Reading a bridge

A bridge answers why a number moved, not just what it did. The habit worth building is reading the pieces before you quote the headline.

Footing

Every bridge foots: its legs sum exactly to the headline change. If ordered sales moved by a given amount, the traffic, conversion, price, and mix legs add up to precisely that amount, no residual left unexplained. That arithmetic is checked server-side before the result ever reaches you, so a number you quote from a bridge is a number the bridge actually carries.

Mix versus rate

A total can move for two different kinds of reasons, and a bridge keeps them separate. Rate is a change in how a given item performed on its own, its price, its conversion. Mix is a change in which items made up the total in the first place, so the total moved even if no single item's own rate changed at all. Collapsing the two together is how a real trend gets mistaken for a shift in product mix, or the other way around; the bridge exists so you do not have to guess which one you are looking at.

Caveats

The engine raises its own caveats rather than waiting for you to notice something is off:

Dark periods: a stretch with a known gap, such as a suspension, called out rather than silently averaged over.
Promotions: a window with promotional pricing, flagged so a rate change is not read as a durable one.
Surge windows: an unusual demand spike, flagged so it is not mistaken for a trend.
Matched windows: when two periods being compared are not naturally the same shape, the engine tells you it matched them so the comparison is like-for-like.
Restatements: when a prior period's data was corrected after the fact, the engine notes that the comparison period was restated.
End-of-life exclusions: items discontinued during the period are called out separately rather than folded into a plain decline.

A real example: a nine-day suspension made a headline change of plus fifty-five percent, until the dark-period caveat surfaced what was actually a like-for-like move of plus thirteen percent. The bridge did not hide the fifty-five; it explained it.

What We Know

Add evidence: true to a request and the result also carries "What we know": measured facts and candidate explanations, each attached to its own evidence, written by the engine rather than inferred by you after the fact.

The method

Attribution runs on HCAM, the Hierarchical Causal Attribution Model. It has two views: the Horizontal Bridge, which shows through which declared metric levers performance moved, and the Vertical Bridge, which shows where across the business hierarchy the movement occurred. The rule behind both is simple to state and easy to violate without an engine enforcing it: never publish a number the artifact does not carry. The full method is written up in the whitepaper, RAG-Enabled Hierarchical Causal Attribution Model.