Agent Lab

Comparing results

Pin a baseline and explain changes using matched data, settings, and trade records.

Use one completed run as the baseline for a focused source change. Keep its identifier available so you can return to the original result after several iterations.

Pin and compare

  1. Open History and choose Use as baseline on the reference run.
  2. Select the candidate run.
  3. Open Metrics to view selected and baseline values.
  4. Compare their revisions and execution settings, then inspect changed trades and logs.

The Lab warns when the runs have different dataset identifiers. A matching dataset does not imply matching capital, quantity, fill delay, or other execution settings.

Interpret the summary

MetricCurrent meaning
tradesCount of returned trade records
netPnlSum of numeric P&L values in those records; missing values contribute zero
winsCount of records whose P&L is positive

Trade records can include entries as well as closes. These metrics are not automatically a count of completed round trips or a round-trip win rate. See reading results before deriving additional statistics.

Change one thing at a time

For an entry-filter experiment, keep the instrument, dates, interval, timezone, capital, sizing, fill delay, and session settings fixed. Review the first trade that differs: which condition changed, what request followed, and when it filled.

Use a prompt such as:

Compare the candidate with the pinned baseline.
Identify the first changed entry and explain it from the source.
Report the saved metrics, but do not infer a round-trip win rate
from the number of returned trade records.

Test another period deliberately

Once the source change behaves as intended, test it on a separate period. That is a new data comparison, so record the changed period rather than treating it as a direct source-only improvement.

Keep the selected revision and the run that supports it clear before publishing.

On this page