The Five Pillars Tell You The Data Changed. They Can't Tell You It's Right.
The question
Sanity Check piece on why the "five pillars" (freshness/volume/schema/distribution/lineage) industry consensus is structurally incomplete — gives MAC a platform to introduce the 2D matrix framing. (Derivative of the 2026-05-11 acceptance-frameworks brief; the deliverable is the argument + evidence for an original re-frame, not a restatement of the pillars.)
What we already know (from the vault)
- [[2026-05-11-data-quality-acceptance-frameworks-vs-mac]] — the five-pillar consensus (freshness/volume/schema/distribution/lineage) is real and unanimous across Monte Carlo, Bigeye, Sifflet, and (operationally) Coalesce Quality. The vault's verdict: the five pillars are "essentially Aggregate × Temporal repeated four ways plus lineage." That single sentence is the spine of this piece.
- [[testing-matrix-template]] — the canonical MAC artifact. MAC is a 3×6 matrix: Scope (Column / Row / Aggregate) × Basis (Absolute / Relative:Source / Relative:Production / Relative:Reconciliation / Temporal / Human). Vendors are dense in the Temporal column and the Column-Absolute cell; sparse-to-absent across the three reconciliation cells and the entire Human column.
- [[2026-04-14-joe-reis-state-of-data-modeling-april-2026]] — MAC is "upstream of detection." Observability tooling (Monte Carlo, etc.) detects decay after the fact; MAC prevents it by making the definition of done explicit at build time. ~90% of practitioners report a modeling pain point — the gap is not boutique.
- [[2026-04-15-commoncog-two-types-of-data-analysis]] — every MAC cell must know whether it is experimental (one-shot) or operational (recurring). The five pillars only live in the operational/monitoring mode, which is part of why they miss the build-time acceptance question entirely.
- The audit-model skill operationalizes the matrix: it walks an engineer through all 18 cells to surface coverage gaps. The product is the gap-finding discipline, which is exactly what a flat pillar list cannot do.
What the web says
- Monte Carlo's five pillars (freshness, volume, schema, distribution, lineage) were coined by Barr Moses in 2020 and "remain the standard evaluation framework... the rest of the industry now uses as standard vocabulary" (montecarlo.ai).
- The pillars detect statistical anomalies and deviations from a learned baseline — not business-logic correctness. Monte Carlo's own framing: tools "use machine learning models to automatically learn your environment... anomaly detection techniques to let you know when things break." Business-logic correctness, reconciliation against external systems, and human plausibility are not pillar functions (montecarlo.ai).
- The industry itself draws the line we want to draw: observability is descriptive (learns patterns, detects deviations); quality is prescriptive (you define standards and validate against them). Observability "identifies that something went wrong"; quality "ensures data is accurate, complete, consistent, and reliable for a specific business purpose" (atlan.com).
- The kicker, in the vendor's own words: "A pipeline can run smoothly (good observability) while producing incorrect results (poor quality)" (atlan.com). This is the structural incompleteness, conceded by the category.
- Honest caveat #1 — the five pillars do NOT claim to be a quality acceptance framework. Monte Carlo positions them as catching "unknown unknowns" that testing (which covers "known unknowns") can't (montecarlo.ai). So the critique is not "the pillars are wrong" — it's "the pillars are a monitoring taxonomy that the market has quietly adopted as its whole mental model of data quality."
- Honest caveat #2 — there is a separate, older industry framing for fitness-for-use: the "six dimensions" (accuracy, completeness, consistency, timeliness, validity, uniqueness), descended from Wang & Strong 1996 (icedq.com, ibm.com). Any piece must acknowledge this exists or a sharp reader will say "the field already has a quality framework." The re-frame survives this because the six dimensions are a one-dimensional list of properties — still not a 2D test-authoring matrix (per [[2026-05-11-data-quality-acceptance-frameworks-vs-mac]]).
Convergences and contradictions
- Convergence: Vault and web agree exactly. The vault said the pillars are "Aggregate × Temporal repeated four ways plus lineage"; the web confirms they are anomaly detection against a learned statistical baseline. Same conclusion, two roads.
- Contradiction to manage (the strawman risk): The five pillars never marketed themselves as acceptance criteria, and the field has a separate fitness-for-use vocabulary (six dimensions). The piece is weak if it claims "the pillars are wrong." It is strong if it claims "the market collapsed observability (did it change?) and quality (is it right?) into one word, anchored on the pillars, and lost the second question."
- The one genuinely defensible MAC wedge: four cells the pillar/observability stack structurally cannot produce — the three reconciliation cells (does it match the source / production / external ledger?) and the Human cell (does someone who knows the business flinch?). A learned baseline cannot reconcile to Stripe or to a sales leader's gut. That is not a coverage gap; it is a category the framing cannot express.
