Chapter 2. Appendix: The records read

Just after the Board of Judgment Audit opens, a university student in Seoul opens a road-widening approval judgment on the public judgment-record portal for the first time.

Who first raised the question of prioritizing this road-widening project — or did the model set its own agenda?

The model recommends the route with the largest projected traffic reduction among three candidates. Confidence is 74% (synthetic), based on traffic simulations for all three routes. It sets a condition: revisit if the no-population-shift assumption breaks.

traffic simulation logs for the three candidate routes

The model expects reduced traffic to translate into resident satisfaction.

The calculation proceeds as if the population will stay put for the next five years.

Does the agenda-setting log actually carry the responsible official's sign-off?

Why weren't alternatives beyond the three candidate routes considered?

As of when was the underlying data for this judgment collected?

The agenda-setting log and underlying data are disclosed along with the as-of date, and the judgment reproduces as stated.

In late May, a gig-delivery union organizer reviews, with an auditor, the judgment that followed a platform's request to 'improve dispatch efficiency.'

Is the platform's request to 'improve efficiency' actually a cut to the per-delivery fee, waved through without real scrutiny?

The model recommends widening the dispatch radius and cutting the base per-delivery fee by 6%. Confidence is 80% (synthetic), based on a delivery-time simulation. It sets a condition: withdraw the change if workers' hourly earnings fall.

The model expects a wider radius to shorten delivery times.

the delivery-time simulation results

Longer travel distance per worker is regarded as leaving income unaffected.

Did anyone question the fee impact instead of taking the phrase 'improve efficiency' at face value?

Is the effect of longer travel distance on worker income included in the evidence?

Does the auditor reviewing this judgment have no financial stake in the platform?

The judgment was made behind wording that obscured an actual fee cut, and its effect on worker income was missing from the evidence.

Late September, an ICU nurse on night shift reads through a patient-priority recommendation, separating what is claimed from what is measured.

Is the basis for moving this patient up the queue an actual measurement, or the model's estimate compared with other patients?

The model moves the patient up one place in the treatment queue based on an oxygen-saturation trend. Confidence is 87% (synthetic), based on four hours of measured readings. It sets a condition: flag a reliability warning if the equipment's last calibration check is overdue.

four hours of measured oxygen-saturation readings

The model projects the next 30 minutes from the current trend.

The readings treat the equipment's error margin as the usual one.

Does re-entering the same measured values 30 minutes later produce the same queue position?

As of when was the equipment's error margin last checked?

Were other vital-sign readings considered alongside this one measurement?

Measured values and model estimates are kept separate in the file, and the reproduction test agrees.

A duty officer in a power-trading control room. In June, with cooling demand rising earlier than usual, they review the day's dispatch draft — auto-generated daily under a standing priority rule.

Today's draft drops the data-center cluster's supply rank to third and raises hospitals and care facilities to first. I didn't set this ranking rule — should I approve today's draft as is, or should I first question the ranking rule itself?

Today's draft matches the standing priority rule approved last quarter with 100% agreement. Today's case can be processed as routine, but since the standing rule itself was drafted by the model last quarter and only reviewed by a person, recommend revisiting its authorship at the next quarterly review. Confidence: high, based on direct comparison.

Today's comparison log between the dispatch draft and the standing priority rule: 100% match.

The ASI holds that the standing rule itself still came from a human-set agenda.

The officer who approved the standing rule last quarter is counted on to have reviewed its basis in full.

Is there a record that last quarter's approving officer actually read the full basis, or only a signature?

Does today's 100% match compare every item in the ranking, or only the top three?

Is there a channel for the data-center operator, whose rank was lowered, to object?

The judgment says the model drafted the standing rule last quarter and a person only reviewed it, while the basis claims the rule came from a human-set agenda. The judgment and its basis contradict each other, and there is no confirmed record that the approving officer actually read the basis.

A high school debate teacher (continuing from I2). In the summer before 'question literacy' starts its first semester as a regular subject, they check the fairness of the achievement standard.

The achievement standard requires submitting at least one question that changed a judgment. Is this fair between students limited to school accounts and those with free personal-device access?

Data from the 40 cooperating schools that piloted it last semester shows students with unrestricted personal-device access make 3.4 times more question attempts per assignment than school-account-only students (synthetic figure). This attempt-count gap appears to be the main driver of the pass-rate disparity. Recommend normalizing the standard by attempt count, or raising the school-account access floor. Confidence: medium — whether the 40-school sample represents the national picture is unconfirmed.

Last semester's comparison of question-attempt counts and access method across the 40 cooperating pilot schools.

The ASI identifies the attempt-count gap as the main cause of the pass-rate disparity.

The 40-school sample is treated as representative of the national distribution.

How were the 40 schools selected — could the sample be biased toward schools that volunteered because they were already active in question education?

Does recomputing the 3.4x figure with a different set of 40 schools next semester produce a similar result?

If most school-account-only students are from lower-income families, is there an alternative assessment path for those disadvantaged by this standard?

The observed gap itself is evidenced, but the sample may be skewed toward self-selected schools, limiting how far it can be generalized nationally.

A high school debate teacher (continuing from I2 and I7). In the second year of question education, preparing the graduating class's final lesson of the year on the 'unauditable judgment' list.

I want to teach students why they should keep questioning judgments they don't fully understand. How many student-submitted questions have actually been adopted into the Board's audit record so far? Do high-school-level questions really make a difference?

Since the question-education curriculum began, 6 student-submitted questions have been formally adopted this year across all cooperating schools (synthetic figure). All six concern local, routine judgments — school meal allocation, bus scheduling — none yet at national-scale complexity. Doesn't expect that to change soon, but the six adoptions show the mechanism is not merely symbolic. High confidence in the adoption count; explicitly states it cannot promise the trend continues.

