"Week 2 — Recognizing the Same Ticket in Different Forms of Data" — @practicaldatamodeling
Why this is in the vault
Second live-cohort post in the Mixed Model Arts re-read (Chapters 4–6, discussion target Oct 28), direct sequel to Week 1: it moves from "which representations must agree" to the sharper operational question of how you decide two records are the same entity when the systems producing them share no identifier, plus a parallel point on attribute provenance that future AI-agent consumers need and usually don't get.
The core argument
Chapters 4–6 cover the building blocks: Ch 4 separates a data form (structured/semi-structured/unstructured/ML-artifact/metadata) from the physical format used to store it; Ch 5 covers entities, instances, and identifiers; Ch 6 covers attributes. Reis reuses the Chapter 12 support-ticket example and tightens it: a customer emails, then chats, then calls — a CRM record, a JSON conversation, and a phone transcript, each a valid instance of the same entity (Ticket, with a channel attribute), not three separate concepts. He adds the harder case directly: a follow-up through another channel where the systems don't share a ticket identifier. Fuzzy matching (e.g., email address) might help narrow it down, but it doesn't settle the identity question — that's a modeling decision, not a lookup.
The attribute half makes a parallel point: a field called date doesn't say whether it means creation, first response, or resolution; a sentiment score extracted from a transcript is categorically different from something a customer explicitly stated, even if both end up as the same-looking value in a table. Reis's framing is that both "what identifies this as one entity" and "where did this attribute's value actually come from" are decisions that have to be made and recorded — otherwise downstream consumers, now explicitly including AI agents and not just dashboards, are left to guess.
Mapping against Ray Data Co
The concrete connection: this is the entity-resolution-without-a-shared-key problem that shows up constantly in client data-modeling work — the same customer, deal, or claim arriving from disconnected systems with no common identifier, where a fuzzy match (email, name, account number) narrows the field but can't settle identity on its own. Reis's attribute-provenance point sharpens into something more specifically relevant to RDCO's agent-deployer thesis: an agent consuming a sentiment_score or a date column has no way to tell derived-from-a-model from stated-by-the-customer unless that provenance is carried as metadata alongside the value. That's a concrete requirement for any data harness feeding agents, not just a modeling nicety for humans reading a report — if the harness doesn't carry it, the agent silently inherits the ambiguity and acts on it with false confidence. This is a genuinely load-bearing parallel to the Week 1 "what must the two representations agree on" diagnostic, not a restatement of it: Week 1 was about operational vs. analytical state diverging; this is about identity and provenance diverging across channels with no shared key at all.
Related
- [[2026-10-16-practical-data-modeling-mma-book-club-week1-chapters-1-3]]
- [[2026-03-02-practical-data-modeling-mma-ch4-five-forms-of-data]]
- [[2026-03-04-practical-data-modeling-mma-ch5-entities-instances-identifiers]]
- [[2026-03-08-practical-data-modeling-mma-ch6-attributes]]