The enabler bench

A match is a claim that two rows are the same thing. So the unit of scoring is the pair: every two rows your key groups together is a claim, every two rows that share a hidden real id is the truth. Three counts come out: predicted pairs P, true pairs T, and correct pairs TP, the claims that were true.

Recall (TP / T) alone rewards a key that joins everything: masked name finds all 339 real customer pairs in the sample, and also claims 2,370,747 false ones. Precision (TP / P) alone rewards a key that joins almost nothing. TP / (P + T - TP) is the share of correct pairs among all pairs that are either true or claimed, the Jaccard index of the two sets. It is high only when a key finds the real links and invents almost none.

Press "Break it" and the bench shuffles the hidden ids across rows, with a fixed seed. The keys are untouched; only the truth is scrambled. If the score measured matching, every grain must collapse, and it does: the best any combination reaches falls to 0.0 percent at customer grain, 0.4 percent at store grain and 0.0 percent at SKU grain. If a grain had held up under the shuffle, the bench would have been measuring the shape of the data, not the join.

The sample keeps Olist's real repeat rate: 250 of the 2,801 customers who bought more than once, and 8,083 one-time buyers, the real ratio of 2,801 to 90,557. That ratio matters, because false pairs are made by strangers who collide on a key, and a sample stuffed with repeat buyers would flatter every weak key.

One honest limit points the other way. False pairs grow with the square of the number of rows and true pairs grow roughly with the count, so at 8,614 orders the fragment keys look better than they would on an enabler's full order history. Treat the customer-grain results as a ceiling.

1. No key combination gets the customer past 23.3 percent. Masked fragments either collide (masked phone finds every real link and 370,684 false ones) or go silent (name, phone and zip is never wrong and finds 19.2 percent). You can say: "on realistic masking, a cross-platform customer join is a guess, not a key."
2. The hashed phone links nothing. 265 orders carried a raw phone and no two were the same person. You can say: "hashing is not a workaround for masking; it needs the value that was masked."
3. Serai's own keys hold. Store id 100.0 percent, SKU 97.5 percent. You can say: "the cross-platform view is real at store and SKU grain, today, on our own portfolio."
4. Owning a key means maintaining it. The 2.5 percent SKU gap is 219 lines on listings missing from the mapping table. You can say: "the mapping table is an owned asset with an owner and a weekly check", which is analyst session 2's job.

For record linkage done properly, with field weights and blocking on millions of rows, see the entity resolution course. The bench asks a narrower question: does a key exist at all.

Order id is unique per order, so at customer grain it claims no pairs and scores 0.0 percent, which is correct: two orders never share one. At store and SKU grain it links the lines inside one order, a handful of true pairs, and scores 0.9 and 7.7 percent. A key that should score near zero and does is the cheapest sanity check a bench has, the same idea as the break button in miniature.

• Pricing first returns $511,972 less to Serai and pays back two quarters later, while adding 3.9 times the GMV. Per live week it is worth $992.62 to Serai and $69,230.77 in GMV to the brands; Serai sees 2 percent of that GMV. It pays the brand.
• The single customer view first costs 16 build weeks and returns nothing, because its grain did not clear the gate. The order ends at $613,747, the lowest of the four.
• Break the identity ladder and nine of ten use cases stop. Hours first falls to -$510,867. An enabler roadmap rests on its joins more than on its models.
• Hours first and brand intelligence first are within 3.2 percent ($1,372,287 against $1,415,738). The gap is made of attach revenue the assumptions block stipulates, so the bench does not rank one above the other. The case for hours first is certainty: it needs nobody's permission and no buyer, which leader session 5 builds on.

The anti-lever carries its own limit, and the bench shows it. Raise Serai's GMV-linked fee to 6 percent and the pricing uplift to 6 percent, and pricing first overtakes hours first, $1,900,438 against $1,612,656. At a 10 percent share it wins from a 4 percent uplift. An enabler paid mostly as a share of GMV is a different business, and for that business the order flips.

This is the playbook's sequencing rule made visible in a single input, and it carries the playbook's status: a working heuristic, uncited, which the bench demonstrates on invented numbers and does not prove about any real company.