Where the human sits

The binding constraint is your processing rate, not the machine's output rate. Tokens, context and parallel instances are abundant and getting cheaper. What is scarce is your attention at the moment a decision is actually made. Everything on this page follows from that one sentence, and the design consequences are less obvious than the sentence sounds.

What the machine is actually for

If the human is the constraint, then the assistant's job is not producing more. It is matching its rate to yours, and there are four distinct ways it does that. They are worth separating, because three of them touch your judgment and one of them does not.

compressreduce the volume that has to pass through you at all. The most obvious function and the one with the sharpest edge, for reasons the rest of this page is about.

shape for a decisioncollapse read, then synthesise, then decide into simply decide. The expensive middle step is the one being removed, and removing it is most of the value.

pacecontrol the rate and the order things arrive in. The map before the detail, because arriving at detail without the shape of the thing costs more than it saves.

presentthe same content at a lower cost to read, through structure and visual hierarchy. The only one of the four that automates no part of a decision, which makes it the safe lever to pull hard.

What can be offloaded at all

One cut decides this, and it is worth making explicitly before delegating anything.

A decision is a branch point. The load sits in evaluating the options, and the preparation for it is delegable. Someone else can gather the options, lay out what each one means, and say what the record argues. That genuinely removes load rather than moving it, because the part that was expensive was the preparation and the part that is yours is the choice.

Judgment-work is judgment entangled with the doing. Writing the paragraph where the argument lives, deciding what to leave out without breaking the narrative, reading a situation. It cannot be offloaded. It can only be supported: clean inputs, sensible pacing, someone else holding the map while you work. Delegating it does not remove the judgment. It relocates the doing and leaves the judgment exactly where it was, now separated from the material it needed.

The governor

Here is the part that is easy to miss. Whatever compresses for you also decides for you. A summary that leaves out an option has made a choice on your behalf, and you cannot audit a choice you were never shown. Compression is delegated judgment that is invisible by construction.

The obvious correction is to surface more. That correction has its own wall, and it is the same failure wearing different clothes.

Both ends of the scale arrive at one place. The left wall is the branch that never reached you. The right wall is thirty branches reaching you at once, which no one deliberates, so they get pattern-matched and passed. The failure is identical from the outside: work going out under your name containing decisions you did not make. Only the middle leaves you deciding what was actually yours.

Compression is what makes governance possible

The two walls look like a trade-off to be balanced. They are not, and seeing why is the useful part.

Attention is a single budget. Every unimportant thing surfaced to you spends attention that a real branch point needed. So compressing the safe half hard is not in tension with surfacing the important half. It is the precondition for being able to surface it at all. Under-compression floods the same channel the branches have to arrive through, which means it is itself a governance failure and not merely an inefficiency.

The precise form is worth keeping. Compression enables governance. It does not constitute it. An assistant can compress beautifully and still quietly settle something that was yours, and no amount of elegance in the summary tells you whether that happened.

The residual nobody can close

There is a hole here and it does not close, so it is better named than papered over.

The dangerous branch is the one the assistant wrongly believed was safe. It is not surfaced, precisely because it was not recognised as a branch, so nothing downstream catches it. A misquoted source can be caught mechanically, because a quote is a string and either appears in the document or does not. A collapsed option is not a string. There is nothing to match it against.

What is available is indirect. When you find yourself telling an assistant that something was not its call to make, that is the leak becoming visible after the fact. The rate of that particular correction is the only real signal on whether any of this is working, which is a good reason to record corrections by kind rather than only counting them.

The dial, not the setting

None of the above has one correct position, and treating it as though it does is its own failure. Heavy support is right in a domain you are new to. The same level of support in a domain you have worked for years takes the reps that build your own capacity and gives them to the machine.

This matters more than it first appears, because of the asymmetry in the pair. Only one party's deep priors keep updating. The machine arrives at each session the same as it was, and improves between releases for reasons you do not control. You are the part of this arrangement that compounds, and permanent maximum compression is a slow trade of that compounding for a smoother week. The adjustable dial is the feature. A fixed setting is a crutch whichever end it is fixed at.

What can be measured of this, and what cannot

Some of it is observable in a record and much of it is not, and the boundary is worth stating plainly before anyone builds a meter.

Load is usefully modelled as weight times position: how much a piece of work holds together at once, and where in a working session it falls. Position is readable from a transcript. Weight is not, and one attempt here to proxy it with a machine-side metric was ruled out on inspection, because how much context the assistant was carrying is not how much the person was holding. That is a cross-axis error, and it is easy to make.

machine context healthhow much the assistant is carrying and how coherent it stays. Readable, and about the machine only.

human loadwhat the work costs the person. Largely not readable from a record, and the axis everything else gets mistaken for.

output fidelitywhether the work is faithful to its sources. Readable, and a trust question rather than a load one.

Keeping those three apart is most of the discipline. A transcript measures the interface between you and the assistant: how much passes each way, what gets volunteered, how often and where you correct it, where the output lands. Arousal state, recovery, perceived load and everything happening away from the keyboard are outside what the record contains. They are not hard to measure from a transcript. They are absent from it, which is a different problem and needs a different instrument.