Chapter 1. One Figure, Two Readings: The Stated Saving and the Executed Run

This chapter shows that thirteen authored records read one figure, the minutes of rework, in two ways; that the reading decides whether the record's own run reports a saving or an increase in work; that where the readings part, the record carries the stated saving through every place where it speaks and leaves the run's result unspoken; and that what the records ask for next is measurement, not a reading of the figure. It contributes to the debates on verification, reproducibility and forecasts a reading of what one designed record set takes its built-in check to cover.

1.1Data and method

AEGIS publishes its authored cases as data files of one record a line, in 11 published parts. This study read all of them once, by GET, on 2026-10-06: 6,540 cases. Some of these cases carry a run, a small calculation that a programme executed on figures the author set, with the input, the output, a hash of the input and a replay. The rule this study reads, work-saving-v1, takes a number of cases, the minutes each case takes before a change, the minutes it takes after, and a figure for the minutes of rework, and reports three totals: the existing work, the new work with rework, and the net saving, where a negative value is an increase. Across the 6,540 cases the rule is run 93 times, in 22 programmes.

The selection rule is one: every authored case whose run executes work-saving-v1 and whose situation or question speaks of rework (재작업). It keeps 13 cases. Nine belong to AEGIS-CALC-v1, a programme of 196 cases that AEGIS describes as executing 12 kinds of calculation rule and keeping the input hash, the output and the replay; three belong to a programme of facilities and energy decisions and one to a programme of logistics decisions. None belongs to a programme that earlier AEGIS papers read, and none is among the cases an earlier paper read in AEGIS-CALC-v1. The other 80 runs of the rule are left out because their situations and questions do not speak of rework; the rule keeps every one of the seven runs of this rule in the whole population whose output is an increase, and none of the 80 left out has one.

The records are what AEGIS says they are. All 13 are typed realistic synthetic (현실형 합성). All give the same source role: an authored question seed whose judgment mark is the author's record, not a factual basis from public records. The nine in AEGIS-CALC-v1 share a programme scope which says that every number is a value assumed by the author and has not been compared with public records; the other four carry only the one-line label of their programme. Every run states in its assumptions that its figures are values assumed by the author, and three add that they are not actual measurements. Every run has been executed as pure arithmetic with no randomness and replayed with the same output; every fact check says there was no verification of real events or authority; every prediction performance is "Not measured"; none has a source link. The deciding roles are 13 different officers and managers, four of them marked as a setting, across five domains.

Each record was saved as a snapshot in this study's folder with two fields removed, the source line and the group label, because some AEGIS records name a model there. The snapshots were checked for model names and none remain. The comparison population is kept in the same folder, reduced to the fields measured, with any model name in a programme's identifier shown as "[model]". The records are Korean. This study wrote the English once, sentence by sentence, from a table of translations kept with the study, and each English seed names its Korean original by hash. Every block quote below is that English.

The fields read are these: the situation (context, which AEGIS repeats as the summary), the question, the three choices, the author's judgment mark (review), the needed inputs, the unknown, the condition for reconsidering (reconsider), the observation, and the run's input, output, assumptions, interpretation and replay. Two things are computed by a script from each record and from nothing else. The first is what the run does: whether its new total equals the number of cases times the new minutes plus the rework minutes, which holds for all 13. The second is what the record's own words do with the rework: a record is counted as reading the rework for each case when a figure it states equals the run's net or the net for a single case, and as reading it once for the whole when a figure it states equals the saving with the rework taken off once, in minutes or in whole hours. Every record falls on exactly one side. Every other count is a stated rule on the Korean words, applied by the same script, and every number in this paper comes from it.

Before any field is used as evidence about the set, it is measured for frame: the share of its five-word runs that recur in at least 5% of AEGIS's items. The situations, questions, choices, needed inputs, unknowns, conditions for reconsidering, observations, run interpretations, run assumptions and link reasons are free or nearly free of frame, at 0% to 3% in Korean and at most 11% in this study's English, except the run assumptions, which reach 21% in English because the same disclaimer closes each list; the author's marks stand at 6% in Korean and 10% in English. Three fields are frame and are not read as findings: the uncertain line (61% in Korean), and the condition and counter-evidence attached to every link, which are the same sentences across almost all of AEGIS's links (99%, and 100% among the thirteen). The programme scope and the source role are one sentence each, repeated wherever they stand; they are read once, as what AEGIS says the records are.

