A way for an agent to work out what a finding means for the person in front of it, decide how much attention that deserves, and then spend no more than that.
This paper takes one idea from an interaction-pattern page and works it through properly: where the idea came from, what already exists that looks like it, what is actually new in it, how it would work as a system, and where it breaks.
Read the story 10 screensExplore the full paper 50 pages
Every contract, name, date and number in the examples is invented for illustration. Every claim about other people's work is marked verified with a source at the end, or marked proposal when it is my own untested idea.
Your wife calls at two in the afternoon. You might answer, or call back later. She calls at three in the morning. You answer at once, half expecting something serious.
Nothing on the phone changed. Same name, same photo, same ring. What changed is everything around the call: the hour, what you expect of her, what you know about her week, what else could be going on. You did not add those up. You read them together, in an instant, and the call meant something different.
This is not a claim that people are predictable. Someone else, on another night, lets the 3 a.m. call ring out. The point is smaller: the meaning of a signal is not in the signal. It is in the situation the signal lands in, and that situation is a web of things that only make sense together. Three in the morning is not urgent on its own. It is urgent because of who is calling and what you know about them.
Agents are now in the position of the phone. They read contracts overnight, connect what they find across amendments and invoices, and reach conclusions nobody asked for. Each conclusion lands in someone's situation. The agent does not know that situation the way you know your wife. It can only see what has been written down. But it still has to decide: say something now, leave a note, or stay quiet.
The rest of the paper takes that question into contract software, where the signals are payment terms and notice windows rather than phone calls. It asks what an agent would need to read a situation well enough to choose its response, and how that response should then appear on the page.
It is 20 May. While the team slept, an agent read every contract in a 640-agreement portfolio, checked each one against invoices, usage data and deadlines, and wrote down what it found. By morning there are 200 conclusions waiting.
Each dot below is one of them. Switch between the two views and watch where they go. Hover or tap any dot to read it.
In the feed, every conclusion got the same card. Nothing was wrong with any of them; the agent did good work. But the reader now has to open 200 cards to find the one where silence will cost them a quarter's figures. If each card takes 20 seconds to read and judge, that is over an hour before they reach it, assuming they reach it at all.
Sorted by Sensitivity, the same 200 conclusions become one thing that stops the reader, three that stand out, eight worth a line, twenty-four quiet notes, and 164 that say nothing because there is nothing to say. Not one conclusion was deleted. Each one was given the amount of attention that missing it would cost.
There is a catch in the second picture. "What missing it would cost" sounds like a number the agent can read off the contract. Often it cannot. The same conclusion costs a lot to miss for one reader this week and nothing for another, because of what is going on around it: who owns it, what is about to happen, what was decided somewhere the agent cannot see. Sorting is the visible half of the job. Reading the situation is the half that comes first.
Nobody asked. A search result lands in a question somebody typed. An agent's conclusion lands in nobody's question. It arrives at 3 a.m., unasked, and the agent has to judge for itself whether the person it concerns needs it today. That judgment used to be made by the person, in the act of looking. Now it has to be made before anyone looks.
The rest of the paper explains both halves: how the agent reads a situation, the five levels, the four questions that start the costing, where the explanation goes when the reader asks "why?", and where the whole idea breaks.
The next screens take the two halves in turn: where the problem came from, how an agent reads a situation, and how that reading becomes something on the page.
In the full paper
How to read this paper · The whole paper in one page · Why this matters now
It is written so that someone who has never designed software can follow it. A few words are used in a specific way throughout. Here they are, once, so you never have to guess.
Software that does work on its own in the background, like reading all the contracts overnight and preparing a short list of things that need a person. In this paper, "the agent" is always that kind of software, never a human.
A point where the screen has something new to tell you. A deadline got closer. A value was read from a document. The agent finished a task. Every moment makes the interface choose how to respond.
Everything around a moment that changes what it means: the state of the work, who owns it, what is about to happen, what was decided somewhere else. The agent can only see the part of this that is written down.
The reader's noticing. It is limited and does not grow when the software gets smarter. Sensitivity treats attention as a budget.
How much attention a moment is allowed to take, on a scale of five steps from Nothing to Interrupt. The reader never sees the word "level". They only feel that some things are easy to ignore and some are not.
Where and how the new information appears on screen. There are four kinds in this paper: Preview, Transform, Expand and Open. Each is explained with a working example you can click.
Where a piece of information came from: read from a document, calculated from other facts, written by an AI model, or confirmed by a person.
A small working example of one moment. The specimens in this paper are live: hover, click and drag them. Boxes marked live respond. Boxes marked static are pictures.
Three labels appear on claims. verified means I checked the source and it says what I say it says. proposal means this is my own idea and has not been tested with users. example data means the numbers exist only to make the example concrete.
Software is getting very good at producing more. More answers, more recommendations, more explanations. People did not get more attention. So the design question changed.
The old question was how much can the screen show? The new one is how much attention does this moment deserve? Sometimes the answer is nothing. Sometimes a small hint. Sometimes a full stop, but only when missing it would cost something real.
Sensitivity turns that rule into four decisions, made at every moment, in this order:
Parts of this exist already. Progressive disclosure, calm technology, phone notification levels and research on the cost of interruptions all overlap with it, and the paper names each one and says exactly what is borrowed. What Sensitivity adds is narrower than the concept page first claimed, and more defensible: one rule that covers both the reader asking for more and the system pushing more, applied inside the work itself, deciding where and when as well as whether, with the software's confidence kept separate from the cost of being wrong, and with the agent's reading of the situation shown on the page so a person can correct it.
The last part of the paper lists where the idea fails and what has not been tested. That part matters as much as the rest.
Sensitivity did not start as a theory. It started as a problem on one screen: a contract page that was redesigned so that the software does the reading and the person only sees what needs them.
A contract management system stores agreements between companies. Think of a master agreement with a supplier, plus the smaller documents that hang off it: statements of work, work orders, amendments. Each one has dates, money, obligations and deadlines inside it.
For years, the page for one contract looked like a filing cabinet: a tree of related documents on one side, typed-in fields on the other, and a stack of collapsed sections for the clauses. Everything was present, and everything was equally quiet. Nothing on it is wrong. It answers what is in the contract and leaves what should I do today entirely to the reader.
