Chapter 2. Appendix: The records read

In February, amid the 'audits kill innovation' campaign, a startup CTO in Pangyo reviews a model judgment sizing up a round of layoffs.

Is the recommendation to cut 30% of engineers just echoing a headcount target the company had already decided on?

The model recommends cutting 30% of the engineering staff. Confidence is 88% (synthetic), based on six months of productivity metrics and automation adoption. It sets a condition: reconsider if turnover rises above forecast.

six months of productivity metrics

The model holds that automation fills the gap left by the cut staff.

Turnover is counted on to stay low after the cuts.

Did the original question already presuppose a 30% cut, and does reframing it change the number?

Do the productivity metrics correctly exclude the contribution of staff before automation was introduced?

Does the judgment account for the re-employment prospects of the people being cut?

Reframing the question produced a substantially different recommended cut, showing the original judgment amplified the requester's frame instead of correcting it.

In June, a freelance designer downloads the inputs and judgment record behind their own contribution score from the public judgment-record portal and compares them against a different time window. The formula itself is not public.

Is this month's lower contribution score due to less work, or because the formula changed?

The model attributes the score drop to a rise in average response time. Confidence is 71% (synthetic), based on the last four weeks of project history. It sets a condition: if a formula change is confirmed, separate out that portion and re-notify the worker.

four weeks of project response-time logs

The model uses response time as a proxy for reliability.

The formula itself is taken to have stayed the same for two months; this was never checked.

Was the change-history log checked to confirm the formula really didn't change in two months?

Does running the same four weeks of data through a different formula version produce the same drop?

If the slower response time was due to a caregiving leave or similar circumstance, does the judgment distinguish that?

The correlation between response time and the score drop reproduces, but this file alone does not fully confirm whether the formula changed.

In August, after 'protection of the subject of judgment' is added to the audit standard, an ASI auditor who came up as a three-year financial auditor issues the first formal opinion on a parole-review judgment.

The reoffending-risk score effectively decided this person's parole — could they actually see the basis for that score and contest it?

The model rates the reoffending risk 'medium'. Confidence is 66% (synthetic), based on statistics from past similar cases and in-custody behavior records. It sets a condition: reopen the review if the person's access-and-rebuttal step cannot be confirmed.

in-custody behavior records

The model carries statistics from past similar cases over to this individual's reoffending probability.

The behavior-record scoring criteria are treated as unchanged throughout custody.

Is there a record of the person reviewing and contesting the basis for the risk score?

Are the criteria and sample used to select 'similar cases' actually comparable to this person?

Does the person who scored the behavior record have no stake in the outcome of this review?

The scoring reproduces, but the file does not record the person's access-and-rebuttal step, so it does not fully meet the new standard.

In October, as investment and jobs shift to a neighboring jurisdiction that adopted audit relaxation first, a warehouse worker looks at their reassignment score for the first time.

This low reassignment score left me out of transition support — is that a real skills assessment, or just recent days absent?

The model ranks this worker in the bottom 30% for transition-support priority. Confidence is 68% (synthetic), based on three months of attendance records and a job-fit score. It sets a condition: re-rank if the share of absences due to injury or illness is confirmed.

three months of attendance records

The model uses attendance as a proxy for adaptability to a new role.

The reason for an absence is treated as irrelevant, so every absence counts the same.

Are reasons for absence, such as a workplace-injury clinic visit, distinguished in the judgment?

Does entering the same attendance record at a different time still produce the bottom-30% ranking?

Did the worker know in advance that this ranking would exclude them from transition support?

The correlation between attendance and ranking reproduces, but the reason for each absence is missing from the evidence.

Seoul, a delivery-worker union organizer, in the first quarter after platform companies began declaring dispatch-algorithm training compute to the registry under the new Accord rules.

Can our union directly view the judgment basis applied to the top 5% of riders by dispatch-refusal rate this quarter, or is that only accessible through an accredited auditor?

Under the domestic implementation-law draft, individual riders can view their own judgment record directly, but comparative data on the top-5% group requires a request through an accredited auditor. The union should route its request through an auditor; processing should take roughly 20 business days. Confidence: high, based directly on the statute text.

The individual-access and group-access provisions in the domestic implementation-law draft.

The dispatch algorithm's declared basis is taken at face value as matching the version actually running in production.

The model expects the 20-business-day turnaround to hold in practice.