1.2One figure, two readings

Everything in this section is read from authored synthetic cases, whose figures, offices and procedures are the author's assumptions. Each case gives four numbers: how many cases a procedure handles, the minutes each takes now, the minutes each would take after the change, and the minutes of rework the change brings. Three of the four have an obvious place in the arithmetic. The fourth does not, and the records show it. The first case, a standard form for computer entry, states its figures and its arithmetic in two sentences:

The author's arithmetic is explicit. Five minutes are saved on each of 800 cases, 4,000 minutes in all, and the rework is taken off once: 40 minutes for the whole reform. The run reads the same figure differently. Its new total is 37,600 minutes, which is 800 cases times the 7 new minutes plus the 40 minutes of rework, so that the rework falls on every case. Against the 9,600 minutes of the existing work, the run reports a net of minus 28,000 minutes: an increase. The author states a saving of 3,960 minutes; the run, on the same input, states an increase seven times as large.

Neither reading is wrong as arithmetic; the sentence admits both. In some records the author's words settle the question. The warehouse case says the rework comes with each job:

The sludge case says the same with the same word. Where the author's words put the rework on each case, the run and the author agree. The maintenance shop's own reading of its run is an instance:

The 35 minutes are the 50 saved on a job less the 15 of rework, and 1,050 is what the run reports. The same holds, by the figures their interpretations give, in the cases of disclosure copies, supply requests, warehouse automation, sludge drying and dispatch planning: six records in all, whose runs report savings. In these six, the rework is between 21% and 32% of the minutes saved on a single case, so that the saving survives it.

In the other seven the author's words take the rework off once, and in every one of them the rework figure is larger than the saving on a single case: 40 minutes against 5 in the computer entry case, 3,600 against 9 in the case of electronic approval. Added to every case, a figure of that size turns the saving into an increase. The electronic approval case states its totals in full:

The record says the rework arises "in the course of the transfer", which reads as a single event, and the author's reading of the result treats it so:

The run's own output for this case is a net of minus 8,618,400 minutes. The interpretation printed beside that output, which is the run's reading of itself, gives the author's 18,000. The pattern repeats in the de-identification case, where the stated result is converted into hours:

The run's net is minus 2,388,000 minutes. In the stock tally case the stated saving is about 253 hours and the run's net minus 1,263,200 minutes. In the four cases of computer entry, the maintenance intake form, the stock survey and the weekly report, the question itself names the author's figure, so that the decision asked of the reader is framed on it. The intake form case shows the shape:

The author takes 120 minutes off a saving of 1,050 and states 930; the run adds 120 minutes to each of 150 cases and reports minus 16,950.

Two features of the split matter. The first is that it is clean. Every one of the thirteen records falls on one side, the six that read the rework for each case all have runs that report a saving, and the seven that read it once all have runs that report an increase. Across all 93 runs of the rule in AEGIS, these seven are the only ones whose output is an increase, and all seven stand in AEGIS-CALC-v1, the programme built around executed calculations. The second is that the agreeing records are not free of the ambiguity; they are only on the side the run happens to take. The supply request case reads its own result this way:

The case's figures give a saving of 17 minutes on each request before rework, and 12 after it. The sentence names 12 minutes as the saving per case, which is the figure after rework, and then mentions the rework of 5 minutes again; read as written, the words could be taken to count the rework twice. The run's total of 10,800 minutes counts it once for each case. The record holds both, and only the arithmetic, not the words, shows which the run used.

This is the point at which the first strand of the debate bears on the records. R. G. Sargent, Verification and Validation of Simulation Models (2013), separates the verification of a model as built from the validity of its data and of its conceptual model. The run in each of these records is verified in a narrow and real sense: it computes what its rule says, on the input it was given, every time. What it does not check is whether "rework of 40 minutes" in the input means what "rework of 40 minutes" in the author's sentence means. In this study's reading, the seven records are a small case of what James H. Fetzer, Program Verification: The Very Idea (1988), calls an equivocation, carried from the program to its data: the formula is a logical structure that can be checked, and the figure it is fed is a description of a procedure that cannot be checked by running the formula again.

