Feedback needs stable identity, exposure attribution, and two separate notions of time

Generate or preserve a stable logical event ID across retries and scope it so different producers cannot accidentally reuse the same key. A transport offset identifies a delivery location, which may differ from the user action's identity. Keep deduplication state for the promised horizon and apply its decision atomically with the relevant aggregate or state update when that guarantee is required. Duplicate suppression in one consumer does not imply that every downstream sink also applies an event only once, especially after replay or retention expiry.

Define whether the target is click count, unique clicked impressions, watch duration, or another outcome. Two valid clicks on the same impression can contribute two to an action count but only one to an at-least-one-click label. A repeated delivery should contribute neither an extra click nor a new impression. Similar timestamps or the same item ID are not sufficient proof of a duplicate. Schema and identity changes need versioned semantics so old and new producers do not silently apply different counting rules to the same evaluation window.

Carry an impression or decision ID into feedback so it can join the original item, position, request context, eligible set, policy version, and recorded action probability. A join to the current recommendation list can attach an old click to the wrong slate. Define what qualifies as an impression: server delivery is not necessarily on-screen exposure or examination. Missing client acknowledgments and incomplete joins affect the denominator and can differ by device or network cohort. A label should preserve those limitations instead of converting every successful server response into certain user exposure.

Attribution rules are analytical choices. A delayed action can follow several exposures, and last-touch, first-touch, or a fixed-window rule can produce different credit assignments. They do not automatically identify which recommendation caused the action. For causal policy evaluation, preserve the action and outcome unit from the experiment or identification design. Record corrections and deletion effects in a way that can reproduce a historical dataset version. A convenient many-to-many join can otherwise multiply one outcome across several impressions and create artificial lift.

A training snapshot claiming to represent what was known at an earlier cutoff must exclude events first available later, even if their action timestamps precede the cutoff. Otherwise features or labels can benefit from future knowledge. Conversely, a mature outcome dataset can intentionally wait for delayed feedback and use a later knowledge cutoff; document that horizon and the censored population. Watermarks express a stream processor's event-time progress policy, not proof that no late event will ever arrive. Choose allowed lateness, corrections, and reprocessing behavior explicitly.

The modular lab illustrates finite labels recurring, not an event-ID or deduplication service. Reset, set both modulus m=12 and additive step a=5, predict residues five after one Step and ten after two, then inspect. Reset and set both m=8 and a=2, predict orbit period four and return to residue zero after four Steps. Each Step advances one addition because it adds 0.5 seconds. The sequence is n×a mod m, starting at zero; it has no event records, time windows, joins, or collision-resistant identity. A recurring small numeric label cannot by itself distinguish arbitrarily many logical actions.