The Personal Information Radar

Priya’s afternoon, as seen from inside the system, begins at 14:40. She is in a design review that runs until 15:30. The Radar, which is what this chapter will call the system that has been assembled over the last seven chapters, the same layer as before now attending to a stream and not a single document, has four candidates waiting to reach her.

A  a new paper that possibly weakens the bounded-delay assumption
B  a major update to a library she uses often
C  an interesting new survey of consensus protocols, mostly material she knows
D  an incident report affecting tonight's deployment

Part I ended with a Semantic Browser, a renderer that treats a source as something to interpret and present. This chapter is the equivalent for attention. It asks: what does the whole attention architecture look like when information is continuously competing to reach one person? The metaphor is chosen with care. A radar watches continuously, signals enter and leave range, some are uncertain, they differ in significance, and not everything becomes an alarm. What the metaphor must not become is surveillance, and the chapter will say what prevents that.

What the Radar is made of

The stages are the ones the part has built:

SOURCES
   ↓
INFORMATION UNITS
   ↓
STANDING SUBSCRIPTIONS AND RELATIONS
   ↓
TASK RELEVANCE
   ↓
PRIOR-KNOWLEDGE CHECK
   ↓
SOURCE AND PRESERVATION REQUIREMENTS
   ↓
QUALIFYING FOR ATTENTION
   ↓
RESOLUTION
   ↓
DELIVERY TIMING
   ↓
READER

One thing about this picture is misleading. It looks like a pipeline that ranks, with each stage nudging a number up or down until a final order emerges. It should not work that way. By now the part has separated several judgements that must stay separate, and a single score would swallow them. So the Radar does not compute a priority. It keeps a record for each candidate:

candidate
  relation               what it is to something she cares about
  task_relevance         how it bears on what she is doing
  source_necessity       what must travel with it to stay faithful
  novelty                whether it is new to her
  evidence_standing      how far the relation is established
  urgency                how costly delay is
  interruption_cost      how costly surfacing now is
  resolution             how much to show first
  delivery               now, queue, digest or available
  reasons                why each of the above

A single score is attractive because it makes ordering easy. It is dangerous because unlike things vanish inside it. The part has been teaching four distinctions that a score would erase:

high urgency    ≠  high confidence
high relevance  ≠  high novelty
high novelty    ≠  safe to interrupt
low novelty     ≠  safe to omit

An urgent item with weak evidence, a relevant item she already knows, a new item that has no reason to take the floor, a familiar item a later unit depends on: each is a different case, and each needs a different answer. The record keeps them apart.

Four candidates, four records

Here is what the Radar does with each, at 14:40. The candidates, the records and every value in them are constructed for illustration.

A  new paper, possible weakening of the bounded-delay assumption
   relation: possible contradiction     standing subscription: match
   task relevance: high                 novelty: high
   evidence standing: medium (preprint, unreviewed)
   urgency: medium (deployment at 18:00)
   interruption cost now: high (presenting)
   resolution: brief
   delivery: queue for 15:35

B  major library update
   qualifies: yes, generally relevant to her work
   task relevance: low for the current deployment
   delivery: digest

C  new survey, mostly familiar
   novelty: low overall; one unit is new (a classification of delay models)
   resolution: glance, with the new unit named
   delivery: available

D  incident report affecting tonight's deployment
   task relevance: high      urgency: high      evidence standing: high
   cost of delay: greater than the cost of interrupting
   delivery: now

Candidate C shows the prior-knowledge rule at work. The survey is mostly something Priya knows, so most of it is not worth her attention, but one section is new, and a glance that mentioned only the familiar parts would have hidden it. Low novelty did not make the whole item safe to drop. Candidate B shows qualification and relevance coming apart: it clearly matters to her work in general and does not bear on what she is doing today, so it waits for the digest. Neither interrupts.

When two items compete

At 14:40 both A and D deserve attention. A system built on a score would say that D scores 0.91 and A scores 0.84, and leave it there. The Radar says why.

