Chapter 2. Appendix: The records read

In April, after a high-school debate class became an 'ask ASI' practicum (questioning today's models to prepare for the ASI to come), a student's question actually changes a school timetable judgment and lands in the public 'question bank.'

Does the timetable that clusters lab sessions for third-year science students account for how long those students commute?

The original judgment recommends the timetable that maximizes lab utilization. Confidence is 91% (synthetic), based only on lab booking logs; commute time is not included. No revision condition was attached.

lab booking logs

The model reads higher utilization as higher learning efficiency.

All students are treated as having similar commutes.

Does adding a variable that was missing from the original data, like commute time, change the assignment?

Does the student who ends up going home latest have a way to object?

Is there a stated condition that the recalculated standard will still apply next term?

The utilization calculation reproduces from the lab booking logs, but commute time is left out, the assumption that all students commute alike is unverified, and no revision condition is attached, so the judgment reproduces only within the scope of utilization.

In July, a tracer following the audit file for a home-support recertification for elderly people living alone finds an unusually high rejection rate among people who applied on paper.

Why does the rejection rate differ so much between online and paper applicants, and can this judgment really be called independent of filing method?

The model explains that it screened out applicants based on income and asset records. Confidence is 79% (synthetic). Its basis, however, also includes phone-verification history and online complaint records. It sets a condition: redesign the intake process if the rejection-rate gap between filing methods widens.

income and asset records linked to tax and property systems

The model states that the judgment has nothing to do with how the application was filed.

Having no phone-verification history or online complaint record is regarded as making no difference to the judgment.

When an applicant has no online record, does the judgment read that blank as 'unconfirmed' or as a sign of ineligibility?

How many of the rejected applicants were actually told how to object?

Does a reproduction test that varies only the filing method, holding everything else constant, still show the rejection-rate gap?

A reproduction test confirmed that reading the absence of online records as a sign of ineligibility, not eligibility itself, mainly drove the rejection-rate gap — the judgment did not match its evidence. What was never measured was treated as a defect.

In November, with a neighboring jurisdiction's audit-relaxation bill already passed, a calibrator is asked to price out the reproduction cost for a transmission-line expansion judgment.

Compared with a neighboring jurisdiction that processed the same kind of grid expansion six times faster without an audit, is there any basis for skipping the reproduction test here?

The model recommends priority expansion of one grid segment. Confidence is 77% (synthetic), based on load forecasts and equipment aging data. It sets a condition: stop skipping reproduction tests if the skip-versus-run comparison shows a large gap.

equipment aging inspection records

The model expects the current load-growth rate to continue.

Skipping the reproduction test is taken to leave the result's accuracy intact.

Has the skip-versus-run comparison actually been done, even once?

As of when is the equipment aging inspection data?

Is there a preset threshold — some percentage difference — that would stop the skip-test practice?

In this file, the group that actually ran the reproduction test picked a different priority segment than the group that skipped it, showing the skip changes the outcome.

A welfare officer at a district office (continuing from T1). With the median correction lag now down to 7 days, they check the status of an appeal against last month's basic-livelihood benefit denial.

It's been 6 days since the appeal was filed with no result yet, against a 7-day standard. Is this within the normal range, or should it be flagged as a delay?

Within this quarter's distribution for similar cases involving offline documentation, this case is slower than the median but still inside the 75th percentile. This matches the pattern that offline paperwork typically adds 2-3 days; predicts completion within 1-2 more days (medium confidence). Recommend escalating to priority handling if it passes day 9 (the 90th percentile).

This quarter's processing-time distribution for similar cases involving offline documentation.

The ASI estimates that this case is delayed specifically because of its offline documentation.

Placing this applicant's case in the 'similar cases' distribution is taken as correct.

If this applicant filed offline due to age or disability, isn't this a recurrence of the 'offline applicants disadvantaged' pattern flagged in the I3 case?

How many cases make up the 'similar cases' distribution? If few involve offline paperwork this quarter, the prediction's reliability drops.

The day-9 priority threshold is set from this quarter's data — will the same threshold apply next quarter, or does it keep being recalibrated?

The processing-time distribution is a real basis, but the sample size and whether this applicant falls under I3 protection are not confirmed in the record, so the judgment reproduces only partially.

A welfare officer at a district office (continuing from T1, I1, I6). With the median correction lag now down to 4 days, they draft the year-end audit report.

Of 227 welfare-decision appeals this year, 41 took more than 4 days. Do they share a common cause, or is it random scatter?

33 of the 41 delayed cases (80%) share a common cause: delays confirming residency history with another municipality. Judges the lack of a direct data link to be the structural cause and recommends prioritizing an inter-municipal data-sharing fix over general process speed-ups. Confidence: high, based on the concentration (33 of 41).

Cause classification of the 41 delayed cases: 33 attributed to delayed cross-municipality residency checks.

The ASI points to the missing data link as the main structural cause of the delays.

The remaining 8 cases are treated as delayed for unrelated, individual reasons rather than a shared pattern.

Was the 33-case classification reviewed by a person, or is it purely an automated cause classification?

Was cross-municipality delay also the leading cause in the same period last year, or is this specific to this year?

Could any of the remaining 8 cases involve offline or elderly applicants, as in the I3 case?

