Lineup training for detection dogs: how to log it, and what the numbers can tell you

Every run comes down to two facts: where the target was, if anywhere, and which positions the dog indicated. Every verdict follows from those two, so there are six possible outcomes:

• ▸Correct — the dog indicated the target and nothing else.
• ▸Found + extra — it found the target but also indicated a position without it.
• ▸False alert — it indicated only positions without the target.
• ▸Miss — there was a target and the dog indicated nothing.
• ▸Blank clear — there was no target, and the dog indicated nothing.
• ▸Blank called — there was no target, and the dog indicated anyway.

Write down where the target was and what the dog said, and let the outcome be worked out from that. If you later decide to score something differently — a false alert followed by the right box, say — the whole history can be re-scored in one go. A log of verdicts can only be typed out again.

Number the positions in the order the dog works them, not by where they stand in the room. Position 1 is the first box the dog reaches, whichever end of the line that is. That is what makes a position comparable across sessions set up in different places.

A handler who places the target brings their own habits along: the middle box feels natural, the end box feels unfair, the same box twice in a row feels wrong. The dog can learn those habits as well as the odour, and per-position statistics then describe the handler rather than the dog.

Drawing the position removes the habit. Rolling a die is not quite enough, though: an independent roll makes every position equally likely, but over sixty runs on six boxes each position is expected ten times with a spread of about three either way, so one box holding the target five times and another fifteen is ordinary. Dealing from a shuffled deck that holds every position the same number of times keeps the counts close to even.

A blank run has no target in the line at all. It is the only way to see whether the dog will leave a line alone — and a dog that always finds something on every line may be learning that something is always there.

How many to run depends on the stage of training: plenty of targets while a new odour is being built, more blanks as the work gets closer to real searches, where most areas hold nothing. Whatever the share, a blank only means something if it is run exactly like every other run — the person holding the lead should not know in advance.

Position bias — the dog favouring the first box, or the last — is one of the most useful things a lineup log can show, and one of the easiest to over-read.

Sample size comes first. A position that held the target ten times and was found eight shows 80%, but the honest range for eight out of ten is roughly 49% to 94%. At fifty runs the same 80% narrows to about 67% to 89%. Look at the count next to a rate before the rate.

Then there is picking the worst of several. With six positions, one of them will look worst by chance alone, and it will tend to look better next time whatever you change — which reads as proof the change worked. Treat a weak position as a question to test over the next sessions, not a conclusion.

Runs in one session are not independent either. They share the weather, the ground, the handler's mood and the dog's warm-up and fatigue, so a hundred runs over three days carry less information than a hundred over thirty.

Two dogs pooled into one success rate are two measurements pretending to be one. The same goes, less sharply, for different odours. Read the figures per dog, and per odour where the numbers allow it.