{ "direction": "short", "leverage": 750, "take_profit_bp": 15, "take_profit_price": 81477.1, "bust_distance_bp": 8.5333, "bust_price_estimate": 81669.13, "expected_net_return_on_margin_pct": 86.79, "confidence": 0.38, "regime": "ranging", "key_levels": { "support": 81460, "resistance": 81632 }, "entry_mid": 81599.5, "data_time_utc": "2026-10-17T02:09:12.948Z", "reasoning": "This fades a failed spike in a quiet weekend range inside a 24h downtrend. The 02:03 UTC push to 81,632 was rejected within a minute, price is pressing the falling 15m EMA21 (~81,580) with the 24h change at -65 bp, and Hyperliquid open interest dropped ~30 BTC while price bounced, which looks like short covering rather than fresh buying. 1m realized vol is 2.1-2.8 bp/min (30-min 1-sigma about 12-15 bp), so 750x is chosen on merit: its 8.53 bp cushion puts the bust ~4.5 bp above the 81,632 rejection high, whereas at 850-1000x the fixed 4.8 bp bust offset leaves 5.2-7.0 bp, which puts the bust only 1-3 bp above that high. The 15 bp target (81,477, just ahead of the 81,452-81,460 range-low bids) keeps ~77% of the raw gain after the haircut for ~87% net on margin, versus 29-55% for 6-8 bp targets that are taxed much harder.", "risks": "The main liquidation path is the 02:00 UTC breakout continuing: price is above session VWAP (+6.5 bp) and the 1m/5m EMAs, with a higher low at 81,566 and a bid-heavy book within 10 bp (imbalance +0.40), so a retest and break of 81,632 toward the 81,700-81,800 overnight highs reaches the 81,669 bust. Thin weekend liquidity means one aggressive order can move the mid 5-8 bp in a single minute, as it did at 02:03 UTC." }

Evidence appendix: BTC-USD perp decision on PaperTrade, 2026-10-17 02:09 UTC

The JSON object above is the decision, with the exact keys the brief asks for. A file containing only that object is at decision/decision.json in the repository. Everything below supports it. Each item is marked Fact (measured, with its source), Inference (my reading of the facts), or Uncertain/Open (not known or not checked).

1. Facts: measured market state

Unless a row says otherwise, the source is the Hyperliquid public info API (POST https://api.hyperliquid.xyz/info), snapshot fetched 2026-10-17T02:09:12.948Z. The raw responses are in evidence/hl-snapshot-0209Z.json, and the derived metrics in evidence/analysis-0209Z.json.

Payoff arithmetic (Fact, from the brief's formula, computed in tools/decide.mjs): at 750x, bust distance = 10000/750 − 4.8 = 8.5333 bp, which puts the bust at 81,669.13. TP 15 bp → 81,477.10. Net move after the 0.2 bp deadband is 14.8 bp, scale × 0.98 = 0.782, so net return on margin = 750 × 14.8 bp × 0.782 = 86.79%. The bust price at each leverage from this mid: 750 → 81,669; 800 → 81,662; 850 → 81,656; 900 → 81,651; 1000 → 81,642.

2. Inferences: why short, why 750x, why 15 bp

• Direction (short), Inference. The higher timeframe is bearish: −65 bp over 24h, 15m EMA9 below EMA21, and lower highs from 82,898 (Fri 11:15Z) to 81,800 (22:30Z) to 81,647 (00:30Z). The 02:01–02:03Z push to 81,632 was a thin-market spike that was sold within a minute. Price is now back at the falling 15m EMA21 (~81,580). HL OI fell ~30 BTC while price recovered, and Binance OI fell 0.9% over 3h. Falling OI on a bounce is more consistent with shorts covering than with new longs building, and such bounces tend not to carry in a weekend Asia session. The 12h and 4h empirical grids (below) also favored shorts. Taken together, I read this as a range fade inside a downtrend. The counter-evidence is real: short-term momentum (+14 bp/30m), price above session VWAP and the 1m/5m EMAs, a higher low at 81,566, a bid-heavy near book, and expanding 1m vol and volume. That is why confidence is low.
• Leverage (750x), Inference. The 4.8 bp bust offset is fixed, so it takes a growing share of the cushion as leverage rises (36% of the zero-equity distance at 750x, 48% at 1000x). At 750x the bust sits ~4.5 bp above the 81,632 rejection high. At 850–1000x it sits only 1.2–3 bp above it, inside one minute's p90 range (3.6 bp). The 30-minute 1σ move (~2.2–2.8 bp/min × √30 ≈ 12–15 bp) is already larger than every available cushion. Under a driftless random walk the expected value at TP 15 bp gets worse with leverage: −0.32 margin at 750x versus −0.35 at 800x. I chose to chase gains through the target size rather than through leverage.
• Take profit (15 bp), Inference. 81,477 sits just ahead of the 81,452–81,460 6h/60m lows, where the 51 BTC bid cluster rests, so the TP fills before that support has to break. At 15 bp the trader keeps ~77% of the raw gain (86.8% on margin). At 6–8 bp the net return on margin at 750x is only 29–42%, the heavily taxed small-target case the brief warns against. A target past the 81,452 low (20–30 bp) would pay 120–185%, but it would need the range to break on a quiet Saturday, and the 4h empirical hit rate for 20 bp was only 5%.
• Regime "ranging", Inference. For 6h price has held 81,452–81,800 with 1m vol about 2 bp/min, roughly 40% of the 12h average. A downtrend dominates at the 24h scale.