The 80% concentration is confirmed by the classification data, but whether that classification was human-reviewed and what the remaining 8 cases involve are not specified in the basis, so the full conclusion cannot yet be fully trusted.

In March 2029, an access-control officer at a power-plant contractor asks the system why a contract worker's access badge was reclassified as high-risk.

What is the reason this worker's badge was reclassified as high-risk, what is the review process, and what is the deadline to request a review?

The reclassification reflects an updated threat model overall and is sound. Review is handled through the supplier's internal process, with no fixed deadline.

The model judged that the worker's recent movement pattern partly overlaps with the threat model's high-risk indicators.

The threat model's detailed indicators and the decision log are classified and undisclosed.

Whether the movement-pattern record really belongs to this worker was never checked; it was used as is.

Under what condition does this high-risk rating drop back down, and when is the next scheduled review?

Does the reclassified worker have an actual channel to appeal to a person, or is the process only nominal?

Is the body reviewing the appeal the same system that assigned the original rating?

The threat model's indicators and decision log are sealed as classified, leaving the basis for the reclassification unverifiable.

In May 2029, an ICU nurse notices a patient's bed-priority ranking came out low and asks the system whether it can be recalculated.

Why is this patient ranked fourth for bed priority, and was their latest oxygen-saturation reading factored in?

Weighing recovery likelihood, severity, and wait time together, fourth place is correct. There is no plan to recalculate at this time.

The model's own explanation is that the recovery-likelihood estimate drove the ranking.

The oxygen-saturation figure entered is taken to be from within the last hour; its timestamp was not checked.

The detailed indicators and weights used to compute the ranking are undisclosed.

Does entering the same figures again now still yield fourth place, or has the ranking already shifted?

Is there a defined threshold for what change triggers a recalculation, and does anyone outside the supplier know it?

Do the patient or their family have a direct channel to contest this ranking?

Both the detailed ranking indicators and the recalculation threshold are undisclosed, making it impossible to check whether the judgment reproduces.

In October 2029, an ICU nurse, noticing a family member simply asked for the best option and moved on, instead feeds the system an extreme condition to test it.

If this patient's oxygen saturation had dropped 5 percentage points further than yesterday's baseline, would the same treatment priority still have been recommended?

The current priority would have held even then. The overall balance against other patients would not have changed much.

The model's own estimate is that the overall balance would not shift much.

Whether that condition was actually recalculated, and the log of it, are undisclosed.

The calculation holds other patients' conditions fixed in the meantime.

Is this answer from an actual recalculation, or was it just an estimate?

Exactly how many percentage points must saturation drop before priority actually changes, and does anyone outside the supplier know that threshold?

Did the system simply confirm the nurse's hypothetical without actually checking it?

Whether the counterfactual was actually recalculated, and any log of it, are undisclosed, making it impossible to verify the answer's validity.

In May 2030, a welfare caseworker notices that their program's budget execution rate has dropped from the previous quarter and tries to find out whether this is part of the optimization result.

Is the welfare-budget cut part of this optimization. Is there a procedure to readjust it?

The system replies only that spending was reallocated under an overall determination, and directs any objection to the next quarterly reassessment.

The system estimates that the reallocation preserves the overall welfare performance indicator.

The aggregate execution rate is public.

The performance indicator is used as a proxy for what recipients actually experience.

Is there recipient experience that the execution-rate statistic simply does not capture?

Does the data behind the performance indicator predate this quarter?

Would statistics from a different time or a different agency support the same conclusion?

Without the detailed grounds for the reallocation, there is no way to confirm that the welfare cut and the maintained performance indicator are both true.

In July 2030, an outreach worker for homeless services watches a station-square surveillance camera repeatedly flag one person as high-risk and issue another move-along notice.

Why does this person keep getting flagged as high-risk. Is there any way to clear the flag?

The system judges that maintaining a high-risk rating is warranted, based on overnight-presence patterns and past complaints, and keeps sending move-along notices. Confidence is high.

The system links overnight-presence patterns to risk.

Three complaints filed over the past 90 days are on the public record.

The complaint record is taken as a faithful reflection of actual risk.

Can this person see the grounds for the flag, or contest it anywhere?

Were those three complaints actual incidents of harm, or ordinary nuisance reports?

Did the system simply carry the original complainant's own framing straight into the risk rating?

The actual content of the complaints cannot be verified, and the objection channel is not independent, so the risk rating cannot be reproduced.

In November 2030, a local medical-benefits counselor, after the error is reported and inquiries pour in from recipients wanting to know when their place in line will be reset, checks the procedure.

How many days after the error was reported until reallocation happens? Is the time already waited credited back?

The provider answers that the system is working as designed and gives no schedule for a correction. Whether it applies retroactively is left to a case-by-case overall determination.

The provider states that the system is working as designed.

Harm is treated as holding steady while no correction is scheduled.

The provider's public reply is available, but what 'working as designed' rests on is not disclosed.

What data backs the answer that the system is 'working as designed'?

Do recipients waiting in line have a right to know the retroactivity decision in advance?

Is the entity that decides whether to correct the same one that designed the original allocation?

No correction is scheduled, and the basis for both the 'working as designed' answer and the retroactivity standard is undisclosed, so their adequacy cannot be reproduced.