40.1Generation and editing costs
A transparent cost account must separate what a project spends at each stage. Generation is the labor of producing a draft. Source preparation covers finding, checking, and organizing the material the draft relies on. Review is assessment against a declared rubric; revision is the work of changing the text in response. Reader effort is what the audience must expend to use the result, and participant burden is the cost imposed on people who supply input or are described by it. These categories are distinct accounting entries, not statistically independent quantities; each can vary, and a project can move effort from one to another without lowering its total.
That separability matters when two encodings are compared. The task and inputs must be held fixed; otherwise a difference in generation effort may simply reflect different work being done. Suppose, as a prospective design rather than an observed result, that one encoding produces a draft more quickly while requiring more review against the same rubric. The reported quantity is then a comparison of elapsed time for draft production under a fixed task, not a measurement of active labor or of overall cost. Review, revision, reader effort, and participant burden remain separate entries, each carrying its own value.
Two further distinctions follow. Summed active labor and parallel elapsed time answer different questions: the first aggregates effort across people, the second measures how long the work takes when activities overlap. Neither substitutes for the other without a stated conversion. Likewise, reader effort transferred away from producers is a relocation of cost, not its disappearance. In a bounded interview study of practitioners, undervaluing data work was reported to contribute to compounding downstream problems (Sambasivan et al., 2021); that finding belongs to its sample and supports sustained attention to preparation, without any prevalence claim about the workflow of this study.
Once the cost categories are fixed, the next question is what a recorded figure actually licenses one to conclude. Suppose an accounting record captures only the bill for generating a draft. That figure measures the generation category alone. It says nothing about source preparation, review, revision, reader effort, or participant burden, whose values may simply be unrecorded. Unrecorded is not zero: an unknown term leaves the total indeterminate rather than small. Consequently, "cheap to generate" is not "cheap to publish" or "cheap to read." The first claim is supported by the ledger entry; the second requires values for every category that enters the total.
Two further conditions keep such a comparison honest. First, the task and inputs must be fixed. Editorial effort varies with the length of the source, the number of distinctions to be retained, the level of review imposed, and the difficulty imposed on readers. If these differ between two workflows, the observed gap confounds the encoding with the job being done. Holding the task constant makes the encodings themselves the object of comparison; leaving it free makes the cost figure uninterpretable.
Second, a named quantity must be identified before anything is summed. As Section 40.1.1 distinguishes, parallel elapsed time and summed active labor answer different questions. Reporting "time saved" without saying which has been measured lets a reduction in elapsed duration pass for a reduction in effort.
Under uncertainty, the appropriate statement is conditional. If every category is known and measured on a common basis, totals can be compared. If one remains unknown, only a partial account exists, and the true total may prove lower or higher than a rival's. A low generation bill then bounds one term of the sum; it is not evidence about the whole. Where a category is missing, the output of the comparison is an open question, not a low total.
40.2A fixed synthetic comparison
Consider a synthetic scorecard, in which three candidate editorial treatments are judged on two declared dimensions: retained distinctions, where more is better, and arbitrary editorial cost units, where less is better. Treatment A retains 10 distinctions at 6 cost units; treatment B retains 9 at 4; treatment C retains 8 at 8. These values are invented for the sake of the example. They are not dollars, measured editing times, or results from any model benchmark, and they support no empirical claim about real workflows.
On these two dimensions, the comparison is straightforward. A retains more distinctions than C while consuming fewer cost units, so A dominates C; the same relation holds between B and C. C can therefore be set aside without further argument, at least within the fixed scorecard. A and B, by contrast, do not dominate one another: A preserves one additional distinction but costs two more units. Both remain Pareto-optimal, and the choice between them is a genuine tradeoff rather than a question that arithmetic can settle.
This is where a value judgment enters. A reader who places high weight on preserving distinctions may prefer A, while one who is more constrained by editorial effort may prefer B. Treating editorial effort as a separate comparison axis makes that pressure explicit instead of letting quality claims float free of the labor they require. It also matters whose labor is counted: an accounting that omits participant burdens understates cost, even if the omission is unintentional.