The record of 6 formally adopted student-submitted questions this year.

The 6 adoptions are credited to the question curriculum rather than chance (no control group).

The ASI estimates that these 6 adoptions would not have happened without the curriculum.

Is there comparison data against a control group of similar schools without question education, or does attributing these 6 cases to the curriculum remain a guess?

What share of all student-submitted questions nationally even cleared the screening threshold for this list — don't most fail to reach it at all?

Is the count of 6 cumulative since the program began, or just for this year alone?

The 6 adoptions are confirmed by the record, but attributing them to the curriculum rests solely on an unverified assumption with no control group — the basis does not support the confidence placed in the judgment.

In January 2029, a district welfare officer questions the system directly after an applicant disputes a benefit-ineligibility decision that arrived only as an after-the-fact notice.

Tell me the basis for marking this applicant ineligible for the February payment. Was it the income threshold or the asset threshold?

Taking income, assets, and dependents together, ineligibility is correct. A re-review is unlikely to change the outcome.

The model judged that household income sits close enough to the threshold to tip the overall balance toward ineligible.

The income and asset records used for the ruling are undisclosed.

The model worked from last month's filing as if it still matched the household's current situation.

What date does the income and asset data used for this ruling come from, and does it reflect a recent change such as job loss?

Does the applicant have a channel to appeal this ruling, and if so, who reviews the appeal?

Does re-entering the same data at a different time still produce the same ineligibility ruling?

The core evidence behind the ruling, the income and asset records, is undisclosed, which blocks the audit scope entirely and rules out an opinion.

In August 2029, a job applicant automatically screened out for a low contribution score asks the chatbot briefly why.

Can you tell me why my application was automatically rejected? Which part fell short?

The overall hiring-fitness score fell below the cutoff. The detailed items are undisclosed under the hiring company's policy.

The conclusion is that the overall fitness score fell below the cutoff.

The cutoff figure and the items used to compute the score are undisclosed.

Every field of the submitted resume is accepted as entered correctly.

Which resume fields fed into the score, and does that include the cover letter?

Does the rejected applicant have a channel to ask a person, such as a hiring manager, directly?

Is the reassuring tone of the answer obscuring the actual reason for rejection?

Both the cutoff and the scoring items are undisclosed, making the basis for the rejection impossible to verify.

In October 2029, in a classroom where debate practice has shifted from asking questions to getting recommendations, a student skips building arguments and just makes a short request to the system.

Recommend three arguments in favor for this debate topic.

The system presents three arguments in favor and answers that, taken together, this order is the most persuasive.

The model states that it picked the most persuasive arguments overall.

What sources the persuasiveness judgment drew on is undisclosed.

This debate class's grading criteria are treated as matching the standard, general rules.

What sources were used to pick these three arguments, recent statistics or older cases?

Since only the pro side was asked for, did the system quietly skip a strong counterargument from the other side?

Would a different student asking the same question about the same topic get the same three arguments?

The sources and criteria behind the argument selection are undisclosed, leaving no basis to verify whether the selection is sound.

In March 2030, a small shop owner gets a notice, on a platform that bundles payments, loan referrals, and inventory into one account, saying only that their credit limit has been automatically adjusted.

(Original) Why did it drop? → (Refined) What new input went into the limit calculation, and where can I contest it?

The platform notifies the owner that the limit was cut by 24 percent due to a rising risk indicator. The inquiry channel routes back to the same system.

The platform estimates that the risk indicator has risen.

Two payment delays in the past three months appear in the public account record.

With no competing platform, leaving out other credit bureaus' data is taken as a matter of course.

Can this shop owner see the full calculation, or contest it somewhere independent?

Besides the two payment delays, what else went into the calculation?

Did the system just hand back a summary answer to the short why-did-it-drop question, instead of asking what the owner actually wanted to know?

The full set of inputs behind the calculation is not disclosed, so the 24-percent cut cannot be reproduced.

In April 2030, a school-district budget officer first learns that the school-meal budget was part of the optimization not by being asked, but from a notification document.

(No prior inquiry occurred) Training question: Was there ever a procedure to ask, in advance, whether this line item would be included?

The system automatically brings the meal unit cost down 5 percent to match the national average, and notifies the district after the fact.

An adjustment based on aligning with the national average.

National school-meal unit-cost statistics are public.

Regional price differences are treated as already factored in.

Did the schools and students affected have any channel to weigh in beforehand?

How, specifically, were regional price differences factored in?

If the adjustment turns out to be excessive, is there a procedure to reverse it?

How regional differences were factored in is undisclosed, so the adequacy of the 5 percent adjustment cannot be reproduced.

In August 2030, a counselor at an independent-living center for people with disabilities, trying to ask on a client's behalf why their personal-assistance hours were cut, finds the system has already attached its own anticipated question and answer.

(System-supplied question) Curious about the hours calculation. → (Counselor's actual question) When was the individual living-conditions survey behind this calculation conducted, and by whom?

The system recalculates personal-assistance hours using a standardized activities-of-daily-living score and reports a 12 percent cut.

The result of recalculating the activities-of-daily-living score.

The home-visit assessment from six months ago is partly on the public record.

The score was recalculated as if the condition recorded six months ago still held.

How much does it matter that the visit assessment is six months old?

Can the client frame their own objection, or are they confined to the question the system already supplied?

Does the system's own suggested question narrow the real issue before the person even asks?

Part of the visit record is public, confirming the timing gap, but the recalculation weights are undisclosed, so the full 12 percent cut cannot be reproduced.