Agent

Your own agent, Hermes, reasoning over a research engine that produces its numbers deterministically. It builds the case for a trade and hands it to you — it does not spend your money on its own.

What makes it different

Most AI trading tools are a language model with a market-data feed attached. Ask one whether a strategy works and it will tell you — fluently, confidently, and without having checked. Agent is built the other way round.

1. 1The research engine computes, first. Backtests, benchmarks, drawdowns, walk-forward tests and confidence intervals are arithmetic — reproducible, and identical every time you run them. No model is involved in producing a single one of those numbers.
2. 2A deterministic verdict decides whether to trust them. Before anything reaches the bot, the result is checked for the failures that make a good-looking backtest meaningless: too few trades, too short a window, a return that dies at realistic costs, a parameter tuned to one lucky period.
3. 3Hermes reasons over the evidence — the findings and the figures, not the raw feed. It explains what the numbers mean and what to do about them. It is never asked to produce a number, and it cannot overrule one.
4. 4You decide. The proposal arrives with the evidence behind it, including the parts that argue against it.

That division is the edge. A model that both produces the numbers and judges them will always find its own work convincing. Separating the two means the bot can tell you a strategy it likes is not trustworthy — and it regularly does.

The trading desk

A play on a named stock sits a simulated firm: four analysts (fundamental, news, sentiment, technical), a bull/bear judge, a trader who leans short / mid / long, and three risk seats. The technical seat runs RSI, MACD, and bands. Sentiment reads the street and the ApeWisdom mention board — attention, not a buy. The trader labels the regime — bull, bear, or chop — before it leans. Cheap plus a bear tape is a falling knife: the judgment is wait, not a long ticket. Deep research sits in a tab on the right: what they sell, what lags, what is next, what failed, then cash, competition, and a verdict — not a recitation of the multiple. The trader can highlight a ticket. It cannot invent a share count, and it cannot send the order. You still Approve.

Each sit is stored as a replayable day: market tape, what already filled, and the last lean. Paper fills score a PRIDE strip (total return, Sharpe, max drawdown, win rate, Calmar). Empty books stay empty. New specialists (macro, prediction markets) register on the firm graph. They do not grow the chat prompt.

Autonomy levels

The level sets how much of the thinking the bot does for you. It does not set whether a human approves the capital, because that answer is always yes.

Before a strategy can go live

Going live is gated in the database, not by a checkbox in the interface. A strategy cannot be promoted until all four are true — the rule is enforced server-side, so it holds however the request arrives.

• A completed backtest that actually traded, run in the last 90 days.
• A stress test over at least 250 trading days. A 95% value-at-risk drawn from a shorter sample is just the worst day in it.
• 14 days of paper trading. Long enough to meet a market that disagrees with you.
• 20 filled paper trades. Fills, not signals — an order that never filled proves nothing.

Risk controls

Every bot carries these regardless of level:

• Maximum position size — caps exposure on any single trade.
• Daily loss limit — pauses the bot once hit, rather than trading through a bad day.
• Volatility guard — reduces size when the market gets rougher than the strategy was tested in.
• Correlation check — stops five positions that are quietly the same bet.

Deploying

alphonce strategy deploy \
  --strategy ./my-strategy \
  --broker t212 \
  --mode paper \
  --allocation 25

Monitoring

Follow a bot's reasoning and actions as they happen. Every decision is logged with the evidence it rested on, so a proposal can be reconstructed later rather than taken on trust.

alphonce bot logs --id bot_abc123 --follow