On this page
I have become increasingly wary of systems that are expected to say something interesting every time they are asked to examine data.
The pressure to fill the screen
There is an understandable expectation that an intelligence product should discover something. A dataset is uploaded, the system works for a while, and eventually a set of colourful cards appears. The user may reasonably assume that the cards represent the most important things in the data.
But the presence of a card is not evidence of importance. A technically correct calculation might be routine. A difference may be real but negligible. Several apparently separate findings may be different descriptions of the same underlying movement. An observation can also be interesting to the system’s author without being useful to the person responsible for the business.
A product designed to always deliver a minimum number of insights is at risk of confusing the need to populate an interface with the existence of something worth a person’s attention.
The distinction I needed to make
As I developed determinIQ, I found it useful to distinguish several questions that are easily collapsed together.
Can the system perform the calculation? Has it obtained a result? Is the result supported by evidence? Does it qualify as a finding? Is it materially significant within the available data? Does it add something distinct from findings already presented? Should it be brought to the user’s attention now?
These questions belong at different stages. A correct answer to the first does not decide the last. The architecture therefore separates governed execution from finding qualification and attention. That decision allows an analytical result to remain valid even when it is not promoted into a prominent Insight.
Research figure 01
Not every valid result needs attention
A simplified distinction between calculation, finding and promotion
- SUPPORTED OPERATIONCalculate a resultThe governed engine establishes the analytical value
- EVIDENCEVerify the basisThe result can be traced and challenged
- QUALIFICATIONAssess the findingGoverned criteria consider whether it merits promotion
- ATTENTIONDetermine presentation membershipCertified attention rules control what is surfaced
- HUMAN INVESTIGATIONInspect, ask or exploreA person retains the choice of what to examine next
The actual qualification and attention contracts are more detailed; no numerical threshold is implied.
I like the consequence of this separation: zero surfaced Insights can be a legitimate outcome. It may mean that the analysis completed and nothing qualified for attention under the current governed policy. That is quite different from an error, a timeout or a failure to investigate the dataset.
Attention is not the same as truth
A user who wants a concise overview has a different immediate need from someone exploring a dataset in detail. determinIQ has used Focused, Balanced and Exploratory modes to describe these different attention postures. But the presentation preference is not permission to rewrite the underlying analytical evidence.
The November 2026 architecture records a particularly useful refinement. The Balanced attention pathway was constrained to the qualification and salience authorities that could actually be supported. A statistical lower-tail device remained available for compatibility and diagnosis, but was not treated as a new authority for declaring findings unworthy of attention. The more general lesson is that a numerical threshold does not automatically establish a business decision objective.
That is a healthy limitation. A system should not claim to know what an organisation ought to care about merely because a computed distribution provides a convenient cutoff.
What the interface should say
There is also a human-computer interaction problem here. An empty Insights screen can look broken even when the absence of findings is correct. A useful interface needs to distinguish an investigation that is still running, an investigation that has failed and one that completed without promoting a finding.
The source documentation for Summary and governed discovery records work on precisely this kind of continuity: preserving the user’s location, waiting for a real presentation artefact rather than inventing progress, and explaining when processing is taking longer than expected. These are not replacements for analytical governance. They help the user understand what the system has, and has not, established.
If nothing qualifies, the product can say so plainly. It can still offer routes into Ask or Explore, where a person may investigate a specific question without pretending that the autonomous discovery process found a significant event.
The limits of a useful silence
I would not claim that returning zero Insights proves that nothing important exists in the data. The system has a defined analytical scope, a particular set of supported operations and governed criteria. Source quality, field meaning, comparison eligibility and the maturity of the implemented analytical families all affect what can be established.
The correct statement is narrower: within the analysis carried out and the qualification rules applied, the system has not found a result that warrants promotion. That leaves room for a human to ask a different question, inspect the source material or challenge the rules themselves.
This is perhaps the aspect I find most interesting. A useful analytical system should reduce unnecessary noise while remaining open to scrutiny. It should not manufacture certainty simply to reward the user for waiting.
What I take from it
The design of determinIQ began with a question about trusting AI-generated analytical answers. The work on Insights introduced a related question: even if a result is correct, why should I be paying attention to it?
I do not think the answer should be “because the system had a card to fill”. Sometimes the most responsible outcome is to show what has been established, explain the limits and leave space for the next question.
Research context
Research programme: Making data analysis understandable, reproducible and accountable
Topic: Governed analytics
Explore further
Supporting material is listed below.
Further reading is listed below.