To find out whether anything needed doing, a person had to open each tab, read each row, remember the dates, and work out the consequences in their head. The software stored. The person thought.
Before this page, the contract page was a filing cabinet: a header, forty fields, a list of documents, and all the reading left to the person. Nothing on it said what mattered. The redesign turned that around.
The redesign turned the page around. An agent reads every document in the family in the background. The page then opens with the few things that need a person today, written as decisions rather than data.
Read the three decision lines in the drawing. Each one is a complete thought: what to do, by when, and what it is worth. Behind each line, the agent did the reading that used to be the person's job. The page says less and means more.
That is where the trouble starts.
The trouble is the "why". Every line the agent writes is a conclusion, and the reader will want its reason before acting on it. The full paper shows what happens on the screen when they ask: the reason costs space, and space pushes the most dangerous item off the bottom. This story goes a different way, to the question the reasons raised. Of everything the agent could say, what does this person need today?
In the full paper
The contract page, before · The moment the reader asks why · Four things pull against each other
A person will act on a one-line recommendation from software only if they trust it. And they trust it only after they have checked it at least once. So the first thing a reader does with a compressed line is try to unpack it: why this? where did that number come from? what changed?
Click the line below once, twice, three times. Watch the word count.
This is the compression problem. The better the agent gets, the more it can say in one line, and the more each line hides. The screen does not get bigger. The reasoning behind every line does.
Once readers start asking "why?", four forces meet on the same screen.
Here is the tension in one gesture. Open the "why" panel on the first decision and watch what happens to the red item below it.
That is the whole problem in one picture. Most interfaces only know how to spend space. When space runs out, the only other budget is time: show one thing, then let it become the next thing, in the same place.
Screens have always been too small for everything. Designers have always had to choose. What is different about software built around agents?
Agents hand people conclusions. Old software handed people data and let them conclude. A field that says "Net 45" needs no explanation; you read it and you know. A line that says "counter at +6%" is a conclusion, and every conclusion owes an explanation. It arrives with a debt attached.
The debt is paid at a different time from the conclusion. The agent finished reading at 3 a.m. The reader checks the reasoning at 10 a.m., on their own schedule, one item at a time. The interface cannot know in advance which items will be checked, so it cannot pay all the debts up front without filling the screen.
Some conclusions are wrong. A model that reports 0.81 confidence should be wrong roughly one time in five, if that number is well calibrated. The reader must be able to find those cases. Hiding the reasoning saves space but makes the wrong cases invisible.
Not every conclusion matters equally. One line may be worth $134K and have three weeks to run. Another may be worth nothing in money but become binding in two days. If the interface treats them the same, the reader has to re-rank them in their head, which is the work the agent was supposed to remove.
Nobody asked. A search result lands in a question somebody typed. An agent's conclusion lands in nobody's question. It arrives at 3 a.m., unasked, and the agent has to judge for itself whether the person it concerns needs it today. That judgment used to be made by the person, in the act of looking. Now it has to be made before anyone looks.
So the question for a screen full of agent output is no longer "where do I put all this?" It is: for this moment, how much attention is right, and what is the cheapest way to spend it? Sensitivity is one answer to that question. The next part looks at the answers that already existed.
Nobody is the first to notice that attention is limited. Before claiming anything new, this part names the ideas that got there first, says what each does, and says where each stops.
What it is. Show people the few most important things first, and offer the rest when they ask. The usability researcher Jakob Nielsen described it in a 2006 article: it "defers advanced or rarely used features to a secondary screen, making applications easier to learn and less error-prone." He adds that the idea is already more than thirty years old. verified
A close relative is a sentence from Ben Shneiderman's 1996 paper on visualisation, often called his mantra: overview first, zoom and filter, then details on demand. verified
It keeps the first view clean and gives the reader control. Every "Show details" button, every "More" link, every expandable row is progressive disclosure, and it works.
Sensitivity uses progressive disclosure as one of its tools. The renewal example in Part 1 (headline, then short reason, then full reasoning) is progressive disclosure. What Sensitivity adds is the rule for how far to go and when to go there without being asked.
What it is. In 1996, Mark Weiser and John Seely Brown at Xerox PARC wrote a short paper called "The Coming Age of Calm Technology". Its idea: technology should live mostly at the periphery of our attention, "what we are attuned to without attending to explicitly", and move to the centre only when it matters. "A calm technology will move easily from the periphery of our attention, to the center, and back." verified
Their best-known example is the Dangling String, an installation by the artist Natalie Jeremijenko: an eight-foot plastic tube hung from a motor wired to the office network. Heavy network traffic made the string twitch and whir; quiet periods barely moved it. Nobody read a screen. People sensed the network from the corner of their eye. verified
In 2015, Amber Case turned this into a set of principles in her book Calm Technology. The first is: "Technology should require the smallest possible amount of attention." Another: "The right amount of technology is the minimum needed to solve the problem." verified
The fairest description of Sensitivity in these terms: calm technology made usable inside a working screen, with named steps, a test for when to climb them, and specific ways to place the new thing.
Two more pieces of existing work come even closer to the "how much attention" question.
Since iOS 15 (announced at WWDC 2021), an app sending a notification chooses one of four levels. verified
| Level | Sound | Breaks through Focus modes | Bypasses silent switch |
|---|---|---|---|
| Passive | No | No | No |
| Active (default) | Yes | No | No |
| Time Sensitive | Yes | Yes, if the user allows | No |
| Critical | Yes | Yes | Yes (needs Apple's approval) |
This is a ladder of how hard something may interrupt you, four steps against Sensitivity's five. The difference is scope. Apple's levels apply to messages arriving from outside the app you are using. They say nothing about a value inside a document, a row in a list, or a button you are about to press. They also control only loudness, not placement.
In 1999, Eric Horvitz, Andy Jacobs and David Hovel at Microsoft Research published "Attention-Sensitive Alerting". Their system, called Priorities, looked at incoming email and decided whether to alert the user now or later. The abstract describes procedures that "balance the context-sensitive costs of deferring alerts with the cost of interruption." verified
Read that sentence slowly. The cost of deferring is what you lose by seeing something later rather than now. That is almost exactly what this paper calls the cost of missing. The idea of weighing it against the cost of being interrupted is more than twenty-five years old.
