Source
Duede, Gross, Crockett and Bergstrom (2026), LIT-492 — read as NOTE-241.
When this applies
You are deploying, evaluating or forecasting a tool that makes people faster at some part of producing something — a coding assistant, a drafting tool, a literature-search agent, an automated reviewer, a faster experiment loop. The people using it choose how much effort to put into each item and how many items to take on. If they are demand-limited rather than rate-limited — they have exactly the work they have and no queue behind it — none of this binds.
Do this
The three-phase split is LIT-492’s, and this practice starts there: the paper applies Charnov’s marginal value theorem to research effort and proves that shortening discovery or the fixed minimum lowers the thoroughness of every developed project, while accelerating discretionary improvement raises it (Props. 5–12). The table below is those propositions read as forecasts.
Attribute the phase before predicting the effect. There are three, and they are distinguishable by a question about the tool rather than about its users:
- Before the item’s value is known. The tool helps find out whether the thing is worth doing — hypothesis generation, triage, search, a quick feasibility check.
- The fixed minimum needed to ship at all. The tool removes required overhead — formatting, boilerplate, figure construction, the first draft that must exist before anything can be submitted.
- Discretionary improvement above that minimum. The tool makes the improving faster — the extra experiment, the sensitivity analysis, the counter-example search, the revision pass.
Then read the prediction off the phase:
| phase accelerated | fraction of started work that ships | thoroughness of what ships |
|---|---|---|
| 1 — discovery | falls (more selective) | falls |
| 2 — minimum to ship | rises (less selective) | falls |
| 3 — discretionary improvement | rises, under a regularity condition | rises |
THEORY-055 gives the reason all three share: the producer’s own rate of return is the price of their time, every acceleration raises it, and the marginal hour of improvement then has to clear a higher bar. Phase 3 is the case where the hour’s value rises faster than the bar.
The negative this replaces
Do not treat time saved as time returned to the same work. “Automating the tedious parts frees us to think more deeply” is the default forecast, it is the one the source was written against, and it is the phase-3 prediction applied to phase-1 and phase-2 tools. If your tool removes overhead, the model says the freed hour goes to the next item, because the next item is now worth more per hour than it was.
And do not run the inference backwards. Observing more output at lower quality after a tool lands does not identify this mechanism. Every phase here predicts lower thoroughness except one, and so do several accounts that have nothing to do with opportunity cost: newly-enabled producers who are worse than the incumbents, readers discounting what they suspect was generated, and skill loss from disuse. The practice is a forecasting discipline, not a diagnosis.
Conditions
This is a model with no measurements in it. Not underpowered — absent. Every claim is a proposition with a proof, and the propositions are about a stylized producer, not about anyone observed. The source’s own description is “a simple mathematical model to illustrate”.
The attribution method does not exist yet, and it is the load-bearing step. The source assigns phases by example and gives no procedure. Since the table above has opposite signs in its first column, a wrong attribution inverts the prediction, and nothing in the source helps you check one. Treat the phase as a hypothesis you state in advance and can be wrong about, which is at least a different epistemic position from having no prediction.
Most real tools act on more than one phase. The source says so and studies them separately anyway. A coding assistant that writes boilerplate (phase 2), scaffolds an investigation (phase 1) and speeds up a refactor (phase 3) has all three signs at once, and the model has nothing to say about the net.
Institutions are held fixed. Peer review capacity, what counts as a shippable unit, and what reviewers demand do not respond in the model. The source notes submissions already straining review as an institutional response beginning, and cannot represent it. The shorter the horizon, the safer the assumption.
The tool is stipulated to be perfect. No errors, no cost. That is a strength of the argument — it isolates the consequence of speed — and a limit on the practice: a real tool’s effect on quality is this effect plus whatever its mistakes do, and this says nothing about the sum.
The source’s own summary overshoots its own result. “Impelling us to do more, less well” appears unqualified in the abstract and again as the closing line, and Proposition 12 proves the opposite for phase 3. The analysis section states the exception. A reader who takes the memorable sentence has taken the part the paper disproved, which is why this practice is written around the table rather than around the conclusion.
Related
SOTA-298 says to measure a writing assistant by the outcome it was deployed to improve rather than by the text features that used to predict it, on the strength of a deployment where the text improved and the outcomes did not. This is the complementary half and a different kind of claim: that one is about what to measure after the fact, this is about what to predict before it. Neither is evidence for the other — LIT-487 measured outcomes and not effort, and this source measured nothing — and the two papers do not cite each other.