D interrupts because:
- it affects tonight's deployment
- the evidence is direct
- delay could make remediation impossible

A waits because:
- it is potentially important
- the relation is still unconfirmed
- there is no material cost in waiting 55 minutes

That makes the allocation of her attention answerable. She can see why D came through and A did not, disagree with either decision, and correct it. At 14:41 a short, plain interruption about the incident reaches her. At 15:35, when she leaves the review, the brief about the paper is the first thing she sees, labelled as a candidate with its standing. The digest she opens the next morning holds the library update, and the survey sits in her list, glanceable and one step from the full text.

What the Radar has not seen

A system like this creates a risk that no individual stage does, and it needs to be stated plainly. When something does not reach Priya, the reasons can be very different:

not surfaced may mean:
- judged irrelevant
- judged already known
- deferred
- withheld
- outside the monitored sources
- never retrieved
- never observed

The last three have nothing to do with judgement at all. The paper that never appeared may not have been in any source the Radar watches. So:

not selected  ≠  does not exist
radar coverage  ≠  the world

The Radar can only reason over what it can see, and the danger is an illusion that a sophisticated filter makes absence meaningful, that silence from the system means there is nothing to hear. A reader should be able to ask, on demand, what sources were searched, which subscriptions were active, what was filtered and why, what was deferred, and what the system never had access to. Not all the time, but whenever she wants to check. The result is a very different kind of tool from a feed. A feed shows what it has. A Radar that is honest about its coverage says what it was looking at, and therefore what it could not have found.

A correction that changes policy, not the person

An hour later Priya tells the Radar: “Don’t interrupt me for library releases unless they affect an active deployment.” This is a correction, and what the system does with it matters.

It updates a declared rule. Library releases no longer interrupt unless they touch something live. It does not conclude that Priya dislikes software-release information, or that she is the kind of person who ignores updates. It makes no inference about her at all.

correction  →  policy refinement
correction  →  not a conclusion about the person

The rule is visible on her list of what the Radar is doing, she can change it, and it is scoped to exactly what she said. Because each stage of the Radar keeps its own judgement, a correction can be pointed at the stage that was wrong: this was not relevant, this was not new, this should not have interrupted. A single thumbs-down on a score could not tell the system which of those she meant.

Adaptation without an unrestricted model

A broader principle has been emerging through this part, and the Radar makes it visible. At each step where the system might have built a large picture of the person, the chapters chose a narrow object instead:

knowledge state  ≠  dossier
explicit standing interest  ≠  inferred identity
observable context  ≠  inferred inner state

The Radar knows which concepts she appears to know and with what evidence. It knows which relations she has asked it to watch. It knows her declared modes and her calendar. It does not know who she is, what she wants from her life, or how she is feeling. And it works anyway. That is the part’s most important finding about adaptation: adaptation does not require an unrestricted model of the person. The Radar is defined by the sources and standing interests she has chosen, not by an unlimited attempt to model her.

What the architecture has not shown

This is an architecture, not a result. The pieces have been argued for individually, and the Radar has been shown working on one afternoon, in a constructed example with the answer in mind. Nothing here measures whether the composition behaves better than a simple chronological feed, whether the stages interfere with one another, or whether the decision records stay honest once many more candidates arrive. Parts that work alone can fail in company. The sort of test that would settle it is one over a fixed set of sources with a known set of things that mattered, so that what the Radar missed can be counted, and any such test is a closed world: it says nothing about an open one, where sources change, evidence goes stale and some of what arrives is hostile. The book offers the design with those limits visible.

Where this leaves us

Part II has taken a flood of cheap representations and put a series of decisions between them and the reader. What is the unit? What matters for this task? What is new to this person? What persistent relation is she watching for? How much of it, and when? Each decision is part of communication now, and each is a separate judgement kept in a record she can read. The channel no longer simply carries a message. It takes part in deciding what becomes a message, for whom, at what resolution and at what moment. That is intelligence entering the channel in a much deeper sense than choosing how a page looks.

Look at what the Radar knows about Priya by the end of the afternoon: what she knows, what she is watching for, what she is doing and what she has asked. And look at how much it has been doing for her. Yet it still waits. It chooses what she sees and when, and then leaves her to act. The next temptation is obvious. If the system knows what matters, what she already knows, what she is watching for and what situation she is in, why should it merely show her things? Why not remember how she responded, adjust what it does next, and change what happens as well as what she sees?

The Radar still waits for Priya to act. The next step is what happens when the channel starts changing not only what she sees, but what it remembers about the relationship between them.