An MT5 trade carries a comment, and the EA chose it. Most vendors use that field to name the setup that fired, so a finished backtest already contains the strategy breakdown that no product page prints. Our audits take the vendor at their word and add up the result behind each label. In the Lizard audit that turns 21,524 trades into six labelled groups, and exactly one of them loses money. In the Prop Firm Gold EA audit all three labels earn, and one of them dug 595.47 USD of the deepest 683.08 USD hole.
The rule is mechanical. We read the entry comment of every closed trade, strip the longest text that all of them share, and treat what remains as the key. Lizard writes Lizard_ in front of everything, so the keys come out as A1, A2, A3, B1, B2 and B3. When nothing remains, the trade lands in a group called default. Each group then gets its own trade count, its own result, its own win rate, its own median holding time and its three most frequent entry hours. The same key is handed to the drawdown decomposition, so the question "which setup dug this hole" is answered in the same vocabulary.
What this is not is our classification. The label is the vendor's text, copied out of their own trade comments. We only do the arithmetic behind it. A vendor who writes nothing, or writes the same string on every trade, produces exactly one group and no information at all.
Group results are net of costs, which matters more than it sounds. The commission retrofit runs before trades are paired, so the 3.50 USD per lot per side is already deducted from every single trade before any label is counted. The proof is arithmetic. Lizard's six groups add up to 183.99 USD, and 183.99 USD is exactly the audited net after 1,721.92 USD of retrofitted commission on 43,048 deals. The tester head for Prop Firm Gold EA shows 7,532.83 USD, the audited net is 6,769.79 USD, and its three labels sum to 6,769.79 USD. The labels always add up to the lower and more honest number.
Lizard earned 183.99 USD across 23 years of gold. The label split shows that this figure is a subtraction rather than a total. B1 is the largest group with 8,944 trades and it lost 1,163.62 USD. The other five labels earned 1,347.61 USD between them, led by A2 with 401.06 USD on 2,451 trades. B1 also wins least often, 21.7 percent against 39.3 percent for A2, and it holds longest, a median of 0.08 hours against 0.03 hours.
Read that as a buyer and the question changes shape. The run is not a strategy that barely works. It is five sub strategies that work and one that does not, bundled into a single product and a single balance curve. A preset switch that disabled B1 would be the most interesting input on the panel, and the audit says so with the vendor's own naming.
Prop Firm Gold EA does not name its setups at all. The shared prefix comes from an EA input called InpTradeComment, and the value that ran in our test is a row of X characters, which leaves a, b and c as the keys. All three earn. Label a made 3,408.95 USD on 3,712 trades, c made 1,931.68 USD on 1,067 trades, b made 1,429.16 USD on 4,759 trades. On the profit side there is nothing to complain about.
The drawdown decomposition tells the other half. The deepest episode on the end of day balance curve runs from 2026-05-08 to 2026-07-01, costs 683.08 USD, and had not recovered when the test ended. Fifty trades closed inside that window. Label b contributed 30 of them and 595.47 USD of the loss, a contributed 12 trades and 86.18 USD, c contributed 8 trades and 1.43 USD. The same label that earns least per trade over the full run is the one that concentrates the damage into the worst stretch of it.
Thirty three audit files in the catalog carry this block. Fifteen of them collapse into a single group called default, which means the EA writes the same comment on every trade it opens. Scalping Robot Pro is one of those. There is nothing wrong with it, and there is nothing to learn from it either.
The other end is just as awkward. Logan splits into 76 groups over 1,372 trades, with keys like "1 1st sell" and "6 1st buy" that encode a strategy number and the entry rank. Fewer than twenty trades per label on average is not enough to judge anything, and the largest single label holds 355 trades for 602.16 USD while the whole run lost 24,334.65 USD after 167.54 USD of commission and 6,448.44 USD of swap. Fine grained labels tell you how the EA thinks about itself. They do not automatically produce a verdict.
There is a flag attached to this section and it has never fired in our catalog. It turns to caution when a run has more than one label and the largest label, measured in absolute result, accounts for more than 80 percent of the summed absolute results. All 33 files stand at ok. Fifteen of them cannot trip it by construction because a single group never qualifies, and none of the remaining 18 concentrates that hard. A lamp that never lights is worth reporting as such rather than leaving in a corner, and Lizard shows why the threshold is hard to reach. Its dominant label is dominant in the losing direction, and the other five are large enough to keep the absolute share below the line.
The label split, the drawdown decomposition and the cost lines are public in every audit we publish. Your own tester report? The browser check is free.