Finally, no missing dimension is declared unimportant. Coherence, accessibility, fairness, and downstream interpretability might reverse the ranking, and the example asserts nothing about them. This is a fixed toy comparison, useful because its tradeoff is visible and its limits are declared; it is not a recommendation independent of values.
The comparison already laid out does its work through structure, not through data. Its numbers are stipulated so that two competing dimensions—retained distinctions and arbitrary cost—can be held in view at once. They are not measurements of a working system, and the cost units are neither money nor hours. What the example licenses is therefore narrower than what an empirical evaluation would, but also clearer about what it does license.
One distinction matters especially. Within the invented set, the options can be ordered: more retained distinctions is preferred over fewer, and lower cost over higher. Because cost is assigned in arbitrary units, only that ordering carries force. Ratios and differences do not: a claim that one option saves "twice" another's cost, or that an extra unit of retention is worth a stated amount, has no content here. There is no exchange rate between the two dimensions, and the comparison supplies none.
Applied, then, the example supports structural inferences that survive arbitrary units. Dominance is such an inference: if an option does at least as well on both stipulated dimensions and strictly better on one, the ranking is insensitive to how the dimensions are scaled, provided preferences on each dimension run as declared. Tradeoffs are the countercondition: when one option retains more and the other costs less, no ranking follows from the two dimensions alone. Choosing between them requires something the toy does not contain—a weighting, a budget constraint, or a threshold—each of which must be stated as an additional premise rather than read off the numbers.
Suppose, by way of illustration only, that cost were recorded in measured hours and retention scored by a stated procedure on real revisions. The comparison would then be partly empirical, and its ordering could fail. In its present stipulated form it cannot fail in that way; its limits are part of what makes it useful as a test of reasoning rather than as a verdict.
40.3Dominance and tradeoffs
Domination earns its place because it separates two questions that a single ranking would blur: which options can be set aside at once, and which remain to be chosen among. On the declared dimensions, A retains ten distinctions at six cost units, B retains nine at four, and C retains eight at eight. With more retained distinctions and less cost both preferred, A is better than C on both counts, and so is B. Each therefore dominates C.
Domination is also doing only one kind of work. Being dominated is sufficient for dropping C under the declared preferences, but non-domination is not sufficient for choosing an option: A and B persist because A retains one more distinction while B spends two fewer cost units. Neither is at least as good as the other on both dimensions, so neither dominates the other. Nor is domination necessary for a reasoned pick; it is a screening relation, not a complete ranking. C's exclusion follows from the stipulated preferences, whereas a preference for A over B is possible but not forced by the comparison.
Keeping editorial cost as a separately declared dimension matters here. "Retained distinctions" compares counts on the toy scorecard, and "cost units" compares the declared cost; no exchange rate between them is supplied, so the pair of values yields a partial order rather than a composite score. As a labelled supposition, if a conversion rate between distinctions and cost units were later introduced, the ordering of A and B would depend on which rate was adopted. C's position would not: as long as more retention and less cost remain preferred, any weakly monotone evaluation still ranks both A and B no worse than C. Strict superiority over C would require suitable strict responsiveness. The tradeoff is thus sensitive to valuation in a way that the dominance result is not.
The fixed comparison set out in Section 40.2 supports a narrow conclusion. C is dominated because each of A and B is at least as good on every listed dimension and better on at least one; A and B are not ordered against each other because each preserves something the other sacrifices. That is a sufficient condition for dominance within the declared comparison, not a sufficient condition for general superiority. The claim is therefore about these dimensions, this task, and these inputs.
Now vary the dimension set. If a dimension were added on which C outperforms both A and B, neither would dominate C and the comparison would become a three-way tradeoff; a dimension on which A and B tie or both exceed C leaves the result intact. But an added dimension cannot make A dominate B: the existing dimensions on which B is better still block that claim. Removing a dimension works the other way, since dropping the dimension on which one option gives up ground can turn a tradeoff into dominance. Whether a change in dimensions overturns the ranking is therefore possible, not inevitable; it depends on which option the added or removed dimension favors.