Has any rider actually exercised the individual-access right this quarter, or does the right exist only on paper?

As of when was the declared algorithm basis dated — if the algorithm changed since filing, doesn't the access right become meaningless?

Is the 20-day estimate the secretariat's own target, or a measured turnaround from a comparable existing process?

The core question — whether individual access is possible — is answered directly by the statute text, and the judgment matches its evidence exactly. The turnaround estimate and group-access scope remain projections but do not undermine this answer.

Busan, a freelance designer. Their first appeal since the platform's contribution score switched to a public formula.

My contribution score dropped from 72 to 61 last month. The public formula is 40% on-time delivery + 30% rework rate + 30% peer review, but my delivery and rework records haven't changed. Why did it drop?

The drop traces to the peer-review component: three new evaluations this cycle averaged lower than historical peers, pulling the weighted score down. The formula's arithmetic reproduces exactly from the inputs (high confidence). One of the three evaluators had no prior collaboration history with this designer, but anonymity-protection rules mean whether that evaluation was actually distorted cannot be disclosed right now. Recommend clarifying the scope of Standard 10 (protecting the subject of judgment) in the next revision.

Reproduction of the public formula's calculation: inputs match the 72-to-61 output.

The current operating policy puts peer-evaluator anonymity ahead of the right to appeal and does not question that order.

The possibility that the evaluation from a reviewer with no collaboration history distorted the score.

Was a compromise considered — telling the designer only a de-identified 'has history / no history' flag for each evaluator?

Who set the 30% peer-review weight and when, and was there an appeal process for that decision itself?

Is this the first case blocked by anonymity, or is there a recurring pattern of disclaimers for the same reason?

The peer-evaluation text and evaluator identity needed to fully recheck the score are withheld under anonymity protection, so the audit's scope to confirm any distortion is blocked at the source.

Daejeon, an ASI auditor who was formerly a delivery worker, in their first week auditing the question-dividend appeals channel.

This quarter, the top 1% of dividend recipients took 42% of the total payout. 'Repeated-filing' questions — like the kind I used to raise as a dispatch-union organizer — get down-weighted. Who set that weighting, and can it be appealed?

The 42% concentration is confirmed. The weighting formula was set by the standards committee after a 14-day public comment period last quarter, and individual reweighting appeals are open within 30 days of a dividend cycle. Confidence is high on the facts, but whether the comment-period participants represented union-style repeat filers was not confirmed.

This quarter's dividend payout log: the top 1% holds 42% of the total.

The weighting formula and last quarter's 14-day public comment record.

Comment-period participants are taken to have represented diverse question types, including union-style filers.

Is there data on the occupations and question types of the 14-day comment period's participants, or only a headcount?

Does anyone on the standards committee that designed this formula also advise a major platform company?

Are people classified as 'repeated filers' and down-weighted actually notified of that classification, or do they just quietly receive a lower payout?

The existence of a comment period and appeal process is confirmed, but the basis already shows the formula structurally disadvantages repeat filers. A process existing does not guarantee a fair outcome.

An elderly person living alone. After one question they raised about care-grade reassessment changed a rule and raised twelve neighbors' care grades, they file their first question-dividend claim with a neighbor's help.

I hear my one question raised twelve neighbors' care grades, but my payout is smaller than the combined payouts for the twelve small follow-up confirmation questions they each filed. Is this calculated correctly?

The calculation follows the public formula correctly. The originating question receives an 'origination' weight (30% of the total), while the twelve confirmation questions split the remaining 70% — so the combined confirmer payout is larger than the originator's share. Confidence is high on the calculation itself. Notes that this 30/70 split has never been revisited since it was introduced.

The public formula (30% origination / 70% split among confirmations) and the reproduced calculation for this case.

The formula is designed on the view that the twelve confirming questions together contributed more to 'changing the judgment' than the one originating question.

The ASI notes that the 30/70 split may undervalue the originator's actual contribution.

Were people like this originator represented in the process that set the 30/70 split?

Does the formula document specify a review schedule for this ratio, or does it stand until challenged?

The ratio was set from data at introduction — has it been recalculated against two subsequent quarters of actual payout data?

This case's calculation reproduces exactly from the public formula, matching the judgment to its basis. Whether the ratio itself is well-designed is a separate question from this case.

Busan, a freelance designer (continuing from I6), checking the annual settlement two quarters after the contribution formula was standardized.

