← Capability library

Semantic Readiness · Context Stability

Time intelligence

Prerequisites: Units/currency/time handling, Grain consistency

Definition

The model's handling of periods, comparisons and as-of questions: fiscal calendars, partial periods, period-over-period logic, and the distinction between what was true at a point in time and what the current record says was true then.

Why it matters

Time is where "correct" and "meaningful" diverge most often. A period-over-period comparison against an incomplete current period is arithmetically correct and analytically worthless, and an agent will produce it without hesitation. The as-of problem is worse: most warehouses overwrite dimension attributes, so a question about last quarter's performance by region is answered using this quarter's region assignments. Nobody notices because the number looks reasonable.

Failure modes

  • Month-to-date compared against a full prior month, presented as a decline.
  • Fiscal calendar implemented in the report layer, so the agent querying the model gets calendar months.
  • No date dimension marked as such, so time functions silently degrade.
  • Type-1 dimensions used for attributes that change (territory, manager, product hierarchy) making all historical analysis retroactively restated.
  • Multiple date columns with no declaration of which is the reporting date for which question class.
  • Event-time versus processing-time confusion at the boundary between the warehouse and the event stream.

Anti-patterns

  • A date table that is not marked as one. The functions still return values; they just stop being right at the edges.
  • Restating history silently. Type-1 overwrites are a legitimate design choice and an illegitimate default. If history restates, the answer surface must say so.
  • Handling partial periods in the visual. A footnote on a dashboard telling the reader the current month is incomplete does nothing for the agent, which will compare it anyway.

Maturity anchors

| Level | Name | What is true | |---|---|---| | 1 | Reactive | Date logic is written per report. No shared calendar. | | 2 | Governed | A conformed date dimension and fiscal calendar exist and are documented, including which date column is authoritative per fact. | | 3 | Contextual | Partial-period comparisons are blocked or explicitly labeled by the model, and attributes requiring historical accuracy are tracked as slowly-changing with as-of query support. | | 4 | Operational | As-of correctness is tested against known historical values, and the rate of partial-period and restated-history answers is measured. | | 5 | Autonomous | The model detects an incomparable period pairing and returns the comparison with its own caveat rather than requiring the consumer to know. |

Diagnostic question

If someone asks the agent how last quarter compared to the quarter before, what stops it comparing an incomplete period, and does it use last quarter's org structure or today's?

  1. Nothing, and it uses today's structure.
  2. We have a conformed calendar and documented conventions; correctness depends on the asker.
  3. Partial periods are blocked or labeled by the model and history-sensitive attributes are tracked as slowly-changing.
  4. As-of correctness is tested against known values and incomparable-period rates are measured.
  5. The model self-caveats incomparable comparisons.

Evidence to request

  • The date dimension definition and its marking, plus the fiscal calendar logic and where it lives.
  • The list of dimension attributes tracked as slowly-changing, and the list of history-sensitive attributes that are not.
  • An as-of query example: the same report run for a historical date, twice, six months apart.
  • The convention for event time versus ingest time at the streaming boundary.

Verification

Ask the agent for a period-over-period comparison mid-month and see whether the answer carries any indication that the current period is partial. Then ask for a metric broken out by an attribute you know has changed: a sales region reorganization is the usual candidate, for a period before the change, and check whether the answer reflects the old structure or the new one. Both tests take under five minutes and neither requires model access.

Article angle

The as-of problem is the better half of this capability and the less written-about. The hook is that most enterprises cannot answer "what did we believe at the time," which is the question every post-incident review and every regulator actually asks. That connects Pillar 1 time intelligence directly to Pillar 2's retrospective causality: the same defect, discovered in two very different rooms.

Sources