2. Ground in evidence and context¶
Module outcome: a contract stating what evidence the decision needs, its reliability, what it cannot establish and what information would change the choice.
Data is not reality¶
Data is recorded evidence about part of reality. It may be accurate but irrelevant, timely but wrongly defined, or associative without establishing causality. Ask: Is it trustworthy and relevant enough for this decision, horizon and cost of error?
Six labels that prevent reasoning drift¶
| Label | Example | Required discipline |
|---|---|---|
| Observed fact | “420 applications were recorded.” | Source, period, population, definition |
| Calculation | “The rate was 12.4%.” | Formula, numerator, denominator, missingness |
| Inference | “The decline appears concentrated in one segment.” | Method, uncertainty, alternatives |
| Hypothesis | “Shorter response time might improve retention.” | Plausible mechanism and future test |
| Recommendation | “Prioritize segment A.” | Criteria, trade-offs, risk, authority |
| Unknown | “We do not know whether customer mix changed.” | Decision impact and gap-reduction plan |
The error occurs when a correlation becomes cause, a hypothesis becomes recommendation, or an AI recommendation becomes fact without notice.
Evidence contract¶
For every material signal document: provenance, definition, quality, population, context, uncertainty, fitness for purpose, counterevidence, and change condition—the information that would change conclusion or action.
Context outside the table¶
Context includes operational history, local definitions, incentives, capacity constraints, prior choices, exceptions and power. Capture provenance: “the team said so” is neither an approved rule nor quantitative evidence. Separate stable context, temporary context, local knowledge, and interests or incentives.
When to stop a conclusion¶
Do not advance to recommendation when a material denominator cannot be reconstructed; definitions changed unadjusted; affected segments are missing; a model is out of population; causal explanation rests only on timing; sensitive data lacks legitimate permission or purpose; or uncertainty changes option ordering. Stopping a conclusion need not stop the decision—it may require a smaller, more reversible option or guardrail.
Applied case · Planning with demographic composition¶
Practice decision. The planning team must choose where and when to expand a service for people aged 65 and over. The regional total establishes scale. To allocate capacity, the evidence contract should specify the target population, service area, planning horizon and review date.
Use the harmonized series to compare countries and check it against the relevant national projection for local planning. Separate estimated years from projected years and record the source, edition and period. If the two series place a material change in different years, record the documented assumptions behind the gap and which part of the decision is sensitive to it.
Exercise. Agree with the decision owner on a threshold for expanding capacity. Compare the decision under both series and record what you would do if they disagree, such as phasing the expansion and setting a review date. Later, check whether observed local demand supports the criterion. This exercise proposes a use for the projections and requires operational data for the actual decision.
Philia Sophia reading (2 Nov 2026) · Latin America is already aging while it is still growing.
Stress test¶
- What is the strongest counterargument?
- Which conclusion depends on one assumption?
- Could the pattern arise without the proposed mechanism?
- What group, period or outcome is invisible?
- Does the source define the concept or only report an observation?
- What uncertainty makes options indistinguishable?
Deliverable¶
Build one row per material item using the nine contract fields. End with three lists: we know, we infer, we do not yet know.
Read The Data Village to see how an ambiguous definition can change an action and what role each part of the architecture plays.