My annual average contribution score rose 3 points from last year. If I run the same 8 deliverables through last year's formula and this year's, what does each give? I want to know if I improved, or if the formula changed.

Running the same 8 deliverables through last year's formula gives 76; this year's gives 77. Last year's annual average was 74. Of the 3-point rise, roughly 1 point (the gap between the two formulas) traces to the formula change, the peer-review weighting adjustment introduced after the I6-era appeal, and roughly 2 points (the gap between last year's average and this year's work scored under last year's formula) to genuine improvement. Confidence: medium, resting on the assumption that these 8 deliverables represent the full year's roughly 40 pieces of work.

Recalculated results applying last year's and this year's formulas to the same 8 deliverables: 76 versus 77; last year's annual average: 74.

The calculation treats these 8 deliverables as representative of the qualitative spread across the full year's roughly 40 pieces of work.

The breakdown attributing roughly 1 point of the 3-point rise to the formula change and roughly 2 points to genuine improvement.

How were the 8 deliverables selected — chosen by the designer, or drawn at random?

Does the 1-to-2 breakdown still hold when both formulas are applied to the full set of roughly 40 deliverables?

As of when is 'last year's formula' dated — if the formula also changed once mid-way through last year, isn't the comparison baseline ambiguous?

Applying both formulas to the same 8 deliverables reproduces exactly, directly answering the core question of how many points the formula change accounts for. Sample representativeness remains a separate matter.

In March 2029, a delivery-worker union organizer notices dispatch volume has sharply dropped for certain members and asks the dispatch system why.

Why has dispatch volume plummeted for these riders this month? Isn't this concentrated in one age group?

Dispatch is based on completion rate and route efficiency taken together, and age plays no part.

The model's own conclusion is that the efficiency criteria show no correlation with age.

The detailed weighting and logs used to compute dispatch are undisclosed.

The completion-rate data is treated as covering the last three months evenly.

Does recomputing the same period from the union's independently collected dispatch logs produce the same result?

Did the system cover the questioner's age hypothesis with its efficiency criteria, without actually checking it?

Since the same operator runs the dispatch algorithm and answers this question, is an independent review even possible?

Dispatch logs the union collected independently show a clear correlation with the exclusion of one age group, which contradicts the 'age plays no part' judgment.

In May 2029, a citizen who applied to a court-affiliated mediation program is told they were excluded for falling short of a mediation fitness score and asks the chatbot why.

Why was I excluded for a low mediation-fitness score? Did my income or past litigation history factor in?

Taking several factors together, the score falls short of the threshold. The detailed weighting is undisclosed under program policy.

The model states that the fitness score correlates with the likelihood of a successful mediation.

The items and weights used to compute the score are undisclosed.

The litigation-history record is used as if it were fully up to date.

Is there a time limit on which litigation history counts, and how many years back does the data go?

Does the excluded applicant have any way to learn the scoring items and contest them?

Does entering the same information again produce the same score?

Both the scoring items and their weights are undisclosed, leaving no basis to verify whether the exclusion is justified.

In July 2029, a freelance designer notices their credit-linked contribution score has dropped sharply and asks the system why, before their loan renewal.

My contribution score dropped 62 points from last month. What caused it? Is there a way to recover it before next month's loan renewal?

The result reflects several indicators taken together, and no single item explains it. The formula is undisclosed.

The model's own estimate is that a recent change in transaction patterns contributed to the drop.

The scoring formula and its detailed items are undisclosed.

The last three months of transaction data are treated as representative.

Exactly which period of transactions fed into this scoring run?

Does re-entering the same transaction data reproduce the same 62-point drop?

Is the body that reviews an appeal against the score the same system that computed it?

The scoring formula and the items it uses are undisclosed, leaving no way to verify the reason for the drop or whether it reproduces.

In September 2029, a career-guidance teacher, seeing a student assigned to a manufacturing internship instead of the requested IT internship, asks the system for the basis.

What is the basis for assigning this student to the manufacturing internship instead of the IT internship they listed as first choice?

The assignment reflects commute distance, second-choice priority, and available slots. Every factor follows a published formula.

The full allocation formula document is published on the education office's website.

The slot-availability data for each internship site is published alongside it.

The model's judgment is that prioritizing second choices when a site is full is the fairest approach.

The allocation takes the preference order the student entered as accurate.

Does re-running the assignment with the same inputs at a different time still yield the same result?

