Agent equity
$999.98Buy-and-hold BTC+ETH
$1,000.72Start
$1,000
Snapshot taken 2026-10-13 23:23 EDT when this page was built. It is a static snapshot, not a live feed; the agent runs every 30 minutes in its own sandbox and results are refreshed on request. Early numbers are fees and noise only: 0 closed trades, so no win rate or calibration yet.
| Time | Pair | Side | Price | USD | Reason |
|---|---|---|---|---|---|
| 10-13 23:00 | BTC-USD | buy | $83,581.71 | $50 | reflex |
| 10-13 23:05 | ETH-USD | buy | $2,510.66 | $50 | reflex |
BRAIN (slow, offline: research, strategy ideas, code review, nightly review)
| proposes parameters; a human approves
REFLEX (every run: state in -> typed decision out) [today: momentum rule, NOT an AI model]
| {action: buy|sell|hold, confidence: 0..1}
RISK LAYER (plain code, owns every limit, can veto the reflex)
|
PAPER BROKER (simulated fills at live Coinbase prices, 10 bps fee assumed)
The reflex is pluggable. The pasted prompt names Jev (TypeSafe AI) for this role; I have no Jev key and have not tested it, so today it is a simple 1-hour momentum rule with an uncalibrated confidence. Code owns every threshold, size and side effect; the reflex only judges.
{"pair":"BTC-USD","action":"buy|sell|hold","confidence":0.0-1.0,"why":"reflex|stop|take_profit"}
Each run pulls live public data (Coinbase ticker and 5-minute candles for BTC-USD and ETH-USD), builds state (price, last 12 closes, positions, entries, equity, day-start equity), then asks the reflex. Not yet used: funding, open interest, dominance, stablecoin liquidity, macro. Those are in the pasted research brief and are listed as next steps, not built.
| Rule | Value | Source |
|---|---|---|
| Max per position / total | $50 / $100 | my choice |
| Risk budget per trade | $0.50, size shrinks as recent volatility rises | responds to paper finding that sizing was volatility-blind |
| Bracket exits | 1% stop, 1.5% take-profit | levels are my choice; the paper reports a mechanical bracket recovered +39 bps per position |
| Daily loss halt | $5, then no new entries until next day | my choice |
| Min confidence to enter | 0.6 | my choice |
| Leverage | none | my choice |
Reports equity vs buy-and-hold, trade count, win rate vs average stated confidence (calibration), halts. It refuses to read meaning into fewer than 50 closed trades.
From the pasted brief: judge the market regime (BTC/ETH trend, dominance, stablecoin liquidity, funding, open interest, macro, narrative), then shortlist 5-10 asymmetric setups with supply/unlock, revenue, TVL, holder concentration and confirmed-vs-speculative catalysts. Status: not built. Honest caution: a slow research story is not an edge by itself, and small-cap "asymmetric" tokens are where retail loses most.
Same three layers and the same code-owned risk layer for every asset class. Only the data adapter and broker simulator differ.
| Asset class | Status | Data / fills | Extra rules |
|---|---|---|---|
| Crypto (BTC, ETH) | Running (paper) | Coinbase public API | 24/7; fee 10 bps assumed |
| Stocks and ETFs | Planned | Source not chosen. A free key-less quote feed I tried (Stooq) did not work; needs a source, possibly the options desk's feed | Market hours only; no leverage or shorting; per-name and total caps; position limit as % of paper equity |
| Options | Run separately by the existing options desk | Per that desk | Executable single legs reported separately from hypothetical spreads (the desk's own audit convention); defined-risk only; max premium at risk per trade; no naked short options |
Unifying with the options desk (plan): keep its bot untouched. Share only a read-only daily export (equity, open positions, closed trades, label executable vs hypothetical). The lab imports it, shows one combined table per asset class with each one's own benchmark (crypto: BTC+ETH buy-and-hold; stocks/ETFs: SPY buy-and-hold; options: cash), and applies one portfolio-level halt on combined paper drawdown. No shared keys or code.
Real trading (OFF, requires Ade's fresh approval): needs a broker or exchange choice, a hard dollar cap, and trade-only API keys (no withdrawal) entered through a vault link, never in chat. Gate before any of it: weeks of paper results beating the benchmark after fees, 50+ closed trades per strategy, and the paper research's warning that LLM agents showed no edge. Order placement code would stay separate from any AI component.
Figures supplied by the separate options-desk bot, as of Oct 13, 11:13 PM ET. Not results: two days of data, too short to mean anything. All simulated; never a promise of returns. Data: Yahoo Finance via yfinance (free, about 15-minute delayed quotes, options chains); crypto spot from Coinbase, Kraken and Gemini public APIs. No keys.
| Sleeve | Start | Value now | Benchmark / note |
|---|---|---|---|
| Trend sleeve (stocks) | $10,000 paper | $10,008.83 | Buy-and-hold $10,016.34 |
| Options sleeve | $100 paper (mirrors a $100 budget) | $86.50 (cash $30.50 + positions $56.00) | Cash = $100 |
| Crypto arbitrage scanner | $1,000 paper float | 0 trades | All spreads thinner than about 0.66% combined fees |
Total paper capital across all bots: $11,100.
Trend sleeve positions (all EXECUTABLE plain buys, Oct 12): AAPL 7.349266 sh @ $340.17; MSFT 4.617573 @ $541.41; NVDA 11.029250 @ $226.67; AMZN 9.555726 @ $261.10. No closed trades. Rules: max 6 positions, volatility-scaled sizing (about $2.5k), 2.5x ATR trailing stop, SPY 200-day regime gate, entry only if price above 50-day average, 20-day momentum positive and RSI below 70; fee $0.005/share.
Options positions:
Options rules: max 2 positions, max 50% of the sleeve risked per trade, $1.15 per side commission (Oct 12 entries were booked at an older $1.50/$3.00 assumption), entries at ask and exits at bid, delta about 0.30, 21-45 days to expiry, open interest at least 100, bid-ask spread at most 20% of mid, implied volatility at most 120%. The $100 sleeve is presumably CAD-ish; currency unconfirmed.