Because the cost units are arbitrary, the cost dimension ranks these options for the present purpose; it measures neither money nor elapsed hours. Editorial effort should be entered as a separate declared dimension rather than folded into the quality scores, so that later additions remain visible in the comparison. Any test of alternative encodings of the same options must hold the task and inputs fixed; otherwise a change in the result is confounded with a change in the problem. On this invented illustration, what is established is conditional: within the declared dimensions, C is dominated while A and B trade off. Beyond them, dominance is an open question, not a settled result.
40.4Whose effort enters the total
A total is a record of whose work was counted, not a natural fact about the project. Before one quality dimension can be weighed against another, the parties who bear that dimension must be named. Generation, preparation, review, revision, reader effort, and participant burden are distinct categories, and each can fall on a different party: a contributor produces material that an editor later repairs, readers absorb the difficulty that remains, and an institution carries scheduling and infrastructure obligations. Folding these into a single effort column therefore makes a distributional claim before it makes an arithmetic one.
Two checks follow. Inclusion is necessary for a claim about total burden but not sufficient for a meaningful one: once several actors' labor sits in a single column, their active hours are being compared as though they were the same kind of thing, and that basis has to be stated. Omission, meanwhile, is not absence. Where a category has not been measured its value is unknown, not zero; silently dropping reader effort manufactures a saving that nobody observed.
The named quantity must also match the question asked. Summed active labor reports how many working hours a task consumed; parallel elapsed time reports how long the task blocked the schedule. The two can move in opposite directions — a review round that adds staff hours while shortening the waiting period — so no single figure stands for both.
Applied to the synthetic case, then, each option's effort declaration has to say which party bears each category and which categories remain unmeasured, before that dimension can enter the dominance comparison at all. A column that silently reallocates labor across actors has fixed neither the inputs nor the task being compared. Whose effort enters the total is a declared choice, not a quantity waiting to be discovered.
Once the boundary of participants has been drawn, scope still leaves a second question open: which quantities the total is allowed to add. Generation, preparation, review, revision, reader effort, and participant burden remain distinct categories even when they bear on the same editorial task. Declaring who is inside the boundary is necessary for a total to be meaningful, but it is not sufficient to make that total comparable across options.
The temptation is to treat a reader's saved time as interchangeable with editor labor: so many minutes saved, so much work released, one sum. That step needs a declared method. Conditional on the project having stated an exchange rule — for example, a rate at which reader time offsets preparation work — the two can be placed on one axis. Absent such a rule, the quantities should remain separate dimensions of the Pareto comparison rather than be merged into a single figure. The same caution governs conversion to money: monetary aggregation is possible only under a declared valuation method, not as a neutral bookkeeping move.
A second check concerns what the named quantity actually measures. Parallel elapsed time records how long the process runs; summed active labor records how much work different people perform. Two designs can share the same elapsed interval while differing sharply in labor, or match in labor while differing in elapsed time. Holding the task and inputs fixed, then varying the encoding of effort is precisely where these quantities part company.
Suppose a wider scope is adopted, so that participant burden is counted alongside editor effort. It does not follow that any option becomes dominated; it remains possible that the added dimension leaves the genuine tradeoff intact. What scope changes is the evidence visible to the comparison, not the outcome it must deliver. Unknown values, meanwhile, stay unknown: treating them as zero would silently delete a burden the boundary has just admitted.
40.5Latency and repeated requests
Latency and cost diverge in a second way once a request fails and is retried. Three quantities must be kept apart: active work time, accumulated across the categories distinguished earlier; elapsed latency, the interval from request to final disposition; and the increment of work performed again after a failure. The third is not a fresh dimension of quality but a consequence of the first attempt, and it must be measured against the counterfactual baseline of a single successful attempt. If a scheduled second review would have occurred anyway, only the additional pass counts as repetition.
