The pipeline, as a data model.

Evidence → decision insight

Four stages — Sources, Decision control, Execution, Products — expressed as entities and relationships.

erDiagram
  SOURCE ||--o{ ASSUMPTION : feeds
  DOMAIN ||--o{ ASSUMPTION : governs
  SCENARIO }o--o{ ASSUMPTION : "selects bands"
  MODEL }o--o{ ASSUMPTION : consumes
  RUN }o--|| MODEL : executes
  RUN }o--|| SCENARIO : under
  RUN ||--o{ OUTPUT : produces
  SOURCE { string provider string type date accessed }
  DOMAIN { string name string steward string cadence }
  ASSUMPTION { string key string value string unit string source float confidence date valid_to }
  SCENARIO { string name string bands string review_status }
  MODEL { string repo string sector string status }
  RUN { string model_version string assumptions_commit string scenario string who datetime at }
  OUTPUT { string artifact string provenance_json }

Read it as the pipeline: a governed ASSUMPTION (sourced, confidence-scored, governed by a DOMAIN steward) is banded by a SCENARIO and consumed by a MODEL. Stages 1–2 are live today. Stage 3 (RUN) is Layer 2 — a lightweight audit record of every execution, pinning the assumptions_commit, scenario, who and when. Stage 4 needs no separate registry: every OUTPUT artifact carries a provenance.json that embeds those same fields.

How models & assumptions meet — the promotion path

A model is born bottom-up (it defines its own assumptions) and matures top-down (it consumes governed ones). Maturity is one journey: an assumption goes local → governed → consumed. Which way you build depends on whether the registry already covers your domain.

Coverage decides what’s even possible: a thin or empty domain forces bottom-up — the model becomes the thing that defines what the registry should hold; a rich governed domain lets you go top-down from day one.

This is what CONSUMED looks like in practice

A 25-year railway feasibility model pulling 12 governed assumptions across 4 domains at runtime. Every input is sourced, confidence-scored, and owned. The model never touches a spreadsheet cell directly — it reads the registry.

Governed inputs consumed by this model

Model output — three scenarios, same governed inputs

The 1% difference between high and medium tariff escalation (5.5% vs 4.6%) moves NPV by R20bn. That one governed assumption is the swing factor. If it lives in a spreadsheet cell, two analysts will disagree on it — silently. In the registry, there is one value, one owner, and one source document.

Where each entity lives — go use it

The model is not a picture. Every entity is a running surface in this intranet.