A rules-first trading engine with a private terminal.

From candle to order

A deliberately simple chain. Each stage can reject a trade, and none can bypass the next. The Gate has the final say.

1. Collect

   4H and 1H candles, 24h volume, spread, open interest and funding for every liquid perpetual.

2. Score

   4H regime (EMA 50/200, ADX, structure) and derivatives flow, computed in code.

3. Adjust

   An AI analyst writes the 4-hour playbook: top 3 to 5 coins, direction and key zones. Code bounds its answer.

4. Gate

   Every 5 minutes the shortlist is scanned. Spread, funding, exposure caps and the -3% daily circuit breaker can each say no.

   Can veto

5. Execute

   Risk-based size, hard stop, TP1 at 1.5R and TP2 at 2.5R. The stop moves to break-even at TP1.

The two-layer model

Regime, levels, scores and triggers are computed in code, so you can reproduce and test them without the AI. The animations are illustrations on generated data.

Bounded AI: a language model writes the 4-hour playbook (its regime and direction must match what code computed, and its conviction stays within 10 points of the computed score) and judges triggered setups (it may veto or refine, never relax a rule). The war-room agents and the news radar can only add caution: a news veto, a risk-agent veto or a weak vote blocks a trade, and nothing can relax a hard rule. They cannot place orders, and they receive market data only, never keys or balances.

1. Risk-based sizingSize comes from a fixed percentage of equity and the distance to the stop.
2. Margin capSize is capped by available margin and leverage, so an order cannot over-commit the account.
3. Exposure limitsA maximum number of open positions and a maximum in the same direction.
4. Protected entriesEvery position gets a reduce-only stop and two targets. If the stop is rejected, the engine closes the position.
5. Kill switchOne action halts automation and stays engaged across restarts until it is deliberately released.
6. Unknown outcomes are quarantinedIf an order result is uncertain, no second order is sent until you verify.

Every decision is on record

Each playbook, trigger and judge verdict is stored with its reasoning and shown in the terminal.

1. 1. PlaybookEvery 4 hours the engine publishes its shortlist, regime and key zones.
2. 2. TriggerEvery 5 minutes it shows how close each shortlisted coin is to its setup.
3. 3. VerdictA six-agent war room debates each triggered setup; code re-counts the votes and logs EXECUTE or SKIP with a reason.
4. 4. GuardianAt 50% of the way to the final target the engine checks momentum and may book half and lock the stop.

Paper first. Live only on purpose.

Switching to real orders is a server-side decision, not a button in the browser.

Every trade is a dot

An illustration of how closed trades build up over time. New dots appear, then join the structure.

A terminal built for one operator

Dashboard data comes from the running engine, not from sample figures. If the engine or exchange is unreachable, the terminal says so instead of showing old numbers.

• Portfolio

  Equity, balance, available margin, realized and unrealized P&L.

• Positions and orders

  Open positions with stop and target, and closed-trade history with fees.

• Health

  Exchange connection, data freshness, uptime, rejected orders and risk-limit breaches.

• Controls

  Start, stop and kill switch, with a typed confirmation for anything destructive.

• Audit trail

  A hash-chained log with an integrity check.

• Analytics

  Win rate, profit factor, expectancy, drawdown and Sharpe, from closed trades with fees.

Three factors. One administrator.

1. 1

   Authorized e-mail

   Only one address can sign in. There are no public accounts.

2. 2

   Admin code

   Checked on the server against a hash that never reaches the browser.

3. 3

   One-time code

   Sent by e-mail. It expires in minutes, works once and allows a few attempts.

4. 4

   The terminal

   The browser never talks to the engine. Requests to it are signed and replay-protected.