Source
Duede, Gross, Crockett and Bergstrom (2026), LIT-492 — read as NOTE-241.
What it explains
| practice | what it says to do | what this says it is |
|---|---|---|
| SOTA-303 | attribute the phase a tool accelerates before predicting its effect on quality | not a caveat about tool quality but a consequence of the producer re-optimizing against a price that the tool itself moved |
The account
Consider anyone who works through a queue of items of varying value: they spend a fixed period finding out what an item is worth, then decide whether to finish it, then choose how much to invest beyond the minimum that makes it shippable. Their payoff over the long haul is a rate — total value produced divided by total time — and the optimal policy maximizes that rate.
The structural fact is that this rate is also the shadow price of an hour. An hour spent improving the item in hand is an hour not spent starting the next one, and the next one is worth the rate. So the policy is: keep improving until the marginal value of another hour falls to the rate, then stop. That is Charnov’s marginal value theorem, and nothing about it is specific to science or to language models.
Now give this producer a tool that makes some part of the work faster. The rate goes up — that is what the tool is for. But the rate is the price of an hour, so the bar every marginal hour of improvement must clear goes up with it. This is the whole mechanism, and it does not require the tool to be inaccurate, to homogenize anything, or to be used badly. It requires only that the producer is choosing.
The sign of the effect on quality is a race, and the race is won by phase.
- Accelerate the finding out what it is worth phase and the marginal hour of improvement is unchanged in value while the bar rose. Improvement falls. And because starting over got cheaper, the threshold for bothering at all rises: fewer items finished, each less finished.
- Accelerate the minimum needed to ship phase and improvement falls for the same reason, but the threshold moves the other way — the cheapest possible item is now cheap enough to be worth doing. More items finished, each less finished. These two are not variations of one prediction; they have opposite signs on the first quantity.
- Accelerate improvement itself and the marginal hour is worth more, by the same factor that raised the rate and then some. This one raises thoroughness. It is the case the paper proves and its own closing sentence forgets.
All three cases are propositions in LIT-492, which applies Charnov’s theorem to research effort, with discovery as travel time between patches and discretionary development as the diminishing harvest. Shortening discovery raises the development threshold and lowers thoroughness (Props. 5–7), and shortening the publishable minimum lowers the threshold and lowers thoroughness (Props. 8–10). Accelerating discretionary work raises thoroughness (Props. 11–12).
The third case is why the account is worth having rather than replacing with “tools make people sloppy”. The mechanism predicts its own exception, from a property of the tool you can state before deploying it.
Why Proposed
Because it is derived and not measured. Every proposition in the source is proved, and none of them is tested; there is no data in the paper at all. A proof establishes that the conclusion follows from the assumptions, and the assumptions here include that the producer optimizes a long-run rate, that item values are independent draws, that the payoff curve is concave above a hard minimum, and that institutions do not respond. Each is an idealization somebody could reasonably reject.
The status is about the account, not about the algebra. The algebra is fine.
What this does not say
It does not say a time-saving tool lowers quality. It says the bar on the marginal hour rises, and that two of three channels lose the race while the third wins it. A summary that keeps the conclusion and drops the phase has kept the part that is not determined.
It does not identify the phase for any real tool. The source assigns phases by example — discovery-like in “some technical fields”, publication-like in “fieldwork-based disciplines” — and gives no method for the attribution. Since the predictions have opposite signs across phases, an account that cannot attribute the phase predicts nothing in particular. That gap is the reason SOTA-303 is phrased as an instruction to do the attribution rather than as a prediction to apply.
It is not a claim about LLMs specifically, and the record should not file it as though it were. Any labor-augmenting tool has this shape. Compilers, autocomplete, automated experiment runners, faster training clusters and evaluation harnesses all remove time from an identifiable phase of somebody’s pipeline, and the argument reaches all of them. The source’s framing is scientific research because that is its authors’ field.
It is one of several accounts of an observation the record already holds. LIT-487 found an AI drafting tool that improved petition text on every measured feature while outcomes stayed flat or fell, and THEORY-051 offered feature-distinctiveness as one explanation while naming diminished author effort as a rival it could not account for. This is an account of that rival. It does not adjudicate between them — both predict the same null — and the two papers were written independently of each other. The mapping of a petition tool onto this model’s minimum-publishable phase is a reading this record makes, and neither paper makes it.