An agent desk that trades its own paper book — and shows its work.
Scheduled signals feed two LLM agents — one proposes, one can veto. Survivors pass a hard risk guard before any order reaches the broker, and every decision lands in an auditable log you can replay.
Live from the research lab
Not a stock photo — the actual walk-forward equity curve of the lab's lead momentum strategy, marked nightly against SPY since inception.
Real output from the research lab's nightly mark job — the same strategy_lab_nav table the operator dashboard reads. Walk-forward, point-in-time data, transaction costs included. Past performance of a paper book; not investment advice.
What's inside
Signal engine
Momentum, sentiment and volatility composites recomputed on schedule over point-in-time data. Two timestamps on every row.
Proposer + reviewer
One model drafts the trade; a second model with a different prompt reviews it. A reviewer veto is final.
Risk guard
Position caps, a daily-loss breaker, and exchange-hours gates run after the agents and before the broker.
Audit trail
Signals, proposals, vetoes, orders and operator actions land in one append-only log the dashboard can replay.
Two models argue about every trade
The proposer never grades its own homework: a second, independently prompted model reviews each thesis — and a veto is final. Most ideas die here. That's the point.
NVDA crossed the composite entry threshold. Proposing a long: ~2% of equity, 15-session time stop, −10% hard stop.
Earnings print in three sessions — the blackout rule forbids new entries into the window. And the momentum term is mostly one gap day, not a trend. Veto.
Illustrative exchange — the shape of a real veto from the paper book.
How a trade happens
Six stages, in this order, every time. Nothing skips the queue — a human can watch (or veto) any of it from the dashboard.
Ingest
Bars, news, social and earnings pulled on schedule — every row stamped with when it happened and when we saw it.
Score
Momentum, sentiment and volatility composites recompute across the universe each pulse.
Draft
An LLM proposer turns qualifying signals into a structured trade thesis with sizing.
Challenge
A second, independent model tries to kill the idea. A veto ends it here.
Check
Position caps, daily-loss breaker, exchange-hours and liquidity gates — no exceptions.
Trade & log
Surviving orders reach the paper broker; every hop lands in the audit trail.
Research lanes on the paper book
Independent strategy lanes run side by side, each firewalled from the live path until the evidence says otherwise.
ScoreMom momentum
Cross-sectional momentum over a 31-name diversified universe, walk-forward validated.
News + social sentiment
LLM-classified headlines and retail chatter, point-in-time joined to the book.
Options regime router
Per-regime structure selection — condors, verticals — forward-papered against an EV baseline.
Earnings vol crush
IV-crush entries calibrated on historical chains, tracked through each earnings date.
Catalyst rank
Event-driven hot-pool ranking with an ML scorer and an LLM judge.
ETF strategies
Sector-rotation and defensive overlays benchmarked against SPY buy-and-hold.
A-share lab
Shanghai/Shenzhen factor panel with its own PIT universe — research-only, firewalled.
LLM strategy loop
Models propose strategy variants; a harness backtests, gates and paper-runs the survivors.
One dashboard over every decision
The operator console shows the book as the agents see it — and the full story of how each position got there.
- Live signals, proposals and reviewer verdicts as they stream in
- Order lifecycle from thesis to fill, with every veto preserved
- Risk state at a glance: caps, breakers, exposure, session gates
- Strategy Lab curves comparing every lane against SPY
Operating principles
Point-in-time or it didn't happen
Two timestamps on every datum. Backtests only see what was knowable at decision time — no lookahead, ever.
Paper first, always
Every strategy earns its place on a paper book with real market data before real capital is even a conversation.
Two models, one veto
The proposer never grades its own homework. An independent reviewer model can kill any trade, and often does.
Everything is auditable
Signals, proposals, vetoes, orders, operator actions — one append-only trail, replayable from the dashboard.
See the book. Question the agents.
Sign in to open the live paper book — dashboard access is approved by an admin.