Sensitivity's rule is "spend no more attention than the moment deserves". Is overspending actually harmful, or just untidy? Three studies say it is harmful, in measurable ways.
Ancker and colleagues (2017) studied clinical alerts in a hospital's electronic records. They found that "likelihood of reminder acceptance dropped by 30% for each additional reminder received per encounter". Roughly a quarter of drug alerts and a third of clinical reminders were duplicates for the same patient within a year. verified
In plain terms: every extra alert made the next one less likely to be acted on. Doctors were not careless. The system had trained them to click past it. This is the cost of a level set too high too often.
Gloria Mark, Daniela Gudith and Ulrich Klocke (2008) measured people doing an email task with and without interruptions. The interrupted groups finished slightly faster (about 20 minutes against about 23) but reported higher stress (9.46 against 6.92 on their scale), higher frustration, more time pressure and more effort. People compensate for interruptions by working faster, and pay for it. verified
Nelson Cowan (2010) summarised decades of memory research: working memory holds about "3 to 5 meaningful items in young adults". verified A screen that asks the reader to remember a number from one panel to understand another is spending one of those three to five slots. This is why the Expand expression in Part 3 exists: when two facts must be compared, keep both on screen so the reader does not have to carry one.
Put side by side, the overlap is large. That is fine. The point of this page is to be exact about it.
| Existing idea | What it already does | What it does not do | What Sensitivity adds |
|---|---|---|---|
| Progressive disclosure | Shows the basics, reveals more on request | Same split every time; only the reader pulls; only adds below | The split moves with the moment; the system can push; a choice of replace, beside, or over |
| Calm technology | Keep most things at the edge of attention; use the least attention | Two zones; no rule for when to move; written for rooms | Five steps; a cost-of-missing test; screen-level expressions |
| Apple interruption levels | A four-step ladder of how hard a notification may interrupt | Only for notifications from outside the app; loudness only | Works inside the page, on a value or a row; also decides placement |
| Cost-of-interruption research | Weighs the cost of deferring against the cost of interrupting | Decides whether and when, not where or how | Where and how; confidence kept separate from cost |
| Context-aware computing verified | Dey (2001) defines context as any information that characterises the situation of an entity, and a context-aware system as one that uses it to give relevant information or services | Describes the user's situation; does not say how a system should judge what a finding means in it, or how that judgment should appear | A named set of situation inputs for agent findings, and a visible link from the judgment to the page |
| Mixed-initiative interaction verified | Horvitz (1999): act only when the expected value beats staying quiet; weigh the user's attention in the timing of a service; allow for uncertainty about the user's goals | Decides whether and when from a utility model; little on where the result goes, how it is worded, or how the person corrects the agent's reading of the situation | The judgment carried through to placement and wording; a one-step correction when the reading is wrong |
| Escalating reminders | A reminder that gets firmer as a deadline nears (calendars, bills, service desks) | Only gets louder; only for deadlines | The same climb applied to any kind of moment, with the option to come back down |
That is a smaller claim than "a new way of thinking about interfaces". It is also one that can be defended line by line, which is what the next part tries to do.
First, a reading: what does this finding mean for this person, in the situation the agent can see? Then three questions the interface answers, in order. How much attention? Where does the new thing go? Where did it come from? Each of the three has a working specimen. The reading is what they all start from.
The three axes that follow (how much attention, where it goes, where it came from) all start from one input: what missing this moment would cost. This page is about where that input comes from. It does not come from the finding. It comes from reading the finding against the situation.
An agent that reads a contract correctly has done the first job. It knows the payment term is Net 45, that it was Net 60 until an amendment last month, and which page says so. The second job is to work out what that means for the person who will see it, today. These are different jobs, and a system can be good at the first and bad at the second.
What can an agent read about a situation? Only what is recorded, but that is more than the contract. Seven kinds of thing, offered as candidates and not as a score. proposal
The finding and its evidence. Net 45, from Amendment 2, section 3. The state of the work. Is this contract in renewal, in dispute, being wound down, or simply running? What is about to happen. A payment run, a signature, a notice window closing, recorded in a calendar, a workflow or a task someone opened. What missing or delaying it would cost. Money, a lost right, a deemed acceptance. Who already owns it. A task assigned, a reply drafted, a colleague who opened the clause yesterday. Whether it can be undone. A draft can; a sent notice cannot. What is not available. The gaps the agent knows it has.
None of these means anything alone. That is the lesson of the 3 a.m. call, and it is why this is not a checklist to be ticked and summed. A payment run next week is a fact. Net 45 is a fact. Put together, they say an invoice may go out on the wrong term. Add that someone in finance corrected the run yesterday, and the same two facts say nothing. Add instead that the run is for $412,000, and they say a great deal. The judgment lives in the relationships between the inputs, not in any one of them, and the relationships move as the situation does.
The output of this judgment is not a yes or a no. It is a reading: this change matters to this person because of that payment, and it matters today. The level, the expression and the wording all follow from that reading. That is why the rest of Part 3 can treat "the cost of missing" as a given: it is a given once the situation has been read. The reading can be wrong, and the page on confidence later in this partscreen 8 says what to do about that. First, the same finding read twice.
Judging whether and when to act is not new. Context-aware computing and mixed-initiative interaction have worked on it since the 1990s, and the full paper lines each older idea up against this one and says exactly what is borrowed.
In the full paper
What already exists, Part 2 · What is borrowed, and what is added · Attention is the budget
A worked example, invented. example data At 9 a.m. the agent's reading of a supplier contract says the payment term is Net 60. At 11 a.m. an amendment that was signed last month is finally uploaded. The agent reads it: the term is Net 45, from the first of this month. Same contract, same agent, better evidence. What should it do with that?
The next invoice on this contract is due in three months. Nobody has the contract open. No task mentions it.
The agent updates the term, keeps the Net 60 reading and the amendment side by side in the history, and marks the change for whoever opens the contract next. Level: Nothing now, a Hint when someone next looks. An interruption here would be noise with a good excuse.
Finance has a payment run scheduled for 12:00 today, and it was built on Net 60. The run is in a system the agent can read.
Now the same finding sits one hour from a wrong payment. The agent brings it to the person who owns the run, with the amendment attached: Interrupt, carried outside the page if they are not in it. After noon it drops back to a record of what was found and who was told.
The finding did not change between the two columns. Its evidence did not change. What changed is what sits next to it: an action, a time, an owner, and the money in play. The relationships between those moved the level from Nothing to Interrupt. Take any one away and it moves back.
Part 4 shows an item moving as a deadline nears, day by day. This page is the other kind of movement. Nothing climbed here. New evidence arrived at 11, and a scheduled action at 12 gave it its weight.
Think of the reader's attention as money. The interface spends it every time it shows something. There are two currencies.
Most interfaces only know how to spend space. The "why" panel in Part 1 spent space and pushed the deadline off the screen. A tooltip spends a little space for a little time. A full-screen dialog spends all the space and all the time at once.
Two things follow from treating attention as a budget.
The rest of this part is the budgeting method: how much to spend (the level), which currency to spend it in (the expression), and how much the reader should trust what they bought (the source).
The first question is how much attention is this moment allowed to take? The answer is one of five levels. The reader never sees the names. Pick a level below and see what the reader actually sees, using the same item each time.
| Level | What the reader experiences | What it costs them |
|---|---|---|
| Nothing | No change. The item is wherever it normally lives. | Nothing |
| Hint | A small, quiet mark that something is there. Easy to pass by. | A glance, if they choose |
| Reveal | The fact is shown in words, in place. Still easy to skip. | A line of reading |
| Emphasise | The item moves up, or stands out in colour or weight. Hard to miss while on the page. | Their eye is pulled |
| Interrupt | It stops them. It appears first and asks for a choice before they continue. | Their task is paused |
The level sets a ceiling. It does not design the interface. Two different moments at level Reveal can look completely different; what they share is that neither is allowed to pull the eye or stop the reader.
What sets the level. Four questions the contract can answer give a first reading: how much money is at stake, how many days are left, whether it can be undone, and whether silence counts as a decision. They are inputs to the judgment on screen 4, not the judgment. Two renewals show why. Supplier A's window closes in two days, Supplier B's in ten, so by days left A is louder. But A is being wound down, a decision recorded on the contract, and B is the one the business wants to renegotiate, which has to start today. Money, days, undo and silence all point at A. The situation points at B. The consequence is not always a property of the information. Often it is a property of the situation the information lands in.
One thing the level is never set by: how sure the agent is. Consequence sets how loud. Confidence sets the wording, and whether the agent may act alone. Screen 8 comes back to that.
In the full paper
What sets the level · Attention is the budget · Four ways to place the new thing · Where did it come from
The level is set by what it would cost this reader, in this situation, to miss this moment. Not by how certain the software is that it found something. Not by how new the information is. Only by the cost of missing, read the way the two pages before this one read it.
How does software measure "cost of missing"? It cannot read minds, but it can read facts that are already in the data. Four questions give a first answer in most cases. They are inputs to the judgment, not the judgment itself. proposal
| Question | Read from | Effect on level |
|---|---|---|
| How much money is at stake? | Contract value, the delta in a renewal, the size of a credit | More money, higher |
| How many days are left? | A deadline or notice window in the document | Fewer days, higher; the climb is steep in the last week |
| Can it be undone? | Whether the action has an undo, a cooling-off period, or a legal lock | Irreversible, much higher |
| Does silence count as a decision? | A deemed-acceptance or auto-renew clause | If yes, the ceiling is Interrupt as the deadline nears |
Run the three Northstar items through it, with today set to 8 May. example data
| Item | Money | Days left | Undoable? | Silence decides? | Level today |
|---|---|---|---|---|---|
| Counter the renewal at +6% | ~$134K/yr | 24 (window closes 1 Jun) | Yes, until the window closes | Yes, it auto-renews | Reveal |
| Sign or annotate the Q2 review | None directly | 14 | No | Yes, silence is acceptance | Hint today, Interrupt by 20 May |
| File the West Region SLA credit | ~$18K, one time | 20 | Not after 28 May | No, it simply lapses | Hint |
Notice what this does to the redrawn page in Part 1. There, the red item sat second. By this rule it sits second today and first on 20 May, because the cost of missing it climbs as the deadline nears while the renewal's cost stays flat. The order of the list is not fixed. It follows the cost.
Two renewals, invented. example data Supplier A's notice window closes in two days; Supplier B's in ten. By days left, A is louder. But A is being wound down: procurement decided last month to let it lapse, and the decision is recorded on the contract. B is the one that matters, because the business wants to renegotiate, and a renegotiation that must be agreed in ten days has to start today.
Money, days, undo and silence all point at A. The situation points at B. A deadline alone cannot set urgency, because a deadline is only a cost while someone still intends to act on it. The four questions find the costs that are written in the contract. The judgment pages before this one are where those costs are read against the work going on around the contract, and that reading is what this page receives as its input. The consequence is not always a property of the information. Often it is a property of the situation the information lands in.
Two different numbers are easy to confuse. Confidence is how sure the software is that it read or reasoned correctly. Consequence is how much it would cost to miss or get wrong. Most systems blend them: lower confidence means more warnings. Sensitivity keeps them apart, and gives each a different job.
Move the two sliders and watch the same payment-terms line change.
Why keep them apart? Because they fail in opposite ways. If low confidence raised the level, a system that is unsure about many small things would shout constantly, and the reader would learn to ignore it. If high confidence lowered the level, a system that is sure about a $2M liability cap would whisper it. Being sure and being important are different questions, and so are being sure and being allowed to act.
There is a second kind of confidence hiding inside the first, and it is worth separating without turning it into a third dial. The slider above measures confidence in the finding: did the agent read Net 45 correctly? There is also confidence in the reading of the situation: given everything the agent can see, is this the moment that finding matters, and to this person? An agent can be at 0.98 on the first and quietly wrong on the second, because the second depends on what it cannot see.
The two renewals on the previous pagefrom screen 6 are the plain case. The agent read both notice windows perfectly. It did not know that Supplier A was being let go, if that decision had only been spoken and never recorded. Its finding was right. Its judgment about who needed it was wrong, and no amount of re-reading the contract would fix that.
What follows for the design is modest. The agent shows what its reading rested on, in a line, not a dashboard: "Raised because a payment run is scheduled for 12:00." Where the reading leans on an assumption, the wording says so, the way the slider above turns "is" into "appears to be". And the person can correct the reading in one step: "Already handled", "We are leaving this supplier", "Not my contract". The correction changes the record and the next judgment on this item. It does not retrain a model, and the design should not promise that it does.
Three ways this fails. When the context was never written down: a decision made in a hallway never reaches the agent, so it raises an item that is already dead and stays quiet about one that is alive. When the agent guesses what you are doing: you opened the contract, so you must be about to act. That inference is often wrong, and an Interrupt built on it is an ambush. And when it is right on the data and wrong for the person: the person already knows, has decided, or is the wrong person. Being right about the contract does not make an interruption right.
In the full paper
Where Sensitivity fails · Preview · Transform · Expand · Open
The second question is where does the new information go, relative to what is already there? There are four answers. Two spend time, two spend space.
The test for choosing between time and space is one question:
Two things to be clear about.
The next four pages show one working specimen of each.
The agent read "Europe" in a request and had to decide what it meant. Hover over (or tap) the word it chose. The other readings appear beside it. Move away and they leave, and the line is exactly as it was.
Two things at once. The original keeps its slot: "Europe" is still the word in the line, and the 23 contracts are still listed. And the preview borrows the reading moment: while it is open you are reading the alternatives, not the line. That second part is why it counts as spending time, even though it technically added pixels. The pixels are returned the instant you look away.
Hover is an input, not a behaviour. The same Preview opens on keyboard focus, on tap, on a long press. On a phone there is no hover at all, so the trigger must be a tap. A Preview that only exists on mouse hover fails every touch user and every screen-reader user.
When the reader might want to check, but probably will not need to. Most of the time "Europe" is fine. The alternatives cost nothing until someone doubts it, and nothing after. This is the cheapest way to pay the "why" debt from Part 1.
Click the value. It becomes the sentence that explains it. Click again and it becomes the clause it came from. Click once more and it returns to the value. Nothing moves on the page except the words inside this one slot.
The old state is given up. While you read the clause, "Net 45" is not on screen. If you need to compare the clause with the value, you have to remember the value. That is the cost of spending time instead of space. In return, the page did not grow by a single pixel and nothing below it moved.
Tabs and segmented controls also replace content in place, so the mechanism is not new. What is narrower here: there is no persistent control and no sibling panels. The object carries its own next level of meaning. A tab switches between equals. A Transform replaces a thing with its own explanation, and the explanation with its own evidence.
Once you have read "Net 45" and want to know why, the number has done its job. It can step aside. The next page shows the case where it cannot.
The reader asked the agent to share a draft with Alex. There are two Alexes. To pick the right one, the reader has to see both side by side, with enough about each to tell them apart. Hiding one behind the other would invite the wrong share. Open the item.
The question stays. Both answers arrive below it, in the same card, and all three are readable at once. The reader can look from "Design · edited this draft twice" to "Legal · signs off contracts" and decide without remembering anything. That is the whole purpose: comparison without memory.
Space. Everything below this card moved down. That is acceptable here because a wrong share sends a draft to the wrong person, and it cannot be unread. It would not be acceptable for a field that was merely read correctly.
The common pattern is autocomplete: type "Alex", the first match is highlighted, press Enter. It makes the choice for you and shows nothing to compare. A Transform would be just as bad here: showing one Alex, then the other, makes the reader remember the first while reading the second. When the decision is between two things, both belong on screen.
Sometimes the new thing is too long or too different to sit beside the old. The full text of an amendment is like that. Open it.
The new content has its own focus (keyboard goes to it), its own scroll (a long amendment scrolls inside the sheet), and its own exit (Close, Escape, or clicking outside). The page underneath does not change at all. When the sheet closes, the reader is exactly where they were.
Long content opens over the page, centred, and nothing below shifts or reflows. This is why Open exists as a separate expression from Expand. An Expand that pushed 400 words of clause text into the list would destroy the list. An Open leaves the list intact and borrows the whole screen for a moment instead.
When the new thing is long, or is a different kind of thing from the original (a document rather than a fact), or needs the reader's whole attention for a while. Open spends the most space of the four, but it spends it in a way the page recovers from completely.
The first version of Sensitivity showed a notice period as a Transform: "90 days" became "the previous amendment", one step at a time. That broke the model's own rule. The reader who wants to know how the notice period went from 30 to 90 days is comparing. Comparison is a relationship, and the rule says keep both when their relationship matters. Try both.
Both, at different moments. The Transform is right when the reader only wants to confirm the current value and its basis. The Expand is right the moment they want the history. The interface cannot know which in advance, so it does the cheap thing first (Transform, zero extra space) and offers the expensive thing on request ("show the full history").
The third question is about trust. Two facts can look identical on screen and deserve completely different amounts of belief. One was copied from a signed document. The other was written by a language model a second ago. The reader needs to tell them apart without reading a manual. Hover each one.
The first version of Sensitivity had two sources, Retrieved and Generated. That missed two cases that matter in practice. Calculated facts are neither read nor invented; they are derived, and they are wrong whenever an input is wrong, which is a different failure from misreading. Confirmed facts are the only ones a person has signed off, and that changes what the system may do with them: a confirmed value should not be silently re-read and overwritten.
A fact can change source. A generated recommendation that a person accepts becomes confirmed. A retrieved value that a new amendment supersedes becomes stale and needs to be retrieved again. The marks should follow. This is why the source is shown as a small live mark rather than as a label typed once.
Asking the agent a question is not a fifth expression. The answer appears somewhere (usually as an Expand) and comes from somewhere (Generated). "Ask" answers who produced it, so it belongs on this axis, not the previous one.
A real moment is rarely one pure expression. Here is the "Ask" surface from the prototype. Type nothing; just press the button and watch what happens inside the card.
| Moment | Attention | Expression | Source |
|---|---|---|---|
| "Net 45", nothing new | Nothing | None (it is just there) | Retrieved |
| Reader wants to know why Net 45 | Hint (the grey line) | Transform | Retrieved, then the schedule |
| Agent recommends +6% | Reveal | Preview for the reason | Generated recommendation, resting on calculated figures |
| About to send to 42 people | Emphasise | Expand | Generated draft, retrieved list |
| Deemed acceptance in two days | Interrupt | Transform (line becomes a question) | Retrieved clause |
The three answers are independent. Change any one and the others can stay. That independence is what makes the model usable as a checklist rather than a style.
This is the part that progressive disclosure cannot do. The item does not change. The moment does. So the same line is quiet on Monday and in your way on Friday. Time is the clearest way a situation moves, so this part follows a deadline. It is not the only way: an amendment arrives, a colleague takes the item, a plan changes, and the same line moves without the calendar turning a page.
"No response by May 22 counts as acceptance." On 8 May that is a fact worth a small note. On 20 May it is a fact that could cost you a quarter's figures. Drag the date and watch the same item climb. proposal example data
| Days left | Level | What the reader sees |
|---|---|---|
| 14 to 8 | Hint | A small grey line under the item: "Reply by 22 May." |
| 7 to 5 | Reveal | The reason appears in words: "No reply by 22 May counts as acceptance." |
| 4 to 3 | Emphasise | The item moves to the top and is marked: "4 days left. Silence means you accept." |
| 2 to 0 | Interrupt | It appears first when they open the contract and asks for a choice: Reply · Accept · Ask for more time. |
| Passed | Nothing | A plain record: "Accepted on 22 May by no reply." No alarm. It is now a fact, not a decision. |
The thresholds (14, 7, 4, 2) are a starting guess. The right numbers depend on how long a reply takes in that organisation and would need to be tested. What should not change is the shape: a slow start, a steep end, and a return to silence once the moment has passed.
If the item jumped from Hint straight to Interrupt, the reader would feel ambushed: "this was quiet yesterday, why is it in my face today?" The steps let the reader see it coming. Each step also carries its reason ("4 days left"), so the climb never feels random.
The clock is the simplest input to the situation, which is why this page uses it. It is not the only one. The next page lists what else moves the level, and in which direction.
The clock is the simplest input to the situation, which is why this screen uses it. Three other things move the same item, in either direction. New evidence arrives: an amendment is uploaded and Net 60 becomes Net 45, with the old reading kept so the move can be explained. Someone takes ownership: the level drops for everyone else and stays for the owner. The plan around it changes: the supplier is being wound down, the run is cancelled, and an item that was climbing goes quiet, provided the change of plan is recorded where the agent can see it. In every case the level is recomputed from the current situation. It is never a flag set once and left on.
In the full paper
What moves the level, down and up · When the reader is not there · The order of the list changes too
A level that only ever rises is a badge counter. It turns into noise. For the climb to mean anything, there must be ways down, and they must be as clear as the ways up. And not every move is caused by the reader or the calendar. The situation around the item moves too, and the last three rows are about that.
| What happens | Where the level goes | Why |
|---|---|---|
| The reader acts (replies, signs, files the credit) | Nothing, immediately | The cost of missing is now zero. The item becomes a record. |
| The reader says "later" | One step down, for a fixed time, never past the deadline | They have seen it. Showing it again at the same level teaches them to ignore it. But it must come back before silence decides. |
| The reader opened the reason (Preview or Transform) | Stays, but the next climb is gentler | They already know. The Interrupt, when it comes, can be shorter and should not repeat the explanation. |
| Someone else acts (a colleague replies) | Nothing, with a note of who | The reader should not be interrupted for something already handled, but should be able to see it was handled. |
| The deadline passes | Nothing | There is no decision left to make. Shouting about it now is punishment, not information. |
| The underlying fact changes (an extension is granted) | Recomputed from scratch | The cost of missing changed, so the level changes with it. |
| New evidence arrives (an amendment is uploaded, an invoice lands) | Recomputed; up or down | The finding itself may change, as Net 60 became Net 45 in Part 3. The old reading stays in the history so the move can be explained. |
| Someone takes ownership (assigns it, opens a draft, starts a thread) | Down for everyone else; stays for the owner | The situation now has a person in it. Interrupting a second person about an owned item is the badge counter again. |
| The plan around it changes (the supplier is being wound down, the run is cancelled) | Recomputed from the new situation, often to Nothing | The contract did not change; what it means did. This only works if the change of plan is recorded where the agent can see it. |
The rule underneath all nine rows is the same: the level is always recomputed from the current situation, the cost of missing read against what is happening now. It is never a flag that was set once and left on.
Everything in this part has a hidden assumption: that the reader opens the product. An Interrupt that waits on a page nobody visits interrupts nobody. If the reviewer is on leave for the week of 22 May, the climb happens in an empty room and silence still counts as acceptance.
So at the top of the ladder, Sensitivity has to step outside the screen. This is the one place where it meets the notification systems from Part 2 directly.
Once the message leaves the page, placement no longer applies: there is no Transform in an email. What remains is the level and the reason. That is fine. Sensitivity is for the working screen. At its top step it hands over to a plain, well-worded notification, and the handover is the design.
The redrawn page in Part 1 had a quiet contradiction: the red item (silence decides) sat below an amber one. Was the list sorted by money, by date, or by colour? No reader should have to ask. If the level is the cost of missing, then the list is sorted by the same thing, and the order moves as the costs move. Drag the date again. proposal example data
The same four questions from Part 3 (money, days, undoable, silence decides) produce both the level of each item and the order of the list. There is no separate "priority" field for someone to set and forget. This matters because a hand-set priority goes stale the moment the situation changes, and nobody comes back to lower it.
A list that reorders itself can be disorienting. If the reader looked yesterday and the third item is now first, they may think something was added. Two things soften it: the item carries its reason for moving ("2 days left"), and items only move when they change level, never for a small change in score. A list that shuffles on every refresh would be worse than a fixed one.
Three moments in depth. For each: the answer most software gives, the answer Sensitivity gives, what each costs in words and attention, and the point where the Sensitivity answer would break.
The moment. The agent has read the vendor's renewal ask and 18 comparable renewals, and recommends a counter. The reader has to decide whether to send it. Cost of missing: about $134K a year, with 24 days to act, and the renewal auto-accepts if nobody does. Level: Reveal today.
If the reader has never seen this kind of recommendation before, a Preview is too quiet for the first time. They do not know the line carries a reason one step away. The first encounter probably deserves an Expand with the reason open, and later encounters the Preview. Sensitivity as written does not yet say how the interface should know it is the first time.
The moment. The agent has drafted a weekly summary from the team's activity. It will go to 42 customers. Cost of missing: a wrong summary reaches 42 people and cannot be recalled after 4:00 PM. Level: Emphasise, at the moment of sending only.
If this card appears every week with the same red line, the red line stops being red in the reader's mind. The Ancker finding from Part 2 applies: repetition cuts response. A better design might show the point-of-no-return line only when something is different from last week (more recipients, an attachment, a blocked item), and stay at Reveal when the send is routine. That is a judgment Sensitivity's current rules do not make.
The moment. A quarterly review has been sent for sign-off. The statement of work says that no response by 22 May counts as acceptance. Cost of missing: the figures become final without anyone having agreed to them. Level: climbs from Hint to Interrupt as the date approaches (Part 4).
If the reader is not the right person (the review should be signed by their manager), the Interrupt lands on someone who cannot act, and Sensitivity has no rule for handing it on. The "fallback person" idea from Part 4 covers absence, not wrong recipient. That gap is real and unsolved here.
The three deep specimens are all high-stakes. Most moments are not. Here are six that live at the quiet end of the ladder, each read with the three axes, so the model is seen doing small work as well as large.
| Moment | Attention | Expression | Source | Why that choice |
|---|---|---|---|---|
| Liability cap · $2,000,000, read from §11.1 | Nothing | Transform on request (show the value before the amendment) | Retrieved | Nothing is new. The old value matters only if someone asks. |
| Both parties signed overnight; contract filed | Nothing | None (the status simply changes) | Retrieved | Nothing for the reader to do. A notification would spend attention on a non-event. |
| 18 unread messages in a thread about a deadline | Hint | Transform ("18 unread" becomes the three lines that matter) | Generated summary of retrieved messages | The reader wants the answer, not the thread. The count has finished its job once summarised. |
| Supplier address changed in the latest amendment | Hint | Transform (the address becomes "changed on 3 May, was …") | Retrieved | The old address matters only if someone asks. Once read, it has done its job. |
| Renaming paused: 14 of 20 files done, 6 need admin access | Reveal | Expand (which six, and why) | Retrieved | The reader needs to see what was done and what was not, together, to decide whether to ask an admin. |
| An email address, "[email protected]" | Nothing | Transform on click (becomes "Copied") | Retrieved | The address has done its job the moment it is copied. A toast would spend space for a fact the reader already knows. |
Two patterns across the six. First, most moments are Nothing or Hint. That is the point. If the gallery were full of Emphasise, the model would have failed. Second, the expression is chosen by the comparison test every time: whenever the reader must weigh two things (done versus not done), it is an Expand; whenever one thing has finished its job (a count, an address), it is a Transform.
A model that only shows its successes is a sales deck. This part lists the places Sensitivity breaks, what has not been tested, and the honest size of the claim.
The notice-period case in Part 3 is the warning. A Transform that shows one amendment at a time looked elegant and hid the one thing worth seeing: the climb from 30 to 90 days. Any time the question is "how did we get here?", the least interface is the wrong interface. The comparison test exists to catch this, but it depends on the designer asking it.
Some information must be shown regardless of what it costs in attention: a regulatory disclosure, a term the other party must be able to see, a warning a court would expect. Sensitivity's "Nothing" level is not available for those. The model needs a floor that the cost-of-missing rule cannot lower, set by someone who knows the law, not by the design.
The source marks in Part 3 are small by design. That is also their weakness. If a reader stops noticing the difference between the teal dot and the violet one, a model's guess reads with the authority of a signed clause. This is the most dangerous failure in the whole model, because it fails silently and in the direction of over-trust. Wording ("appears to be", "the agent suggests") has to carry the difference as well as the mark, and even then it may not be enough.
Every rule in Part 3 that raises the level has to be matched by a rule that keeps things down, or the system drifts upward until it is a wall of Emphasise and the reader tunes out. This paper names that risk but does not solve it. A real system would need a budget per screen, something like "at most one Interrupt and two Emphasise at a time", and a way to decide which items lose when the budget is exceeded.
Preview is the cheapest expression and the one most tied to hover. On a phone, every Preview needs a visible trigger and a tap. For a screen reader, a Preview that appears visually but is not announced is simply absent. Each specimen in this paper has a keyboard path, but that was added one by one; the model has no general rule for it yet.
The whole model assumes the reader has learned which quiet cues carry a reason. A first-time reader has not. Sensitivity may need a louder first week and a quieter second one, and nothing here says how to do that.
The judgment in Part 3 reads only what is recorded. A decision made in a hallway, a supplier being replaced after a call, a colleague already on it: none of these reach the agent. It will raise an item that is already dead and stay quiet about one that is alive. The one-step correction is the only defence, and it only works if correcting is cheaper than ignoring.
Situation B in Part 3 works because the payment run is in a system. The tempting next step is to infer activity: you opened the contract, so you must be about to act. That inference is often wrong, and an Interrupt built on it is an ambush. The model should raise the level on recorded actions and treat inferred ones as a Hint at most. Nothing here says where that line sits.
Everything in the four questions can be true and the interruption still unwelcome: the person already knows, has decided, or is the wrong person. Being right about the contract does not make an interruption right. This is the failure the 3 a.m. call warns about, and the one a first test should look for hardest.
This is a design paper, not a study. It is important to be exact about what kind of evidence sits under each part of it.
| Claim | Evidence | Status |
|---|---|---|
| Too many alerts reduce response to each one | Ancker et al. 2017, hospital alert data | verified |
| Interruptions raise stress even when work gets done | Mark, Gudith, Klocke 2008, controlled study | verified |
| People hold 3 to 5 things in working memory | Cowan 2010, review of memory research | verified |
| The ideas in Part 2 exist and say what this paper says they say | Primary sources, listed on the last page | verified |
| Five levels is the right number | None. Chosen because four (Apple) felt too coarse for in-page use and six had steps nobody could tell apart. | proposal |
| The four cost-of-missing questions set a sensible level | None. Worked by hand on three examples. | proposal |
| The 14 / 7 / 4 / 2 day thresholds | None. Placeholders. | proposal |
| Readers understand the four source marks | None. | proposal |
| A list that reorders by cost is better than a fixed one | None. Could easily be worse if it moves too often. | proposal |
| The seven situation inputs, and that their relationships set the level | None. Argued from two invented cases and one phone call. | proposal |
| A one-step correction is enough to repair a wrong reading of the situation | None. | proposal |
| The specimens in this paper are usable | Built and used by the author. Not tested with anyone else. | proposal |
Take the three deep specimens from Part 5in the full paper (counter the renewal, send to 42 customers, sign or annotate by 22 May). Show six contract reviewers the usual version and six the Sensitivity version. Measure three things: how long until they act, whether they act correctly (counter, review recipients, reply), and whether they can say afterwards why the system recommended what it did. The third measure is the one that matters most, because it tests whether the "why" debt from Part 1 was actually paid.
A fourth measure, for the judgment pages in Part 3: wherever the agent's proposed level and a reviewer's own call disagree, write down the fact that explains the gap. Those facts are the real list of situation inputs. They will be more useful than the seven guessed here, and some of them will turn out to be things no system can be told.
Until something like that has been run, every rule in Part 3 and Part 4 should be read as a starting position, not a finding.
The size of the claim. Not a new way for software to judge situations; context-aware computing and mixed-initiative interaction have worked on that since the 1990s. The claim is seven specific, checkable design decisions, the first of which is that the agent's reading of the situation is carried through to the screen: it sets the level, shows in the wording, and can be corrected in one step. Each could turn out wrong in testing, and each is small enough that being wrong about it would not sink the others.
In the full paper
Counter the renewal · Send to 42 customers · Sign or annotate by 22 May · The honest size of the claim · Beyond contracts · Sources
The concept page this paper started from said, in effect, "here is a new way to think about interfaces". After Part 2, that is too big. Here is what can actually be defended.
Part 2 already lined up each existing idea against what Sensitivity adds. Stripped of the comparison, the claim is seven specific decisions:
That list is the claim. It is a set of specific, checkable design decisions, not a philosophy. Each one could turn out to be wrong in testing, and each one is small enough that being wrong about it would not sink the others.
What is not claimed: a new way for software to judge situations. Context-aware computing and mixed-initiative interaction have worked on that since the 1990s, and the seven inputs in Part 3 are a designer's list of what an agent would need to read, not an algorithm for reading it. The claim is only that the judgment should be visible, placed, worded and correctable on the page where the work happens.
Principles cannot be patented. Specific mechanisms sometimes can. Of the seven, the items most like mechanisms are the Transform-to-Expand switch triggered by back-and-forth reading (Part 3, flagged as untested), the two-input rendering where confidence and consequence control different outputs (Part 3), and the climb-with-descent rules (Part 4). I am not a patent attorney and have not run a search. Two cautions before anyone does: most of the examples here are contracts, so an employer may have a claim on contract-specific inventions, and the closest existing work (Horvitz et al.) is from a company with a large patent portfolio in exactly this area.
Contracts are a good place to find this problem because the cost of missing is often written in the document. But the rule does not depend on contracts. Three other places where an agent hands people conclusions, and what the three axes would do there.
An agent reviews a change and leaves twenty comments. Most are style. Two are a bug and a security hole. Level: style comments at Hint (a dot in the margin), the bug at Reveal, the security hole at Emphasise. Expression: the comment Transforms into the failing case on click; the security hole Expands to show the line and the exploit side by side, because the reader must compare them. Source: "the linter found" (retrieved, from a rule) is marked differently from "the model thinks" (generated).
An assistant drafts replies. Level: a reply to a newsletter is Nothing (sent, logged). A reply that commits money or a date is Reveal before sending. A reply to a legal thread is Emphasise. Expression: the draft Previews under the original message so the reader can check it against what was asked. Source: any figure the model inserted that was not in the thread is marked generated, so the reader checks it before it goes out. This is the "generated looks retrieved" failure from the limits page, in the place it does the most harm.
This is where the evidence in Part 2 came from, and where the cost of getting the level wrong is highest in both directions. An alert at Interrupt for every drug interaction trains clinicians to click through (Ancker's 30% per repeat). The Sensitivity reading: level from the severity of the interaction and whether the order can be changed later, not from the confidence of the match; expression Transform for the common case (the order line becomes the warning in place) and Interrupt only for the rare severe one; source a rule-based match (retrieved) marked differently from a model's inference (generated). I would not design this without clinicians, and the paper does not claim the model is ready for it. It claims the questions are the right ones to ask.
In all three, the specific levels and thresholds would be different. The three questions would not.
Software that does the work for people owes them two things: the conclusion, and a cheap way to check it. Sensitivity is a method for paying the second debt without bankrupting the first.
Four questions, asked at every moment: what does this mean for this person now, how much attention does that deserve, where should it go, and where did it come from. A ladder with five steps and ways down as well as up. A placement test with one question. Four source marks. Most of the time, the right answer is nothing at all.
None of it is finished. The limits page is longer than the claims page, and that is the correct ratio for an idea at this stage.
That was the story, in ten screens
Explore the full paper from the start · What already exists · The four expressions · The same moment, twice
Amrit Kumar · 2026 · Sensitivity is a personal project and is separate from any employer's product work. The contract data throughout is invented.
Each entry was checked against the source itself while writing this paper. Where a claim rests on an abstract rather than a full read, the text says so.
The redrawn pages in Part 1 and the specimens throughout are the author's own work. Earlier design-direction documents by the author ("Think less", 2026, and "Reduce doubt, not data", 2026) supplied the four-layer view of agent work, the "value × time pressure × irreversibility" ranking that Part 3's cost-of-missing questions extend, and the "67% asks, 94% acts" rule that separates confidence from permission.