GuideSecondary informationLocal-only model

How Smartly cashback attribution works

This page keeps policy notes, statement caveats, and confidence language out of the main analyzer while preserving the reasoning behind every verdict.

Back to app
Evidence hierarchy
Statement totals are stronger than MCC guesses
The CSV tells us merchant, amount, date, and MCC. The statement reward summary is the source of truth for what U.S. Bank actually awarded.
1

Match statement period purchases against the CSV window.

2

Compare base points and Smartly bonus points against the model.

3

Separate adjustments, deals, refunds, and travel bonuses where possible.

4

Only then use residual math to infer which rows likely earned 4% or 2%.

Privacy
The app path stays local
Production users upload a CSV in the browser. Development admin mode can preload ignored local files for faster analysis.

Keep real statements in data/. Committed fixture files belong in tests/data/.

Verdicts
Use precise confidence labels
The UI should separate model-only output from statement-backed evidence.

Statement-backed 4%

Statement math supports base points plus Smartly bonus.

Statement-backed 2%

Statement math supports base points only.

Modeled 4% / 2%

The MCC model has a direction, but there is not enough statement evidence yet.

Partial or cap-limited

Some of the purchase may have received the Smartly bonus before a cycle cap or boundary effect.

Unresolved

The statement does not reconcile cleanly enough to trust row-level attribution.

Statement eras
Do not learn every era the same way
Policy and reward-bucket wording changed over time. Era context should affect confidence.

Legacy flat 4%

Useful for that historical statement, but weak evidence for current MCC learning.

Transition bucket era

Statement labels changed, so attribution should stay cautious until residual causes are understood.

Post-policy split bonus

Best source for current 4% / 2% learning when purchase totals and points reconcile.

Workflow
What belongs in the app
The primary analyzer should stay task-first. This guide can carry the deeper product and rewards-accounting context.

Import

CSV first, then statement data when available.

Reconcile

Compare modeled points against closed-statement totals.

Inspect

Review transaction-level verdicts and confidence.