Live Now — lablab hackathon account (paper, $100k)
Risk Guardian's 24h hold cap hit at 12:07 PM PDT on Sep 3. The position closed by 12:22 PM.
Portfolio Value
—
Unrealized P&L
—
Cash
—
Trade Result
+$172.83
Agent Control — Autonomous Runner
CHECKING…
Gates new entries only — stop-loss / hold-cap enforcement always runs.
Applies within one scan cycle (~5 min).
Applies within one scan cycle (~5 min).
| Symbol | Side | Qty | Avg Entry | Current | Unrealized P&L |
|---|---|---|---|---|---|
| No open positions. The spread closed on Sep 3 at 12:22 PM PDT for a realized +$172.83 — see Recent Orders below. | |||||
This account has no open positions right now. Values above are live from the Alpaca paper API, polled every 60 seconds; the trade card and log below show exactly what happened while it was open.
The Trade — SPY debit call spread, placed and closed by the agents
Net Debit Paid
$481
Risk Guardian Stop
$580 / 24h
Realized P&L
+$172.83
Status
CLOSED — SETTLED
Closed Sep 3, 2026 · 12:22 PM PDT
Entry Order
c49e917f-5ad5-404a-bdfe-5af3524dcb0a
Long Leg Entry → Exit
SPY260916C00765000: $7.37 → $12.33
Short Leg Entry → Exit
SPY260916C00775000: $2.56 → $5.79
Spread Width
$10 wide
Reward:Risk (at entry)
~1.08:1
Held
Sep 2, 12:07 PM → Sep 3, 12:22 PM PDT (~24.3h)
Risk Guardian's 24-hour hold-time cap hit at 12:07 PM PDT on Sep 3. A Codespace outage delayed the automated close by about 15 minutes. Both legs closed by 12:22 PM for a realized +$172.83. Full sequence in the Agent Decision Log below.
| Submitted | Symbol | Side | Qty | Status | Filled Avg |
|---|---|---|---|---|---|
| Sep 3, 12:22 PM PDT | SPY260916C00765000 | sell | 1 | filled | $12.33 |
| Sep 3, 12:18 PM PDT | SPY260916C00775000 | buy | 1 | filled | $5.79 |
| Sep 2, 12:07 PM PDT | SPY260916C00765000 / SPY260916C00775000 | buy / sell | 1 | filled | $4.81 net |
Agent Reasoning — actual pipeline transcripts
> scout --signal SPY-LIVE-FIRST-REAL
thesis: “confirmed market structure shift and strong displacement suggests upward pressure toward the measured move target”
confidence: 0.78 · structure: debit_call_spread
thesis: “confirmed market structure shift and strong displacement suggests upward pressure toward the measured move target”
confidence: 0.78 · structure: debit_call_spread
> risk_guardian --review
[VETO · Sep 1] — daily P&L −$50 plus the backtest tail stat that 8.5% of winning trades still lose more than $50. Scout’s $300 loss ceiling got rejected. Executor: NO_ORDER, zero wasted calls.
[APPROVE_MODIFIED · Sep 2] — “$480/contract… stop must cover the full debit plus tail risk… $580, 24h hold”. Executor placed the order, both legs filled in ~100ms.
[VETO · Sep 1] — daily P&L −$50 plus the backtest tail stat that 8.5% of winning trades still lose more than $50. Scout’s $300 loss ceiling got rejected. Executor: NO_ORDER, zero wasted calls.
[APPROVE_MODIFIED · Sep 2] — “$480/contract… stop must cover the full debit plus tail risk… $580, 24h hold”. Executor placed the order, both legs filled in ~100ms.