Because these quantities move differently, a failed request need not raise them equally. Suppose a revision is rejected at review: generation and review are performed again, but preparation may not be. If a different reviewer must be found, the repeated labor falls partly on categories not exercised the first time, and the elapsed interval lengthens by more than the repeated active minutes. This is a prospective design inference, not an observed result.
Retrial is also conditional, not inevitable. The same failure may end the request, or the degraded output may be accepted, in which case the rework term is zero while latency still includes the failed interval. Unknown failure probabilities therefore cannot be entered as zero, and a design that assumes every failure leads to repetition will overstate rework.
A named quantity needs a stated comparison. "Active work time" is a total only if categories share a unit and an aggregation rule; otherwise the report is a vector of category amounts, and adding them silently imports a weighting. Elapsed latency compares calendar durations, not burdens on particular people. The next question is whether a faster option is also cheaper; that cannot be read off latency alone.
Once active work, elapsed latency and repeated work are distinguished, a further distinction appears within latency itself. Parallel activity can shorten the interval between start and completion while leaving summed active labor unchanged. The two quantities describe different features of the same process: elapsed latency compares how long a task takes to finish, while summed active labor compares how much work is performed. Holding the task and its inputs fixed, compressing a schedule changes how work is arranged; it is not evidence that less work was done.
A faster completion is therefore neither sufficient nor necessary for a cheaper one. It is not sufficient: two options may differ in speed while consuming the same active labor, and cost may include categories that speed does not touch, such as the review and reader-effort categories kept distinct in the illustrative project. It is not necessary either: an option finishing later could involve less active work. Speed settles neither direction without an added assumption linking the two.
The inference has a conditional side as well. If concurrent activity requires additional coordination or a repeated review, summed active labor may rise even as elapsed latency falls. Whether it does is a feature of the particular design, not something parallel scheduling itself guarantees; where the coordination requirement is unknown, it cannot be recorded as zero.
Comparisons must fix the task and inputs before encodings are set against each other. If one option handles a broader corpus or a heavier revision, a difference in elapsed latency mixes scheduling with scope, and the comparison no longer tests the encodings alone. What parallel latency supports is a claim about completion times for the same work. What it does not support, by itself, is a claim about labor, total cost, or which option to prefer; declaring editorial effort as a separate cost dimension keeps that judgment from being made by substitution.
40.6The value of stopping
Once a manuscript is workable, the question is marginal: does the next revision earn its cost? The comparison is between increments rather than totals. On one side stands the expected benefit of one further pass—less ambiguity, better fit to the intended readership, a tighter link between the claimed contribution and its evidence. On the other stands what that pass declares: editorial labor, and the elapsed time it consumes, recorded separately rather than folded into a single quality score. When the expected gain is smaller than the declared cost, stopping is rational; further work is warranted only when the comparison runs the other way.
The word expected does work here. Suppose a hypothetical pass over an already revised section is expected to remove one remaining ambiguity while demanding substantial editorial attention and postponing delivery. The improvement is possible, not inevitable: a revision may change nothing, or trade clarity in one dimension for awkwardness in another. Expected benefit is thus a claim about the range of outcomes under the stated assumptions, not a promise about this particular manuscript. Stopping on a negative margin is therefore a sufficient reason under those assumptions, but not a necessary one; a deadline, an exhausted editor, or an unavailable reader can justify the same decision without altering the estimate. The converse also holds: a positive margin makes continuing worthwhile in expectation, not assured of success.
The margin must name what it compares: the expected gain from one further revision against the cost that revision declares, for the same manuscript, the same target readership, and the same revision task. Fix the task and the inputs; the way the decision is recorded may vary without changing the comparison. Nor is the margin stored once for all. If, as a supposition, a fresh reading surfaces a concrete defect, the expected gain rises, and a pass that was not worth making becomes worth making. Stopping is a decision about the next increment, not a verdict on the manuscript.
What this establishes is a decision principle, not a stopping threshold optimized once for all manuscripts. The principle tells a decision-maker what must be supplied before any verdict on further revision follows: an estimate of what another round would change, a declared cost for that work, and some way of weighing the two. The principle supplies none of these; the decision-maker does. Its force is therefore conditional. If the declared costs and the benefit estimate are accepted, the verdict follows; if they are disputed, the verdict inherits the dispute.
