Compound ideas get decomposed, and the decomposition has a blind spot.

Why decomposition is forced

A scenario that moves two things explains neither.

Run a test that changes branch weights and stop policy simultaneously, observe an improvement, and the result is uninterpretable in a specific and permanent way: there is no arithmetic that recovers how much of the improvement came from each change. The improvement may even conceal a deterioration — one change helping more than the other hurt. Because a scenario's whole purpose is to attribute an effect to a cause, a multi-variable contract does not produce a weak answer. It produces no answer, dressed as a number, which is worse because it will be quoted.

The chain's construction

Same baseline, same evidence window, same seed — every link.

A chain is only comparable if its links differ in exactly the intended way. Each contract runs against the identical Standard baseline rather than against the previous link's result, over the same evidence window, and where the workbook permits it, on the same random draw — because a fixed seed means the difference between a link and the baseline is attributable to the change rather than to the sampling. Chaining against the previous result instead of the baseline is the common error: it produces a cumulative story in which every subsequent link inherits the noise and the assumptions of the one before it.

The blind spot, stated honestly

Sequential isolation cannot detect interaction.

Here is the limitation the discipline buys its clarity with. Three variables tested one at a time answer three questions about individual effects and are silent on whether the effects combine, cancel, or compound. A looser filter may be neutral alone and harmful only when paired with heavier weighting toward a high-variance branch, because the pairing changes what the filter is admitting. Nothing in a chain of single-variable contracts will surface that. This is not a flaw to be engineered around at the operator level — it is the known cost of attributable testing, and the correct response is to know it rather than to forget it.

What to do about it

Promote sequentially, and treat each promotion as a new baseline.

The practical mitigation is procedural rather than analytical. Where a chain produces two favourable verdicts, they are not promoted together. The first is promoted, becomes the new Standard baseline, and the second is then re-tested against that baseline before its own promotion. The re-test is where interaction becomes visible: if the second change's advantage evaporates or reverses once the first is live, the two were interacting and the chain's original verdicts were incomplete. This costs an extra persistence window and is the only place the interaction question gets asked at all.

The ordering question

Which link goes first is itself a decision with consequences.

Because sequential promotion makes each change a new baseline, the order of promotion can change the outcome. Two changes that both look favourable in isolation may resolve differently depending on which is adopted first, and there is no neutral ordering. The defensible default is to promote in order of evidential strength — the change with the widest margin over the longest persistence window goes first — with the reasoning recorded in the verdict, so a later reviewer can see that an ordering was chosen rather than defaulted into. An unrecorded ordering is an unexamined assumption sitting underneath everything downstream of it.

• Every link runs against the Standard baseline, never against the previous link.
• Favourable links are promoted one at a time, with a re-test against the new baseline between them.
• The promotion order is a recorded decision, justified by evidential strength.

The key idea

Attributable answers are worth what they cost, and they cost interaction.

The single-variable rule is not fussiness. It is the difference between a research programme that accumulates knowledge and one that accumulates confident numbers nobody can trace to a cause. The chain is how a compound idea is made testable, sequential promotion is how interaction gets its one look, and the blind spot in between is a real limitation that the discipline is better for naming out loud. A method that concealed its own limits would be a poor advertisement for a system built on stating them.