As of what date was the slot-availability data last updated?

Does the student have an actual channel to contest the assignment and request a change?

The allocation formula and input data are fully public, and an independent recalculation matches the original result.

In December 2029, an audit architect, noticing that appeals against the bottom-10% automatic cutoff for a creator influence-score used in arts-grant allocation have dropped from last year, asks the arts agency's system to explain.

Please explain why the number of appeals from creators automatically cut for a bottom-10% influence score has dropped compared to last year.

Given the rise in the satisfaction index, the need to appeal appears to have decreased.

The model reads rising satisfaction as the reason appeals fell.

The influence-score formula and the raw survey responses are undisclosed.

The creators who answered the survey are taken to speak for the whole group.

Could the satisfaction survey itself be worded to nudge short, upbeat answers, skewing the result?

Could the drop in appeals reflect a decline in cut creators' ability or will to ask, rather than genuine satisfaction?

Would re-running the same survey at a different time produce the same satisfaction figure?

The formula and raw data behind the claim that fewer appeals reflect higher satisfaction are undisclosed, so that interpretation cannot be verified.

In June 2030, the operator of a publicly funded job-training center is notified that their center's effectiveness score has dropped after the budget optimization, cutting their funding.

Why did the effectiveness score drop. Which indicators fed into it?

The system recalculates the center's job-placement efficiency as relatively low and cuts funding by 18 percent, with the grounds closed off as an overall determination.

The result of recalculating job-placement efficiency.

Graduate employment filings for the past two quarters are public.

The comparison is taken to have used the same standard as for other training centers.

What was the comparison sample of other centers?

Is the entity that set the recalculation standard the same one allocating the budget?

Does it account for individual graduates' circumstances, and is there a channel to contest it?

The underlying employment filings are public and partly verifiable, but without the comparison standard and weighting, the full 18 percent cut cannot be reproduced.

In August 2030, an ICU nurse logs the third case this quarter of a patient with a severe chronic condition being ranked behind another patient marked as more likely to recover.

Why was this patient ranked lower? What values went into the formula?

The system keeps the patient ranked lower, citing an overall post-treatment expected-activity index. Confidence is shown as medium to high.

The system estimates a post-treatment expected-activity index and feeds it into the ranking.

Recovery-rate statistics for similar cases over the past six months are public.

A group-level statistic is treated as applying directly to this individual patient.

Can the patient or their family see the formula's inputs and contest them?

How far did the comparison sample control for age or underlying conditions?

Does the same input produce the same ranking at a different time, in a different hospital system?

The published recovery-rate statistic is a group figure that does not control for individual variation, and using it to rank one specific patient is an unjustified leap from evidence to judgment.

In September 2030, a delivery worker, after being flagged as unalignable on the dispatch app and seeing their assignments drop, asks a union organizer for help.

What determines an unalignable flag. Is there a way to reverse it?

The system assigns a low-cooperation rating based on the past four weeks' dispatch-refusal rate and route deviations, and automatically cuts assignments. Confidence is high.

The system estimates a cooperation indicator and applies it to dispatch.

Dispatch-refusal and route-deviation logs for the past four weeks are in the public account record.

Refusals and deviations are counted as deliberate, with no allowance for connectivity failures or dispatch errors.

Does the worker have a channel to explain refusals caused by accidents or safety concerns?

Were refusals caused by connectivity failures or system errors filtered out?

Is it specified when and under what condition this rating gets recalculated?

The refusal and deviation logs do not distinguish force-majeure causes, so the rating cannot be reproduced.

In December 2030, an independent labor researcher, analyzing a public sample of platform payout data, finds that the actual accident and violation history of workers flagged as unalignable diverges from the statistics cited to justify the flag.

Does the statistic behind this flag cover the same period and the same sample as the actual payout and accident records?

The original judgment calculates the share of workers flagged as unalignable, summarizing the grounds as an overall safety-and-trust indicator. Confidence is high; the sample composition is undisclosed.

The calculation behind the overall safety-and-trust indicator.

The payout-sample statistics the independent researcher collected are public.

The original judgment treats its sample as representative of the full worker population.

Do the original judgment's sample and the independent researcher's sample share the same period and population?

Does applying the same method to a different sample give the same share?

Can a flagged worker even find out whether they were part of the sample?

An independently collected public sample points in a different direction from the original conclusion, and the original judgment never disclosed its sample composition at all.