Every trade is linked to the AI decision behind it, the Alpaca order id, the broker's fill time, and realized P&L. We deliberately show losers, cancelled and dead orders, model errors, and decisions the AI made but never acted on. Each day's record is hashed and anchored to the Bitcoin blockchain (OpenTimestamps) so it can't be backdated. Generated from the database every publish — not hand-authored.
The day's audit record is anchored to Bitcoin via OpenTimestamps (we pay nothing; you need no wallet). Confirm it existed and wasn't edited:
{`pip install opentimestamps-client
curl -O /data/audit/${today}.json
curl -O /data/audit/${today}.json.ots
ots verify ${today}.json.ots`}
OpenTimestamps proves existence-by-time, not correctness, and is forward-only from launch. A "pending" result just means Bitcoin confirmation (a few hours) hasn't landed yet.
| Decision time | Signal | Ticker | Side | Qty | Price | Status | P&L | Alpaca order |
|---|---|---|---|---|---|---|---|---|
| {r.decision_time ? r.decision_time.replace("T", " ").slice(0, 16) : "—"} | {r.signal || "—"} | {r.ticker} | {r.side} | {r.qty || "—"} | {r.price ? "$" + r.price : "—"} | {r.status} | = 0 ? "var(--pos, #2e7d32)" : "var(--neg, #b3462d)" }}>{r.realized_pnl != null ? pnl(r.realized_pnl) : ""} | {r.order_id ? r.order_id.slice(0, 8) + "…" : "—"} |
Showing the most recent {Math.min((rec.rows || []).length, 60)} of {s.trades_total ?? "—"} trades. Decision link: FK = hard foreign-key to the decision row; matched = best-effort by ticker+date; backfill = reconciliation row with no stored decision (older/manual/orphan fills). Full set at /api/reconciliation.
Runs on a separate machine — real-money keys never touch this site. Data appears once it publishes its first snapshot.
A real-money account that mirrors the LLM fund in real time, scaled to its own capital — each LLM trade replicated as the same % of NAV it was for the LLM fund. It copies trades forward from its own start (no back-fill), runs no debates or discipline of its own, and mirrors only the LLM fund. Returns are measured against total deposited capital, and the S&P benchmark invests each deposit on the day it landed.
A separate $5,000 Alpaca paper account running 50% SMH / 40% QQQ / 10% WTAI with a 5pp drift rebalance gate. The 10% WTAI sleeve (WisdomTree Artificial Intelligence & Innovation Fund) broadens AI exposure beyond pure semis/Nasdaq into the wider AI software + application layer. Evolved from 60/30/10 QQQM/SMH/XLK → 50/50 QQQ/SMH → this mix. No LLM, no signal feeds, no debates — pure allocation discipline. The LLM portfolio's alpha (or lack thereof) measured against this one is the real edge test.
| Ticker | Target | Actual | Drift |
|---|---|---|---|
| {t} | {tgt.toFixed(0)}% | {act.toFixed(2)}% | 5 ? "neg" : ""}>{drift >= 0 ? "+" : ""}{drift.toFixed(2)}pp |
{strategy?.rebalance_rule || "Rebalance when any leg drifts >5pp"} · daemon fires every 30 min · trades only during market hours when drift exceeds threshold
| Date | NAV | Δ from start | QQQ close | Drift |
|---|---|---|---|---|
| {h.date} | {fmtUSD(h.nav, { decimals: 2 })} | = 0 ? "pos" : "neg"}>{ret >= 0 ? "+" : ""}{(ret * 100).toFixed(2)}% | {h.benchmark_qqq_price ? `$${h.benchmark_qqq_price.toFixed(2)}` : "—"} | {(h.drift_max_pp || 0).toFixed(2)}pp |
No history yet — daemon writes a snapshot every 30 min during market hours.
)}The LLM-driven portfolio (other tab) makes daily allocation decisions using a multi-agent debate over its signal layer. This sister fund makes zero discretionary decisions — it buys QQQM/SMH/XLK at the listed weights and only rebalances when drift exceeds 5pp. If the LLM fund's NAV doesn't materially beat this one over time, the debate framework isn't adding alpha. That's a real question this site answers in public, day by day.
A separate Alpaca paper account that fronts-runs the order flow of @aleabitoreddit (Serenity) — the photonics / AI-supply-chain analyst whose calls (NBIS, AXTI, AAOI, SIVE) have run 3–15× off the bottom. The pipeline scrapes her live timeline, parses every $TICKER, and scores conviction from mention frequency, recency, and tone.
Each name is then run through a Serenity-tuned TradingAgents debate — a multi-agent bull/bear book that ingests her conviction context alongside price action, AV news sentiment, and StockTwits retail flow. Position sizing is a function of both her conviction and the debate verdict: OVERWEIGHT/BUY → full size, high-conviction starter → 5% NAV tranche, thesis-tail → tiny 0.5% lotto on names the tape hasn't confirmed yet. We risk-filter her, not replicate her: chase the structurally durable, fade the blow-up-prone.
