XTester 0.0.70: AI Research Lab Stops Dead Branches

XTester 0.0.70 is now available. The main change is a more reliable AI Research Lab. The lab is better at recognizing when a strategy branch has stopped producing useful evidence, so it should no longer spend the next runs extending the same dead end.

The release is available for Windows as an installer, a portable build, and an automatic update.

Download XTester 0.0.70 for Windows

What exposed the problem

One live research task had accumulated 53 strategy versions. Versions v50 through v53 all produced the same degenerate result: zero fitness, no signals, and no trades.

Version v29 was still present in the history. It had completed 223 trades and produced a positive result, although it was not yet acceptable because its drawdown exceeded the research limit. The correct next step was to return to the working version and reduce its risk—not to keep generating descendants of a branch that could not enter the market at all.

The investigation showed that the trading idea was not the only issue. Generated code could contain a guard that stopped processing precisely while the strategy was flat. The version compiled and looked plausible, but it could never open its first trade.

What changed in 0.0.70

The lab evaluates more than the final score

When Research Lab chooses a base version, it now considers:

• whether the hard research requirements were met;
• whether the strategy produced real signals, orders, and trades;
• whether runtime evidence is healthy;
• whether the version made objective progress;
• whether the result passed a fresh Walk-Forward or Out-of-Sample check.

Expected risk stops are no longer classified as software exceptions. Genuine runtime failures still invalidate the evidence.

Problems are detected before an expensive backtest

XTester now checks generated strategies for several high-risk patterns before running them:

• unreachable entry logic;
• look-ahead in a Donchian breakout;
• counters updated in the wrong order;
• inconsistent ATR units;
• risk scaled twice;
• no protective exit.

These checks do not replace backtesting. Their purpose is to avoid spending time and model tokens on a candidate that cannot produce trustworthy evidence in the first place.

The user remains in control at a plateau

When a research run reaches a plateau, the lab now waits for the user’s decision. Questions and normal chat messages are preserved, while a direct instruction can be carried into the next hypothesis. Automatic continuation is allowed only after the normal iteration limit.

The cumulative LLM-token counter is also restored correctly when a research task is reopened.

Working and failed versions are easier to compare

The version tree now has filters for:

• all versions;
• working versions;
• failed versions.

A matching candidate can be viewed in the tree with its ancestor path preserved, or in a sortable table. The table shows summary metrics, fitness delta, and genome parameters, making it easier to see where progress occurred and where the research moved into a dead end.

Other improvements

Git history now includes a colored line-by-line diff preview with correct line numbers and a stable changed-file table. The main Commit action stays visible, while additional commit options are grouped in a compact menu.

The release also fixes:

• a false “m15 data unavailable” message after warm-up;
• attempts to parse a plain-text C# response as JSON;
• the Director-chat message queue;
• selected-version and pause-state restoration;
• window geometry after monitor or DPI changes;
• diagnostics for step-size, minimum-notional, and actual fills.

Release verification

XTester 0.0.70 passed 478 automated tests. The XTester, Launcher, and MCP Host builds, Velopack packaging, and installer release all completed successfully.

The regression suite covers the causes identified in the dead branches. The next level of validation continues in longer live research runs: the lab must not only reject technically invalid candidates, but also improve the risk profile of a working strategy on unseen data.

Download XTester 0.0.70

Download the latest XTester release for Windows

XTester helps evaluate trading ideas on historical data, but it cannot guarantee live-market results. Always review drawdown, out-of-sample stability, and execution risk before deploying a strategy.