Run Discriminating Experiment First

Run the discriminating experiment first

Two things are almost the same. One behaves, the other does not. The instinct is to fix by the strongest hypothesis, verify, and if that fails move to the next hypothesis. That loop has a terrible payoff curve: a guess costs one full verification round and returns information only when it happens to be right.

A discriminating experiment costs the same round and returns information either way, because you decide before running it what each outcome rules out. Swap the two things' positions. Trade one attribute between them. Run both variants side by side in one build. Whatever comes back, half the candidate causes are gone.

The rule of thumb: count the variables that still differ between the working and the broken side. If that count is greater than one, you are not allowed to write a fix yet.

The recorded case

In a multiplatform media app, a header carried two controls using the same visual-effect helper: a 48dp round icon button on the leading edge and an elongated multi-icon pill on the trailing edge. The pill showed the glass rim. The button showed nothing.

Four verification rounds went to hypotheses — change the button's shape, remove a wrapper layer, feed the effect real content instead of a flat colour, copy the structure from the other orientation. All four were wrong, and the person running the builds had lost patience before any result arrived.

The fifth round swapped the two controls between the leading and trailing edges. The effect followed the pill, not the position. Position was eliminated in one round, and what remained was the widget's own geometry: the rim style in use was lit along a single direction, so a long edge catches the sweep and a small circle catches almost none of it. Recorded history for that area now says the four guesses were burned "before anyone read the library source".

Traps

Fixing while more than one variable differs. This is the whole failure. Two controls that differ in shape, size, nesting depth, sampled surface and position give you five candidates; a fix aimed at any one of them has a one-in-five prior and tells you nothing when it fails. Write the list down first — the act of listing is usually what reveals that you have five and not two.

Designing a test whose outcomes you cannot tell apart in advance. Before running anything, finish both sentences: "if the effect moves with the widget, then ___" and "if it stays at that position, then ___". If either sentence is hard to finish, the experiment does not discriminate and you are about to spend a round on nothing.

Changing two things inside the "swap". A swap that also re-indents, renames or tidies is no longer a swap. Move the two call sites and change nothing else, even things that look obviously wrong on the way past. Fix those in a separate change after the result is in.

Deleting a live experiment during a tidy-up pass. Recorded in the same episode: one screen's variant was removed while "making things consistent", and the build it was riding in came back useless. An experiment is not untidy code — it is an instrument with a reading pending. Mark it so a cleanup pass can see it, and do not remove it until somebody has read the result.

Running one variant per round when the round is expensive. If a build is the bottleneck, put several variants in the same build — one per screen, one per row, one per flag — and label them. Four hypotheses in one round beats four rounds. This only works if the previous trap is respected.

Treating a partial improvement as confirmation. In the recorded case, widening the round button to roughly twice its size did make the effect appear. That looks like a solved bug and is not: the longer edge simply caught more of the directional sweep. A change that improves the symptom without your being able to say why is a second observation to explain, not a fix to ship.

Believing a comment or a changelog entry about which variable matters. Prose in the repository records what somebody concluded at the time, and the code beside it may have moved on. Treat it as a hypothesis to test, in exactly the same queue as your own.

Asking a human for a verification round you could have run yourself. Each round you delegate costs someone else's attention and arrives minutes-to-hours later. Before requesting one, check whether the question can be settled by reading an artifact instead — the dependency's own source, the resolved artifact list, the generated output. The four wasted rounds above were all answerable from the library source without a build.

Verifying it

1. List the differing variables and check the count. Pull both call sites of the shared helper into one view and diff them; every surviving line is a candidate:

   HELPER=TheSharedComposable          # the widget or function both sides call
   grep -rn "$HELPER(" --include='*.kt' .
   

   More than one difference means: design an experiment, do not write a fix.

2. State the outcomes before you run it. Write the two "if … then …" sentences into the change description or the ticket. If you cannot, redesign.

3. Confirm the experiment is still alive before the result is read. Tag every variant with a marker word and check that no cleanup pass has quietly removed one. Bare git diff compares the working tree to the index only, so a cleanup that was staged — or committed, the ordinary case once somebody has tidied and moved on — passes it silently:

   grep -rn 'EXPERIMENT' --include='*.kt' . | wc -l
   git diff HEAD -- '*.kt'                   | grep -E '^-[^-].*EXPERIMENT'
   git log -p <branch-point>..HEAD -- '*.kt' | grep -E '^-[^-].*EXPERIMENT'
   

   Pass condition: the count matches the number of variants you planted, and both greps print nothing. ^-[^-] rather than ^-, so the --- a/path diff header cannot match instead.

4. After the result, prune the hypothesis list explicitly. Name which candidates the outcome eliminated. If it eliminated none, the test did not discriminate — that is the finding, and the next step is a better test, not another fix.

5. Remove the markers in their own change, once the cause is known, so the diff that proves the fix is not tangled with the instruments that found it.