The control pattern
Data asset
Evidence, metadata and model trace data
Knobs
Provenance depth, Factor visibility, Lineage, Citation granularity, Confidence reporting, Version disclosure
Outcome
Auditable decisions and understandable scores
Measurement
See how to measure this below
What it is
Explainability shows how an output was produced and what evidence supports it. Evidence and metadata are the data asset; the depth of provenance, factor visibility, citation granularity, confidence reporting and version disclosure are the knobs.
An explanation should be faithful to the actual system rather than a persuasive story generated after the fact. For deterministic scores, expose the inputs and calculations. For model interpretations, show supporting passages, uncertainty and reasonable alternative readings.
Good explainability does not claim that a model's internal reasoning is fully knowable. It gives users the external evidence, transformations, assumptions and limitations needed to audit the result.
The knobs in detail
Each row is one adjustable property of the data asset, and what moving it tends to do.
| Knob | What you adjust | Likely effect |
|---|---|---|
| Provenance depth | How far back the trail goes | Determines what can be audited |
| Factor visibility | Which contributions are shown | Makes a composite score legible |
| Lineage | Which transformation produced the value | Connects display to source |
| Citation granularity | Section and passage vs. homepage link | Decides whether evidence is checkable |
| Confidence reporting | Decomposed or single number | Conveys why certainty is what it is |
| Version disclosure | Which formula and model produced this | Enables reproduction later |
Applied: Stocks Assistant
The Stocks Assistant Health Score is naturally explainable because it decomposes into Momentum, Fundamentals, Sentiment and Outlook. Display each sub-score, its weight, major contributing factors, peer group, data timestamp and formula version. A counterfactual view can answer questions such as how the score would change if outlook weight were lower, or if the company were compared only with sector peers. For earnings summaries, citations should open the exact filing section or transcript passage rather than merely link to a company homepage.
How to measure it
Evidence that the knob produced the intended behaviour, rather than shifting the problem elsewhere.
- Explanation coverage across outputs
- Citation validity
- User comprehension
- Whether stated factor contributions reproduce the displayed score
- Confidence decomposed into source quality, completeness, extraction certainty and signal stability
Common mistakes
Post-hoc narratives that sound convincing but do not describe the actual computation.
One opaque confidence percentage where four components would say far more.