A recipe without measurements?
If a meal turns out well but you never recorded the quantities, repeating it may be difficult. An attractive trading report without the robot version or settings has the same problem. Today we build a test record, not a contest for the prettiest equity curve. Trading-screen photographs provide visual context; they are not results from an account or test run by us.
What must be recorded?
Record the program version, input settings, symbol, broker, requested dates, test capital, costs and data mode. Compare the requested start with the actual test start and retain the journal. Insufficient history for initialization can shift the effective beginning. A screenshot of the final balance does not preserve these settings or explain the conditions under which the test ran.
Example one: two apparently identical tests
Two people run an Expert Advisor on the same symbol and obtain different results. Before declaring the robot broken, compare versions, dates, data and costs. One may use real ticks while the other uses open prices. For a strategy that reacts inside a candle, that difference can matter. Investigate the mismatch rather than selecting whichever output looks better.
Example two: change one thing
Save the first run as the baseline. In the second, change only the cost assumption while holding the program, inputs and interval fixed. The difference now has a clearer interpretation. Changing ten parameters and the dates at once obscures which change mattered. This procedure cannot guarantee returns; it makes comparisons easier to explain and audit.
Protect the held-out period
Reserve a period before tuning. If you repeatedly inspect it and adjust parameters to improve it, it is no longer untouched evidence. A historical forward test is also different from running a demo account in real time. Label each stage accurately. A favorable report chosen after hundreds of trials is incomplete without acknowledging the selection process.
Assignment and acceptance check
Create a test folder containing settings, version identity, full report, journal and assumptions. Repeat the baseline without intentional changes and record any differences; do not assume exact equality without checking. Then make one controlled change. A useful answer states what changed and what remained fixed. Next we read the report using several simple measurements that answer different questions.