The trading engine #

This is a research and paper-trading engine. It collects market data, screens a broad US stock and ETF universe, runs strategies in a simulator, and records enough evidence to ask a hard question: did a decision process add value after realistic costs, or did it merely get lucky?

“Paper” is the important word. Orders, fills, cash, positions, and costs are simulated. The engine has no broker connection, no broker credentials, and no authority over real money. A future broker paper account is described in an internal proposal, but it is explicitly not work for now.

The engine runs unattended on a Linux server. An LLM can make bounded research decisions through a model gateway, while deterministic code controls the data it may see, the candidate list, position size, risk limits, fills, accounting, halts, and evaluation. The model can express a view. It cannot invent a price, bypass a risk check, send a broker order, or turn a weak result into a pass.

What happens each day #

On a normal market day, the engine:

1. checks pending paper entries before the market opens;
2. observes intraday bars, headlines, and scheduled market snapshots without granting intraday order authority;
3. collects end-of-day prices and other facts after the session;
4. advances every active paper book using the same next-open fill rules;
5. asks the daily AI policies to score candidates after the deterministic data is ready; and
6. joins older decisions to outcomes that have now matured.

Weekly jobs verify the store, refresh liquidity, rerun walk-forward checks, and summarize the state of the research. The daily cycle gives the exact schedule in New York time.

What the engine is trying to prove #

A positive return is not enough. A policy is compared with a frozen control on the same dates and candidates. Costs are charged to the strategy. Attempts are counted. Results must survive data that was not used during development. A policy that fails its registered test is stopped rather than tuned after the fact.

The live profitability evidence loop scores a wider candidate set and runs three matched paper books: AI-ranked, rule-ranked, and a rule book with an AI veto. It has run since September 2026 and is still collecting evidence. Its scoring, books, and pass/fail gates have not changed since it started; later registration revisions only fixed the infrastructure around them.

The challenger lab, text edge, and evaluation science plan has built stronger evaluation, shadow challengers, filing and replay paths, portfolio construction, and execution measurement. Those parts remain inert until their own registration, rehearsal, and activation gates pass. The shared backtest core is complete and available for public, reusable studies.

No policy has yet beaten its frozen control prospectively. That is the honest current status. The system has working infrastructure and accumulating paper evidence; it does not have a proven edge.

What is running now #

As of 22 October 2026:

• Engine v2 is live on paper. Since 18 October 2026 the engine has been the single executor and ledger for every paper book and paper account. From 19 October each book pays a registered commission schedule, and its evaluation clock restarted on that date so old and new cost rules are never mixed in one result.
• 17 paper books are active. On 21 October eight books that showed no edge after costs were retired. Their history stays in the record; they simply stop trading.
• Three frozen forward records keep accumulating: a weekly SPY Monday experiment, a sector momentum rule, and a 12-1 stock momentum comparison. None has reached the date when its verdict is allowed.
• A senior-quant review is under way. This in-depth technical review of the whole pipeline fixed data-integrity and statistics defects, made the nightly jobs, simulator, and test suite faster while proving identical outputs, and repaired stored price history on 22 October. A storage reorganization from the same review is still being rehearsed and is not live.

Public engine, private research #

This repository contains public infrastructure: data handling, simulation, evaluation statistics, replay tools, runners, schemas, and example policies. It is intended to be inspectable and useful to anyone who wants to test their own ideas.

Private research strategies run in a separate private repository and reach the engine only through the shared backtest core and paper accounts. Their rules, prompts, parameter sets, registrations, trial counts, findings, and results are not published here. That private programme exists and is active, and none of its strategies has a proven edge either. Engine v2 already gives each private strategy version its own separately funded paper account, so it can be observed in real time without revealing its logic.

Where to go next #

Read Paper books and the league to understand what a simulated trade means, AI agents to see the model's narrow authority, and How research earns a verdict for the safeguards against overfitting. The glossary translates the internal plan codes and statistical terms.