> position_monitor --watch
enforced max_loss=$580 · max_hold=24h · cap hit Sep 3, 12:07 PM PDT → both legs closed by 12:22 PM → +$172.83 realized
enforced max_loss=$580 · max_hold=24h · cap hit Sep 3, 12:07 PM PDT → both legs closed by 12:22 PM → +$172.83 realized
| Timestamp | Signal | Scout | Risk Guardian | Executor |
|---|---|---|---|---|
| Sep 1, ~9:44 PM PDT | Bullish SPY FVG, spot $762.15, target $777 | debit_call_spread, confidence 0.78 | VETO $300 loss vs. -$50 daily P&L, per the backtest tail-risk stat | NO_ORDER (correct no-op) |
| Sep 2, ~12:01 PM PDT | Bullish SPY FVG, spot $765.18, target $773.18 | debit_call_spread, confidence 0.78 | APPROVE_MODIFIED $580 cap set: the spread costs $480/contract, and the stop has to cover the full debit plus a tail-risk buffer | Order built: long 765C / short 775C, 1 contract, dry_run=true |
| Sep 2, 12:07:12 PM PDT | Same signal, live submission | — | Same as above, real run | FILLED Real Alpaca order c49e917f…, both legs filled simultaneously (~100ms) |
| Sep 3, 12:07–12:22 PM PDT | 24h hold-time cap reached | — | Cap enforced by Position Monitor (deterministic, no LLM) | CLOSED Short leg bought back @ $5.79, long leg sold @ $12.33. Net +$172.83. |
How It Works — deterministic math, AI judgment, nothing blurred
Deterministic core
FVG Engine
ATR-based gap filter, MSS + displacement confirmation. Fires the raw directional signal.
→
LLM Agent · Groq
Scout
Contextualizes the FVG signal into a structured thesis and confidence score (openai/gpt-oss-120b).
→
LLM + Rules Agent
Risk Guardian
Hard limits in plain code first, fail-closed, then Groq reasoning. Sets the real stop-loss and hold-time cap, can veto outright.
→
Deterministic script
Executor
Alpaca CLI. Target-aware debit-spread strike selection, multi-leg order placement. No LLM in this step.
→
Deterministic script
Position Monitor
Polls the open position against Risk Guardian's max-loss and max-hold-time, and closes it when either trips.
Built the Hard Way — real bugs found live, fixed, committed
Silent Misroute Bug
str(Enum) would have flipped every bullish signal into a put spread, silently. Caught by diffing the real schema before the first live run, not after.
Cost-Blind Veto Bug
Risk Guardian kept setting a $50 stop no real spread could satisfy. It only surfaced once we ran it against live Groq; the fix went straight into the prompt.
Zero-Collision Filter
43.7% of backtest trades priced entry and exit off the same bar, faking $0 P&L. We caught it by distrusting a win rate that looked too clean, not by luck.
Enforcement State Loss
A Codespace rebuild wiped the Position Monitor's gitignored database, so it silently reported zero open positions while the spread was live. Caught by cross-checking the account directly and restoring the record.
Backtest Evidence — 2.5 years, real Alpaca options prices
Cumulative Backtest P&L — 1,569 clean trades
Mar 2024 – Aug 2026 · SPY / QQQ / XLV / XLF / IWM. Not a smooth rocket. The strategy dug a -$1,957 hole through 2024, peaked at +$2,090, and survived a -$2,179 peak-to-trough drawdown before finishing +$1,295. We show the drawdowns because that's the honest curve.
48.9% Win Rate
768 profitable / 801 losing trades across 1,569 clean trades. Avg win $42.33, avg loss $38.97: close to a coin flip with a small edge, which is what a real, non-curve-fit options strategy looks like.
Why the Risk Guardian Exists
19.4% of winning underlying signals still lost more than $20 on the options leg (8.5% lost more than $50). Directional accuracy alone doesn't protect the trade, so the Risk Guardian sets its own loss and hold-time stops instead of trusting the FVG signal.
Mean P&L per trade — symbol × direction
Clean dataset, real historical Alpaca options prices (intraday hourly bars, target-aware strikes). QQQ puts are the strongest replicated signal; the live strategy stays symmetric across all directions and symbols.
SPY, QQQ, XLV, XLF, IWM · Mar 2024 – Aug 2026 · zero-collision-filtered. Same-bar entry/exit pricing artifacts removed. That's a measurement fix, not a tuning change. Total P&L +$1,295 across 1,569 priceable trades.