Synthesis for RDCO
The re-frame, in one sentence: The five pillars answer "did the data change?" — a monitoring question — and the industry has quietly mistaken that for the data-quality question, which is "is the data right?" — an acceptance question the pillars are structurally unable to ask.
The piece works because the incompleteness is conceded by the category itself, not asserted by us. Monte Carlo's own materials say the pillars detect "unknown unknowns" via ML anomaly detection; Atlan's own comparison says "a pipeline can run smoothly while producing incorrect results." Freshness, volume, schema, and distribution are all the same shape of check — watch one metric over time, alarm when it deviates from a learned norm. Lineage is the map of where the alarm propagates. None of them encodes a business rule (ending_arr = starting + new + expansion - contraction - churn), reconciles to an external ledger (warehouse vs Stripe), or asks a human to look. They detect that a number moved; they are silent on whether it was supposed to. A revenue model can pass all five pillars every night — fresh, normal volume, stable schema, in-distribution — and still double-count a fan-out join into a number no human ever sanity-checked. The pillars would never flinch.
The honest framing — and the one that keeps the piece from being a strawman — is that the five pillars are an excellent monitoring taxonomy and a terrible acceptance taxonomy, and the industry has been using a monitoring taxonomy as its entire definition of data quality. That's the structural incompleteness: not that the pillars fail at their job, but that they only have one axis. They vary what surface gets watched (and even that mostly collapses to the aggregate/dataset level), but they have no axis at all for what you're checking against. Every pillar checks against the same thing — the data's own recent history. That is exactly one cell of MAC's matrix (Aggregate × Temporal), dressed up four ways.
That gap is MAC's platform. MAC's load-bearing move is adding the second axis — Basis: what are you evaluating against? Absolute (a fixed rule), Relative:Source (the upstream layer), Relative:Production (the existing report/ledger), Relative:Reconciliation (an external system), Temporal (history — this is the only basis the five pillars cover), and Human (someone who knows the business). Cross that against three Scopes (Column / Row / Aggregate) and you get 18 cells. The five pillars occupy roughly one of them. The piece should end not by claiming MAC has better checks — most individual checks exist somewhere in the vendor landscape — but by claiming MAC makes the missing axis visible, the way Kimball's bus matrix made dimensional-modeling gaps visible without inventing a single new primitive. The reader's takeaway: "I've been buying a smoke detector and calling it a building inspection."
Arc for the issue: (1) open with the failure mode — the model that passes all five pillars and is still wrong; (2) name the quiet category error — observability ("did it change?") masquerading as quality ("is it right?"), citing the vendors' own words; (3) introduce the missing axis (Basis) and the 18-cell matrix as the fix; (4) land on the four cells a learned baseline can never produce (the three reconciliations + Human) as proof the gap is structural, not incremental. Avoid the strawman by explicitly granting that the pillars are great at what they do — the indictment is what the market did with them.
Open follow-ups
- Should the piece visually render the five pillars inside the 18-cell matrix (one shaded cell) as the hero graphic? That single image may be the strongest MAC marketing asset we own — candidate for a tufte-viz pass.
- Does the "six dimensions" (Wang/Strong descendant) framing deserve its own rebuttal paragraph, or does mentioning it once inoculate enough? Risk: over-explaining hands the reader a competing framework mid-piece.
- Is "observability is a smoke detector, not a building inspection" the right central metaphor, or does it undersell observability's genuine value (catching unknown-unknowns)? Test a less adversarial metaphor.
- Worth a companion Data Dots micro-post: "5 pillars = 1 of 18 cells" as a standalone visual hook for distribution (LinkedIn/X)?
- Does Monte Carlo's 2026 extension into "Data + AI Observability" (agent behavior, output drift) change the gap, or just add more Temporal-axis cells in a new domain? Quick check before publishing so the piece isn't dated.
Related
- [[2026-05-11-data-quality-acceptance-frameworks-vs-mac]]
- [[testing-matrix-template]]
- [[2026-04-14-joe-reis-state-of-data-modeling-april-2026]]
- [[2026-04-15-commoncog-two-types-of-data-analysis]]
- [[2026-03-30-founder-data-quality-framework]]
Sources
Vault:
- /Users/ray/rdco-vault/06-reference/research/2026-05-11-data-quality-acceptance-frameworks-vs-mac.md
- /Users/ray/rdco-vault/01-projects/data-quality-framework/testing-matrix-template.md
- /Users/ray/rdco-vault/06-reference/2026-04-14-joe-reis-state-of-data-modeling-april-2026.md
- /Users/ray/rdco-vault/06-reference/2026-04-15-commoncog-two-types-of-data-analysis.md
- /Users/ray/.claude/skills/audit-model/SKILL.md
Web: