I taught my bot to confuse important with urgent

The triage queue I started on 13 August holds eighty-four analysed items. Sixteen came back as "relevant now". Thirty-two were dropped as noise. The remaining thirty-six went onto a "come back to this" list. The rule that sorted them read: does this content change what I do within fourteen days, measured against my project priorities. The model applied it correctly. The rule was the bad part, because it welded two independent things into one verdict, weight and deadline.

My claim is that a tool built around your own attention does not increase the supply of it, it moves the cost of judgement outside you, and only then can you see how much of that cost there was and what shape it had. A decision hangs on this, easy to state and uncomfortable to answer: keep such a prosthesis running, or switch it off. Three numbers will settle it. Two I have only begun to collect; the third I do not yet measure.

Attention is a budget, and Simon supplied the unit

Herbert Simon set this out in 1971 in his lecture "Designing Organizations for an Information-Rich World". The sentence about a wealth of information creating a poverty of attention gets quoted to death. What Simon did immediately afterwards gets quoted far less. He said an allocation problem has to be posed properly, which needs a measure of the scarce resource that cannot be stretched at will, so he rejected Shannon's bit: bit capacity depends on how a message is encoded and is therefore not an invariant. What was left was the simplest unit available. Simon proposed the time a recipient spends on a message (Simon 1971, pp. 40–41).

This is where the line runs between a metaphor and a model. If attention is a budget, a link dropped into a queue is not a saved expense but a deferred one. You drop it because right now you have nothing to pay the judgement with. The bill waits.

Where the debt comes from when you defer judgement

Sophie Leroy showed in 2009 that moving from one task to another leaves attention residue: part of your cognitive resources stays with the previous task and does not work on the current one. The residue is larger when the previous task is left unfinished (Leroy 2009). I treat an item dropped into a queue as such an unfinished task: you watched enough to call the thing potentially valuable, and too little to do anything with it.

More important for this argument is what Leroy published with Theresa Glomb in 2018 in Organization Science. Data from four studies supported the authors' hypotheses: people who expected to resume interrupted work under time pressure did worse on the task that interrupted them, and part of their attention stayed behind. A brief plan for returning eased this. The plan did not finish the interrupted work (Leroy and Glomb 2018).

From this I draw a hypothesis of my own, one these studies did not test. I assume the mechanism is not that the prosthesis reads for you. A prosthesis is meant to turn an open loop into a recorded decision, the counterpart of a plan for returning, and such a record may pay down the attention debt without the reading being finished. The authors studied a plan for returning to interrupted work under time pressure, not a verdict on a deferred link with no deadline, so carrying their result over to a bot's queue remains my assumption. The working criterion can be derived from this in one sentence: the system works when it makes a decision you do not have to make a second time.

A rule written into code shows what you cannot choose

Separating importance from urgency is not a discovery. Covey described it in 1989 as the second quadrant of the matrix: things that are important and not urgent, the ones that lose every single day and win the decade. The maxim underneath is older. Eisenhower delivered a version of it in Evanston on 19 August 1954 and credited the words to an unnamed "former college president" (via Quote Investigator 2014), so treat the named attributions in circulation today with care. The point is that I knew the distinction and still wrote a rule that papered over it. I wrote it out of my own reflex, not out of the book.

The same reflex came out a second time, in another tool and in a harder form. I run a weekly radar that picks three things to do out of everything that arrived in a given week. Priority was impact times confidence, divided by effort, and effort entered as a linear divisor rising from 0.5 for half a day's work to 5 for work longer than three days. With that divisor the best possible large move scores 5 times 1 over 5, which is 1.0. Average small stuff scored 2.8 to 3.6. A large correction could never win, whatever it was worth. This was not a model error or a data error. The ceiling sat in arithmetic I had written myself.

By construction, both rules pushed the same class aside: things important, slow and without a deadline. The categories I invented turned out to be a map of my own gaps. I saw them only once I had to write them in a form a machine executes, because a machine makes no exception for a thing whose author privately knew it mattered.

One design decision from that period held up completely. A rejecting verdict still writes one line to the wiki with a link and a reason. The rejection is therefore reversible and countable, and I can check after the fact what the system threw out and why. A bad system is one that deletes quietly.

You can call this procrastination and sometimes you would be right

The strongest objection to this thesis is not "tools do not help". It runs: building a system for reading is a form of avoiding reading, only with a better justification. The objection lands and it narrows the conclusion, so I will not leave it in a footnote.

A prosthesis defends itself under two conditions at once. First: the stream is larger than the attention budget. Eighty-four items in eighteen days is not quite five a day, and for material that has to be watched and judged that is more than I have to pay with. Second condition: the cost of keeping the prosthesis is falling. Eighteen days of triage cost 3.14 dollars in model charges, roughly 17 cents a day, and it fits inside a hard cap I set for myself. A radar run cost 17 cents, the same as one day of triage.

Honestly: those figures speak to money, which was small from the start, rather than to the cost of my attention, which is the thing actually in dispute. I do not yet measure upkeep in hours, and after eighteen days I have no right to claim the curve is falling. If the cost of upkeep is rising then it is a hobby, and a hobby is allowed as long as nobody calls it infrastructure.

I split the axes and fell into the same hole one floor up

In the radar the axes are no longer welded. A data point gets a class, impact from 1 to 5, effort, confidence from 0 to 1 and a gate flag for product ideas, each on its own. No deadline is written into any of those axes. That looks like a repair and at the level of classification it is one.

What settles the outcome is not the split but the formula that welds the axes back together. It contains a rule of mine: a product idea whose only source is content from the internet gets confidence of at most 0.4, and the rule is a sensible one, because a single creator's video does not justify changing a product. In the first real run the effect came out like this: two product ideas with impact 4 got a priority of 0.36. They landed at the bottom of the backlog, below small stuff with impact 2. Of the fourteen proposals the radar picked three: two small things and one medium, the last of them only because I had added a separate rule guaranteeing one slot to something bigger.

The class the first version rejected outright, the second version no longer rejects. It pushes it to the end of the list with a multiplier I judged reasonable myself. The scope of the thesis therefore has to narrow: separating the axes is a necessary condition and not a sufficient one. What survives is decided by the recombination rule and by who is obliged to look at the result.

The obligation is in worse shape than the arithmetic. The radar has had one run so far, on 24 August, and it produced fourteen proposals. Six days later not one of them carries a recorded decision, and the trust ledger meant to hold my acceptances and rejections is empty. Thirty-six items on the "come back to this" list plus fourteen proposals without a decision makes fifty deferred judgements. I started with one queue of unjudged things. I now have two, neatly sorted.

I wrote the condition for my own failure into the spec before I saw these numbers: if I accept fewer than 30 percent of proposals after six runs, the kill criteria apply. I have one run and zero acceptances, so the counter has not started.

Of these three numbers, only that counter has a threshold, and only it decides whether to switch off. The other two have none: together with the counter they show whether what stays switched on is a prosthesis or a hobby. The first measure, the acceptance rate, does not, however, measure the mechanism I have been describing. It says how many proposals I accept. It does not say whether the system closes matters. I could reject every proposal and thereby settle each one.

So the second measure stands apart: what share of deferred judgements ever received a recorded decision. Both acceptance and rejection count. I adopt the Pareto principle: about 20 percent of deferred judgements carry 80 percent of the value. I therefore do not need to settle all of them, only the right ones. As of 30 August none of the fourteen proposals has a recorded decision. The third measure is the time I spend on the system's upkeep. I do not measure that time yet. Until I do, I cannot say whether the prosthesis relieves me or only moves the work from judging to upkeep. Only these three numbers together will let me tell a prosthesis from an expensive hobby.

References

1. primaryHerbert A. Simon, "Designing Organizations for an Information-Rich World", in M. Greenberger (ed.), Computers, Communications, and the Public Interest (The Johns Hopkins Press, Baltimore, 1971) — the quoted passage and the argument that an allocation problem needs a measure of the scarce resource, with the bit rejected and time proposed instead, pp. 40–41. Read in the original scan.
2. primarySophie Leroy, "Why is it so hard to do my work? The challenge of attention residue when switching between work tasks", Organizational Behavior and Human Decision Processes 109(2), 2009, 168–181 — attention residue and its dependence on how the previous task was left.
3. primarySophie Leroy and Theresa M. Glomb, "Tasks Interrupted: How Anticipating Time Pressure on Resumption of an Interrupted Task Causes Attention Residue and Low Performance on Interrupting Tasks and How a 'Ready-to-Resume' Plan Mitigates the Effects", Organization Science 29(3), 2018, 380–397 — data from four studies supported the authors' hypotheses; for people who expected to resume under time pressure, a brief plan for returning to the interrupted work reduced attention residue, and performance on the interrupting task did not suffer.
4. primaryStephen R. Covey, The 7 Habits of Highly Effective People (1989) — the important–urgent matrix and its second quadrant.
5. secondaryQuote Investigator, "What Is Important Is Seldom Urgent and What Is Urgent Is Seldom Important" (2014) — Eisenhower's address in Evanston on 19 August 1954 as the earliest evidence the page's author knows of: he credited the words to an unnamed former college president, not to himself; wordings with "seldom" may stem from faulty transmission of the speech or from Eisenhower's use of more than one version.

Figures on the triage queue, its costs and the radar run come from the database and repository of my own tools, as of 30 August 2026.