Exits are symmetric — hard −12% stop, 50% trim into binary earnings prints, 25% trim when she turns bearish on a held name, and a daily re-debate sweep that cuts positions the agents flip to SELL. Non-US names she pushes (SIVE, LPK, SOI, IQE) get debated for the analyst read but aren't tradeable on Alpaca. Fully autonomous: scrape → debate → size → fill, fired on a schedule and on every laptop wake.
No positions yet — initial TradingAgents debate may still be running or hasn't been triggered yet.
) : (| Ticker | Qty | Avg cost | Last | MV | PnL |
|---|---|---|---|---|---|
| {p.ticker} | {(p.qty || 0).toFixed(4)} | ${(p.avg_cost || 0).toFixed(2)} | ${(p.last_price || 0).toFixed(2)} | {fmtUSD(p.market_value || 0, { decimals: 0 })} | = 0 ? "pos" : "neg"}> {(p.unrealized_pnl_pct || 0) >= 0 ? "+" : ""}{((p.unrealized_pnl_pct || 0) * 100).toFixed(2)}% |
Debate hasn't run yet. Will populate after first daemon fire.
) : (| Ticker | Signal | Date | Rationale (click to expand) |
|---|---|---|---|
| {ticker} | {sig} | {(rec.date || "").slice(0, 10)} | {rec.error ? error: {rec.error} : (rec.decision_excerpt || "—")} |
Click any row to expand the full debate excerpt. Re-debate fires every Monday on all 11 candidates.
The universe extends beyond pure-CPO names to include Leopold Aschenbrenner's Situational Awareness LP top-5 holdings — BE (Bloom Energy, his #2), CORZ (Core Scientific), CRWV (CoreWeave) — under the thesis that AI scale is power-constrained, not chip-constrained. COHR and LITE are also in his fund. The TradingAgents debate ultimately decides which names earn an allocation; the universe is wider than pure CPO to let bull/bear arguments span the broader AI-infrastructure thesis.
Top mentions (14d window), a {pulse.window_days_heatmap}-day mention heatmap, and the raw tweet stream — each ticker chip colored by our most recent TradingAgents debate verdict. Click any ticker to filter the stream. {stats.tweets_total} tweets · {stats.debates_total} debates · {stats.tickers_tracked} tickers tracked
{selected && (| Ticker | {days.map((d) => { // Show every 5th day label const dayNum = parseInt(d.slice(-2), 10); const showLabel = dayNum % 5 === 0 || d === days[days.length - 1]; return ({showLabel ? d.slice(5) : ""} | ); })}
|---|---|
| setSelected(isSelected ? null : row.ticker)} style={{ position: "sticky", left: 0, background: labelBg, padding: "4px 8px 4px 8px", fontWeight: 600, fontSize: 11, cursor: "pointer", color: row.non_us ? "var(--ink-3)" : (isBuy ? "var(--pos, #2da465)" : "inherit"), textDecoration: isSelected ? "underline" : "none", borderLeft: isBuy ? "3px solid var(--pos, #2da465)" : "3px solid transparent", }}> ${row.ticker} | {row.cells.map((cell) => (0 ? `${row.ticker} on ${cell.day}: ${cell.count} mention${cell.count > 1 ? "s" : ""}${cell.debate_signal ? ` · we debated → ${cell.debate_signal}` : ""}` : ""} style={{ padding: 0, minWidth: 16, height: 16, background: cellColor(cell.count), border: "1px solid rgba(0,0,0,0.15)", position: "relative", }}> {cell.debate_signal && ( • )} | ))}
Cell intensity = mention count that day. Dot inside cell = we ran a TradingAgents debate on that day (color = signal: ● BUY/OW · ● HOLD · ● SELL/UW).
{t.text}
{t.tickers.length > 0 && (Daily scrape of @aleabitoreddit (Serenity — the original WSB / AI-Semi supply chain analyst) plus accounts she's endorsed. Tickers mentioned across the feed get filtered to US-tradeable names, deduped against current holdings, and the top {2} new candidates run through a TradingAgents debate. Only BUY/OVERWEIGHT confirmed names get bought at 10% NAV. The rest is research, not signal.
| Ticker | Mentions | TA signal | Action | Notional |
|---|---|---|---|---|
| {a.ticker} | {a.mention_count}× | {sig || "—"} | {a.action || "—"} | {a.notional ? `$${a.notional.toFixed(0)}` : "—"} |
Deferred to tomorrow (over daily cap): {run.candidates_deferred.join(", ")}
)}No tweets pulled yet — first daemon run hasn't completed.
) : (New @-mentions surfaced in today's feed (potential additions to watch list — review manually):{" "} {run.new_mentioned_handles.slice(0, 15).map((h) => ( @{h} ))}
)}{fund.name} is a transparent paper-trading experiment built on{" "} TradingAgents{" "} — an academic multi-agent LLM research framework (arXiv:2412.20138, Tauric Research). The reasoning model is GPT-5.5 via OpenAI's Codex CLI. Every trade, every thesis, every source — published live. Not real money, not advice.
Transparency note: the {metrics.losing_trades} losing trades shown above are a feature, not a bug. A portfolio manager — human or AI — that never loses is either lying or not trying.
A deliberately simple process, because simple processes are the ones you can actually audit.
{text}
Click any position to read the thesis.
Trades and unfiltered thinking — in the same feed, same time, same tone. Published automatically, not curated after the fact.