On this page
A financial director and an operations manager can look at exactly the same movement in a dataset and ask quite different follow-up questions. One may want to understand its financial exposure; the other may be interested in service capacity or the pattern of work behind it. Neither needs a different calculation. They need an explanation that respects the responsibilities they bring to the result.
This distinction became an architectural question for determinIQ: how can a system be attentive to the person reading an analysis without allowing that context to change what the analysis says?
One analytical result, several legitimate perspectives
A result has at least two parts. There is the governed analytical fact — what was calculated, over which records, with which operation and constraints — and there is the account we give of that fact. The second part is useful precisely because people do not arrive with identical experience, terminology or interests.
Consider an illustrative increase in a monthly measure. An executive briefing might begin with the scale and direction of the change. An analyst might want the calculation basis and the coverage of the underlying records. An operations manager might want to investigate whether the change was concentrated in a particular period or process. Those are different routes into the same evidence, not separate permissions to produce different answers.
In the design, the user’s narrative role, organisation context and narrative priorities can affect wording, depth and emphasis. They must not change a result value, analytical ranking, materiality, salience, evidence identity or verification state. This is not a small editorial preference; it is an authority boundary.
Keeping explanation on the presentation side
In the determinIQ documentation, a narrative context resolver brings the permitted contextual information together, and a presentation contract binds it to a particular output surface. The important public principle is that context is applied after the analytical basis has been established. No audience profile should become an alternative calculation engine.
Research figure 01
One result, different explanations
Audience context belongs downstream of the governed facts
- ANALYTICAL AUTHORITYCalculate once under governed rulesA supported operation establishes the result
- STABLE EVIDENCEPreserve the analytical basisValues, verification and identity remain anchored
- PRESENTATION CONTEXTResolve the reader’s needsRole, organisation and narrative priorities
- CONTROLLED EXPLANATIONAdjust language and emphasisTone and depth may change; facts may not
- HUMAN JUDGEMENTDifferent readers, same resultThe person can inspect the shared evidence
Public conceptual model; it does not reproduce the internal runtime sequence.
That arrangement helps answer a subtle question: which part of the system is allowed to change when someone changes their role? The explanation can become more concise, more technical or more relevant to an organisation. The underlying result cannot quietly change with it.
This also prevents a tempting misuse of configuration. An organisation might reasonably say that it is particularly concerned about operational reliability or customer experience. That can guide the explanation towards those aspects when the evidence supports them. It must not instruct the system to report that performance has improved, or to bury an inconvenient result.
A role is a communication preference, not a new source of truth
I wanted to avoid implementing a different analytical personality for every job title. That would introduce parallel rule sets and make it harder to explain why two people received different conclusions. Instead, the documented model resolves named roles to communication profiles within a common narrative boundary.
For example, a governance reviewer may want the lineage of a result brought forward. An executive may prefer a shorter account of its possible significance. A data analyst may appreciate more detail about the underlying measures and limitations. In each case, there should still be one place to inspect the evidence.
The same distinction applies across the product. Summary has an orientation purpose; Ask presents the result of a particular question; Insights introduces findings that have already qualified for attention; Reports communicate selected analytical material. Those surfaces need different language. They do not need competing analytical truths.
Testing the invariant, not just the wording
A convincing demonstration would not merely show two differently phrased paragraphs. It would establish that the governing analytical identity and factual content remain invariant while the presentation changes. The development documentation describes a role-preview facility for doing exactly this: examine the same governed material through alternative communication profiles and organisation-context states.
That is a useful internal test boundary. It is not the same as an independently conducted study showing that every kind of business user understands the result better. The latter would need its own research design, participants and evidence.
There are also genuine usability questions. How much detail does a reader need before an explanation becomes overwhelming? When does simplification conceal uncertainty? Can emphasis alter the apparent significance of a finding even if the numbers are unchanged? These questions matter because an explanation can mislead without ever making a numerical error.
What the implementation contributes
The November 2026 codebase contains the narrative context and contract components, while the maintained documentation defines permitted adjustments and prohibited changes. The October research paper describes the broader reason for separating interpretation from analytical truth.
The significant engineering choice is to make contextual presentation a named, constrained responsibility rather than an open-ended permission for AI to decide what matters. It leaves room for more useful explanations while preserving a stable route to evidence and correction.
For me, the continuing question is not whether the same result can be written in two ways. That is easy. It is whether those two explanations can serve different people without becoming two different accounts of the facts.
Connected research
This capability study belongs with the work on separating analytical authority from natural language, maintaining evidence and communicating governed findings in reports. Those boundaries are what make contextual explanation a useful extension rather than a competing authority.
Research context
Research programme: Making data analysis understandable, reproducible and accountable
Topic: Evidence and verification
Explore further
Supporting material is listed below.
Further reading is listed below.