Two distinctions keep the scope clear. A modest expected gain is not by itself sufficient reason to stop: when the revision addresses a defect that the declared priorities rank highly, the work may still be warranted. Nor is a large gain sufficient to continue, since the cost of the effort and its alternative uses belong to the comparison. What the principle makes necessary is only that both sides be declared before a choice is made.
The result is also possible rather than inevitable. Stopping at a given point can be rational; it need not be, and a decision to continue is not thereby shown to be wasteful. The quantity compared is anticipated change against declared effort — two judgements attached to one manuscript and one revision goal — not a measured improvement rate that transfers between manuscripts. Editorial effort enters as a separately declared cost dimension rather than as an unspoken constraint.
If the benefit estimate is uncertain — a conditional supposition, not an observed result — the principle calls for showing how far the verdict depends on that estimate. It does not yield a cutoff valid across manuscripts. Holding the task and inputs fixed, the comparison holds; changing either requires the inputs to be declared again.
40.7A decision with declared priorities
The synthetic comparison's dominance relation settles one thing only. Because A and B each outperform C on both declared dimensions, C is dominated and drops out of further comparison. That elimination does not rank A against B: A preserves more distinctions, B incurs fewer arbitrary cost units, and the example supplies no rate at which one dimension may be exchanged for the other. The preference directions—more retained distinctions preferred, less cost preferred—are sufficient to identify domination, but not sufficient to produce a unique choice.
A selection between A and B becomes determinate only once a priority or constraint is declared, and the declaration must do real work. (Conditional supposition.) Suppose the priority is to retain ten distinctions whenever any admissible option does so; A then follows. This is a possible outcome of that rule, not an inevitable consequence of the scorecard. (Conditional supposition.) Suppose instead the constraint is a cost ceiling of four units; B is then the only admissible option. The same pair supports different selections, which shows that the decision depends on the declared rule as much as on the comparisons themselves.
Necessary and sufficient conditions can be separated here. Declaring some priority or constraint is necessary for a justified, determinate ranking of A and B; naming any priority is not sufficient unless it ranks the two or excludes one. A threshold requiring at least nine retained distinctions leaves both admissible, so it narrows the set without deciding. Only a rule that explicitly orders the dimensions or rules out one option yields a selection.
Two qualifications attach to this toy scorecard. Its units are arbitrary, so the comparison of ten retained distinctions against six cost units is not a measured gain, and no missing dimension has been declared unimportant. The chosen rule is therefore a design commitment made explicit before the choice, not a verdict recovered from Pareto efficiency alone.
In the synthetic illustration, editing cost stays its own axis. That separation governs how the surviving options may be compared. Pareto dominance asks only whether one option does at least as well on retained distinctions and on cost, and better on at least one; by that test A and B remain unordered. Collapsing the two axes into a single number introduces a different decision rule, not a sharper version of the same one. A ratio of retained distinctions to cost units would ask how many distinctions an option preserves per unit of editing effort — a rate of return rather than an overall ranking.
Such a ratio needs its own justification because its denominator is the cost axis and its units are the arbitrary cost units of the illustration. The quantity compares distinction-retention per arbitrary cost unit. It compares neither money nor measured hours, and it leaves the separate cost ranking available only as something the ratio has absorbed. If this rule were adopted as a supposition, it would favour B over A, since B spends less while giving up one distinction. That outcome is possible rather than inevitable: it follows from the chosen ratio and from the equal value that a per-unit comparison implicitly assigns to each distinction, and a stated weighting or a cost ceiling could select differently.
Conditions should also be kept distinct. A declared priority — cost minimisation subject to retaining at least some number of distinctions — is sufficient to choose between A and B here, and a cost cap can exclude A while admitting B. Neither dominance nor any ratio is sufficient on its own: dominance only removes C, and a ratio determines nothing until its denominator and the trade-off it assumes are accepted. Cost therefore remains a separate axis unless an additional rule is explicitly justified.