06-reference

practical data modeling mma book club week4 chapters 10 12

2026-11-08·reference·source: Practical Data Modeling (Substack)·by Joe Reis

"Week 4 — Keeping Track of What We Knew and What We Meant" — @practicaldatamodeling

Why this is in the vault

Fourth live-cohort post in the Mixed Model Arts re-read (Chapters 10–12, discussion target the week of Nov 11): it names the exact failure mode behind one of RDCO's own already-logged operational mistakes — overwriting a status instead of keeping both when it was true and when it was recorded — so this is a direct theoretical grounding for a lesson already learned the hard way.

The core argument

Chapters 10–12 move from "what does this field mean" to "when did we know it, and does our model let us keep both answers." Reis's running example: a support ticket arrives at 11am and is marked medium priority; at 2pm a manager notices the customer has Premium Support and the ticket should have been high priority all along. If you overwrite the priority field, the report looks corrected but you've silently erased the fact that the team spent three hours working with different information. Chapter 10 is the fix: distinguish valid time (when a fact is true in the world) from system time (when the system recorded it), so you can ask both "what do we believe now" and "what could someone have known when they acted."

Chapter 11 layers on a second problem that time alone doesn't solve: two people can agree on a field's value while disagreeing about what it means. Does "resolved" mean the agent closed the ticket, the customer confirmed satisfaction, or a fixed period passed without reopening? A timestamp can't settle that — it needs an explicit definition that fits the question being asked. Chapter 12 synthesizes: start from the business question, then work through data forms, shared entities, grain, time, and semantics before picking an implementation pattern, because a decision at any one layer (like correcting that priority field) can silently change which records belong in an SLA calculation or how an agent's performance reads.

Mapping against Ray Data Co

The single most concrete hit: this is the formal name for the bug already logged in feedback_verify_blockers_against_source_not_notes — a 🔴/blocked status got written into a note, the underlying situation changed, and the stale note was read later as a live blocker twice before the fix (re-check at the source before anything downstream reads it). That's Reis's ticket example exactly: a status got corrected at the source, but the old note had no mechanism to distinguish "true as of when this was written" from "true now," so the overwrite (or the unrefreshed copy) silently erased the fact that an earlier decision was made on now-outdated information. The fix the feedback memory landed on — re-verify at the source, don't trust the note — is a system-time patch; Chapter 10's actual prescription is a valid-time one: keep both the correction and the window during which the team legitimately believed the old value, so a later audit of "why did we treat this as low-priority for three hours" has an answer instead of a gap.

Chapter 11's semantics problem maps just as directly onto feedback_workflow_agent_output_integrity's catalogue of factory failure modes — "false verified stamps" and "proposal-as-fact" are exactly the ambiguous-field-value trap Reis describes with "resolved": a workflow or sub-agent writes done or verified into a result, and nothing forces agreement on whether that means a critic checked it, the founder saw it, or the agent merely finished running. The AGENTS.md hard-rule chain's own "newer wins" memory-conflict rule is a live instance of the same discipline Reis is naming: a memory gets superseded, and the correct move is to log the correction rather than silently overwrite, which is bitemporal thinking applied to the vault's own memory system without ever calling it that.

Related