legs: alt 0.09% of notional plus BTC about 0.09% x beta, so about 0.20% of alt notional, plus 3 days of funding on both legs, which roughly nets out. Against an expected realised capture of about 1.0% per alt leg after the cap (half the pre-study gross, to allow for close-vs-fill slippage on crash days, the 2026 decay and in-sample optimism), the edge is about 5x the round trip. PRE-REGISTERED DIAGNOSTICS for the analyst. (a) The same rule WITHOUT the OI gate, run over the same span: if the OI-gated per-trade return is not clearly above the ungated one, the deleveraging premise is not shown. (b) Hedged versus unhedged per-trade return: the claim is about the residual. (c) Count of distinct entry dates, not only per-leg trades (L238).

Point-in-time panel of 55 Hyperliquid alt perps on 1-DAY bars, hedged with BTCUSD.HYPERLIQUID. At each 00:00 UTC close, for each alt: beta = 90-day OLS slope of its log return on BTC's, lagged one bar and clipped to [0.5, 2]; residual e = r - beta*r_btc; z = e / std of the residual over the prior 60 days (lagged one bar). A shock needs z <= shock_z (-2.0), r < 0, and a 24h log change in the same coin's Binance USD-M open interest (coin units) of at most -oi_drop (-5%). The OI is read at or before the close and at or before close-24h; each reading may be at most 3h stale, otherwise there is no trade. A name is eligible only while it is listed and at least 120 days after its HL listing. Qualifiers are ranked most-negative-z first, with at most 6 open alt legs. The strategy buys the alt at 10% of MTM equity and shorts beta x that notional of BTC through ONE netted BTC short, which is re-targeted to the sum of the open legs' hedges on every entry and exit. Each alt is closed at market at the close of the 3rd daily bar after entry. There is no stop and no target (risk route b: time exit plus capped sizing; leverage 2x, gross about 1.2-1.4x). legs_balanced() is checked every bar, and every leg is flattened on a failure. Execution substitution (next_bar_open_fill = use_equivalent): orders are market orders at the signal bar's close, which on the 24/7 HL market is the next bar's open.

Implements the hypothesis rules as written, with the optimization-plan parameter names. Fixed values: leverage 2, position_pct 0.1, beta 90/[0.5, 2], z_std 60, max_concurrent 6, min_listed_days 120. Tunable: shock_z, oi_drop, hold_days. OI WIRING (important for the analyst): the platform attaches per-leg open interest (open_interest_by_instrument) only for legs in extra_instruments on BINANCE/BINANCE_SPOT/BINANCE_CM. HYPERLIQUID legs and signal-only legs get none. The 55 matching <COIN>USDT.BINANCE perps are therefore listed in extra_instruments plus their 1-DAY bar types, as signal sources that are NEVER traded. This is the same wiring as the earlier HlBinanceOiFlushCapitulationPanel. Side effect: the runner funds a second, idle $100k BINANCE venue account, so the reported total_return and max_drawdown are measured on $200k while all trading and sizing use the HL account (get_account_equity reads the primary venue). Reported percentages are therefore about half of what the HL book earned. Per-trade returns, PF, win rate and Sharpe are not affected. OI is read point-in-time: the latest hourly observation at or before the close and at or before close-24h, bisected against the bar clock through an incremental cache that grows with the point-in-time view. It is not read through supp_as_of(), because that accessor serves only primary-symbol series. Eligibility uses is_listed() and listed_since() + 120 days. When there is no listing record, it falls back to 120 of the name's own bars. Only alt bars stamped exactly at the BTC close contribute; a name with no fresh bar is marked not-ready, so a stale z is never reused. Hedge: BTC quantity = beta x intended alt notional / BTC close, fixed at entry and released at exit. One netted short; a tolerance of 0.5% of equity avoids dust rebalances. Diagnostics in entry_diag: shocks_price (price-only shocks among eligible names), shocks_no_oi, shocks_oi_not_falling, shocks_oi_confirmed, entries_<year>, distinct_entry_dates and skipped_max_concurrent. These support pre-registered diagnostics (a) and (c). The ungated rule can be rerun by setting oi_drop to a large negative value; a missing OI reading still blocks the trade.