Assistants

When a backtested strategy is ready for the next stage, the Paper Testing Assistant takes it directly. Once the algorithm is deployed, this assistant watches the strategy as it runs and asks three sharp questions in order, because each one builds on the last and a strategy that fails an early check is not worth grading on a later one.

The first question is whether the algorithm is actually doing what its code says. Orders should fire on the expected symbols, at the expected times, in the expected direction and size, with no surprise fills and no missing trades. The second question is whether the live behavior lines up with a backtest run over the same dates as the paper trading period. Small differences from fill slippage or timing jitter are expected, but missing trades, opposite directions, or entirely different symbols are structural problems that need explanation. The third question is whether out-of-sample performance still fits inside the statistical envelope the backtest established — a paper drawdown that doubles the backtest’s worst case is a strong signal the strategy is overfit, not a streak of bad luck.

When a check fails, the Paper Testing Assistant does not throw up its hands. It re-reads the project source against the observed behavior, looking for state-dependent bugs, timezone issues, universe drift, or data differences between the backtest and live environments. If the cause is not obvious from the code alone, it adds targeted instrumentation — tags on orders to capture the decision variables at the moment of submission, or logs on each time step when the issue is that no orders are firing in the first place. It then redeploys, watches again, and iterates up to three total attempts before escalating rather than spinning indefinitely.

When escalation is the right call, the Paper Testing Assistant stops the algorithm and tells you exactly which goal failed, what it observed versus what it expected, what it tried during diagnosis, and its best hypothesis for the root cause. It is careful to separate what it knows from what it suspects, so you are never left guessing which parts of the report are facts and which are interpretations.

A strategy that has either earned the confidence to be considered for live capital or has been stopped early with a clear explanation of why. The Paper Testing Assistant treats every stop and start as a real cost on the clock, batches instrumentation changes to avoid unnecessary cycles, and defaults to asking rather than letting a misbehaving algorithm keep running. The outcome is the same kind of honest verdict the research stages produce, only now backed by live market behavior instead of historical simulation.

The Paper Testing Assistant has access to the following tools:

• create_compile
• read_open_project
• update_project
• read_project_nodes
• create_backtest
• read_backtest
• list_backtest
• update_backtest
• delete_backtest
• search_backtest_logs
• create_live_algorithm
• read_live_algorithm
• liquidate_live_algorithm
• datasets
• user_input
• environment_library_support
• object_store_get
• object_store_set