Read-Points in Abundance, Attribution Still Impossible: The Shrink Ceiling

You solved tagging years ago. Item-level RFID is mature in your chain: the tags are on the goods, the readers are at the doors and the transition points, cycle counts run daily, and your on-hand accuracy sits at numbers your predecessors would not have believed. And the accuracy gains have plateaued — because read-point counting has said everything it can say, and the shrink that remains lives in the one dimension your read-point data does not carry.

An honest label before the argument, because this post's credibility depends on it: general merchandise and apparel carry no compliance clock and no citable third-party demand artifact for what follows. The food segments this journal also serves have public rollouts and a regulation to point at. You have neither, and we will not manufacture one. This argument stands on structure or it does not stand. Here is the structure.

The investigation transcript

Every LP investigation you run ends the same way, so run one on paper. A tagged item — pick a $400 jacket — moves through your chain, and the read history is genuinely excellent:

Annotate what each line answers. The inbound read answers where and when. The sorter read answers where and when. The floor transition answers where and when. Then the item is seen here, then not — and the investigation dies on the only question it ever actually had: between Friday's count and Saturday's, who handled it? Six readers produced a flawless account of the item's places and none of them can name a single pair of hands. Add a seventh reader and you will know the places more precisely. You will still not know the hands.

That is the ceiling. It is not a gap in your deployment; it is a gap in the record your deployment writes. EPCIS 2.0 — the event standard item-level programs speak — defines an event in five dimensions: what, when, where, why, how (§7.2.2). No performer is among them, and the party fields it does provide are organisation-grain: a company, a PGLN (CBV 2.0 §8.7.1). The record can say your company moved the item through zone C. It structurally cannot say who carried it out of zone C. The section numbers and the ten-minute self-check are laid out in Five Dimensions, No Performer — the fact is checkable without believing a word of ours.

The remaining shrink is concentrated exactly there

Your read-point investment already sorted your shrink for you. External theft announces itself at exits your readers watch. Vendor and process loss surfaces in reconciliations your counts catch — and the dispute half of that story, where both sides have data and neither has evidence, is its own post. What is left — the residual your plateaued accuracy numbers cannot touch — is concentrated in handoffs: back-of-house to floor, floor to stockroom, store to store, the transfer that left complete and arrived short. A handoff error is a who problem by definition. Two people, one item, one moment — and a record with no field for either person.

So the marginal dollar of read-point spend now buys the answer to a question you already have, while the question you actually have goes unanswered at any spend. That inversion is the purest form of the argument this whole journal makes, with no regulatory scaffolding to lean on: attribution is not a better read-point counter. It is a different dimension.

What performer attribution adds to the same history

Now replay the transcript over a record that carries the two grains the standard lacks: who, the attested observer on every event — a person, an agent, or a machine, resolved through id.org.ai — and capturedBy, the warrantor account that stands behind the capture. Never one field, never collapsed; the distinction and its failure modes are worked in who Is Not capturedBy.

The same investigation now reads: Friday's count — counted by a named, attested observer. Saturday's — a different observer, whose count is short by one. Between them, the floor-to-stockroom handoffs, each carrying its observer and its warrantor. The dead-end question — who handled it? — becomes a list. Not an accusation: a list of attested handlers between last-seen and first-missing, recorded at the moment of each act rather than reconstructed from schedules and camera timestamps a week later. Every investigator on your team knows the difference in hours; your exception views render it directly, an attributed timeline per item instead of a read-point scatter — a queue your own analyst can hand to a deputized agent once the volume proves out. And because the attribution rides in the record itself — conformantly, as a superset of EPCIS 2.0 whose projection still validates against the official schema — it holds up in the two places a shrink finding eventually travels: the HR conversation and the claims dispute.

The purest test

Strip this segment of everything the food posts carry — no FDA rule, no supplier re-marking wave, no third-party rollout to cite — and what remains is the differentiator with nothing propping it up: your read-point data answers where; your losses happen in who. If that sentence is wrong for your chain, nothing here applies to you. If it is right, no additional reader density will ever make it wrong, and the question becomes when your record gets a field for the answer.

The door here is the get-started interview: your email first, then a short interview that branches on your answers — the retail branch asks how you decide today whether a loss is theft, damage, or a scan error, what evidence you attached to the last claim you filed, and how many of these you see a month. We answer in writing.