3. Uncertainty: what the numbers say about the odds

• Random-walk baseline (computed). With an 8.53 bp bust and a 15 bp TP, P(TP first) = 8.53 / 23.53 = 36%. The breakeven win probability is 1 / (1 + 0.868) = 53.5%. Without a real directional edge, every allowed combination of leverage and TP has negative expected value under these mechanics. The bust offset and the win haircut make sure of that. The user's constraints require a trade, so the stated confidence (0.38) is only slightly above the no-edge baseline. The implied expected value is about 0.38 × 0.868 − 0.62 ≈ −0.29 of margin. That is an honest estimate, not a promise of profit.

• Empirical grid (computed, tools/analyze.mjs). Each 1m open in the last 12h was replayed for 30 minutes. A bar that touched both TP and bust was counted as a bust (conservative). Trades still open after 30 minutes were marked at close.

  Window	short 750/15: P(TP) / P(bust) / still open / EV	long 750/15	short 1000/15
  12h (n=720)	0.33 / 0.43 / 0.24 / −0.13	0.21 / 0.52 / 0.27 / −0.33	0.24 / 0.60 / 0.16 / −0.30
  4h (n=239)	0.20 / 0.21 / 0.59 / −0.00	0.05 / 0.35 / 0.60 / −0.31	0.17 / 0.42 / 0.41 / −0.17
  2h (n=119)	0.15 / 0.39 / 0.46 / −0.23	0.11 / 0.38 / 0.51 / −0.22	0.10 / 0.59 / 0.31 / −0.44

  These are backward-looking and overlapping, so the samples are not independent. In the 2h window long and short are about equal, which is why the directional call rests on structure (the failed spike, the downtrend, OI) more than on the grid. In every window, 750x beats 1000x.

• Time to resolve (computed). In the 4h grid, winning short 750/15 trades took ~17 minutes on average. Many trades resolve in neither direction within 30 minutes at the current vol, so this position may stay open past the brief's ideal 30-minute window.

4. Open questions and limits

• Not checked: liquidation heatmaps (no free primary source was used), and other venues' trade-level order flow beyond the one Binance/Coinbase snapshot. A web search for BTC news on 2026-10-17 found no event-specific report. The latest dated item was a BGeometrics daily report putting BTC at ~$81,900 on 2026-10-12 (bgeometrics.com). No catalyst was identified. US macro data are not scheduled on a Saturday, but I did not verify this against an economic calendar.
• Coinbase −17 bp vs Binance USDT spot: I did not check the USDT/USD rate, so I cannot say how much of this gap is a stablecoin premium rather than weak US spot demand. It got little weight in the decision.
• Book data: Hyperliquid returns only 20 levels per side. At full precision that covers about ±2.5 bp. Resting size can be pulled, and the near-book imbalance flipped from −0.43 to +0.40 within 3 minutes, so it is a weak signal.
• OI change: this is three point readings over ~3 minutes, not a history. Hyperliquid's public info API gives no OI time series.
• Latency: the entry is the 02:09:12Z mid. PaperTrade fills at the live mid when the order is placed, so if the mid has moved, the TP and bust prices shift by the same number of bp. If price is already above ~81,632 at entry, the thesis (a failed spike) is void.
• 1m OHLC cannot show the order of highs and lows inside a bar. The grid's same-bar rule is conservative for both sides.

5. Reproduce

node tools/fetch.mjs <dir>            # HL candles 1m/5m/15m/1h, metaAndAssetCtxs, l2Book, fundingHistory, recentTrades
node tools/fetch_extra.mjs <dir>      # coarser HL books, premium history, Binance/Coinbase quotes
node tools/analyze.mjs <dir>/snapshot.json <dir>/analysis.json
node tools/decide.mjs <dir>/snapshot.json short 750 15 81460 81632

Sources: Hyperliquid info API https://api.hyperliquid.xyz/info (request bodies in tools/fetch.mjs); Binance https://fapi.binance.com/fapi/v1/premiumIndex?symbol=BTCUSDT, https://fapi.binance.com/futures/data/openInterestHist?symbol=BTCUSDT&period=5m&limit=36, https://api.binance.com/api/v3/ticker/bookTicker?symbol=BTCUSDT; Coinbase https://api.exchange.coinbase.com/products/BTC-USD/ticker; BGeometrics daily report 2026-10-12 https://bgeometrics.com/reports/daily/2026-10-12.html. This report is a model-generated trading decision under the brief's forced-trade constraints, not financial advice, and no independent reviewer checked it.