1.3What the replay checked, and what the records ask to measure

The cases in this section are, again, authored synthetic settings, and their offices and minutes are the author's assumptions. The nine records of AEGIS-CALC-v1 share one sentence of programme scope that says what the executed calculation is for:

Each run keeps its input hash and a replay, and in all 13 the replay reports the same input and the same output. What the replay shows is the property Roger D. Peng, Reproducible Research in Computational Science (2011), proposes as a minimum standard where full replication is not possible: that the computation, run again, gives the result it gave. In the seven records where the readings part, that minimum is met, and the result it confirms is the one that no sentence of the record repeats. The run's interpretation gives the author's saving in all seven, five times in minutes and twice in hours. No sentence in any of the seven, in its situation, question, choices, mark, unknowns, conditions, observations, assumptions or interpretation, gives the run's net or speaks of a negative result or of the total work increasing. (One condition in the stock tally case speaks of rework rising, which is a different thing.) The only place the increase stands is the output itself, a number under a label.

The author's recommendation follows the stated figure. In five of the seven the recommended option names it as its ground, as in the computer entry case:

In the electronic approval case the author's reason for that option speaks of the rework directly:

The half is the author's: 18,000 minutes against 36,000. Every mark then closes with the same sentence, which keeps the author's judgment apart from the calculation:

The sentence separates the mark from the calculation in so many words, and in the seven records the two then point in opposite directions without remark. Nothing in this study says which direction is right; the records do not say either. What they show is that the built-in check and the stated reasoning run in parallel and never meet. Raymond R. Panko, "What We Know About Spreadsheet Errors" (1998), holds that assumption sections may matter far less than testing after development. The records have an assumption section in every run, and it is careful:

It lists the four figures and disclaims them as measurements. It does not say whether the 40 minutes fall on each case or on the reform. The replay is the only testing the record carries, and it tests repetition.

Where, then, do the records place their doubt? Almost entirely in the world. Twelve of the thirteen ask, among their needed inputs, for a measurement or a record of time or rework; eleven name rework or errors among what is unknown; all thirteen ask that time or rework be recorded after the change; eleven tie the condition for reopening to rework. The alternative the computer entry case sets beside its recommendation is to wait and see:

And its condition for reopening repeats the figure as a threshold:

The other three cases whose questions name the author's figure do the same, with 120, 200 and 300 minutes:

These conditions inherit the ambiguity they are meant to watch: forty minutes of rework, measured in the field, would meet the condition on one reading and fall far short of it on the other, since the run already counts 40 minutes on each of 800 cases. None of the thirteen asks how the figure is to be read.

The records' doubt about the size of the rework is the doubt the third strand of the debate studies. Daniel Kahneman and Dan Lovallo, Timid Choices and Bold Forecasts: A Cognitive Perspective on Risk Taking (1993), and Roger Buehler, Dale Griffin and Michael Ross, Exploring the "Planning Fallacy": Why People Underestimate Their Task Completion Times (1994), both locate optimistic forecasts in plans and scenarios that neglect the record of the past. The cases that come nearest to that concern are the ones whose runs agree with their authors. The dispatch planning case recommends a pilot on the ground that full introduction without measurement produces results different from what is expected, and its interpretation keeps its figure in its place:

Its condition for reopening is stated in terms that admit only one reading:

The warehouse and sludge cases give their thresholds as a rate of rework, 20 and 25 percent, which also has one reading. The disclosure copies case goes furthest, and rules out the option of leaving rework out of the reckoning altogether:

The set, then, knows that rework matters and asks for it to be measured. Winston Harrington, Richard D. Morgenstern and Peter Nelson, On the Accuracy of Regulatory Cost Estimates (2000), trace the difference between their total-cost and unit-cost results to frequent errors in estimating the quantities, the effects of individual rules. In this study's reading, the seven records show the same structure in miniature, with a difference of its own, which is not drawn from that article: the unit figures, five minutes saved and 40 of rework, are stated and clear, and what decides the total is a quantity the records never state, the number of times the rework is counted.

This chapter has shown that one figure is read two ways in the thirteen records, that the reading decides the sign of the run's result, that the records carry the stated saving through every sentence while the run's result stands only as a number, and that the records place their doubt in the size of the rework rather than in its reading. Its contribution to the debates is to show, in a designed record set where every step is visible, a built-in check that verifies the arithmetic and a reasoning that never consults it.