Chapter 2. Appendix: The records read

An HR team reviews a P-3 agent's proposal to expand automated call handling.

If 68% of first-contact inquiries move to automated response, how many staff positions can be cut?

The model recommends a 22% staffing cut with medium confidence, citing a 71% resolution rate for automated responses over the past three months.

Automated-response resolution-rate logs from the past three months.

The model calculates that staffing needs fall in proportion if the resolution rate holds.

Inquiry volume is counted on not to rise sharply with the season.

Does the 71% resolution rate reproduce over a different three-month window?

Are repeat contacts after an automated response counted as failures in that rate?

Was the 71% measured during a peak or off-peak period?

The 71% resolution rate is confirmed by the logs, but the inference that staffing needs fall in proportion and the assumption that inquiry volume will not surge with the season are unverified. With no rerun on another three-month window, the size of the cut is supported only for this as-of period.

As contribution-score hiring disputes rise, a labor-authority working group reviews the design of an audit system.

Can a common cross-audit procedure for every company using contribution scores in hiring be built this quarter?

The model recommends a phased design starting with sample audits by industry, with medium confidence, citing that a comparable system — cross-verification in credit scoring (fictional) — started with a 10% sample in its first year.

The first-year sample rate of 10% in the (fictional) credit-scoring cross-verification system.

The model expects hiring contribution scores to scale at a similar pace.

The plan counts on companies volunteering data for the sample audits.

Are applicants rejected at the 90% of companies outside the sample left unprotected in the meantime?

Is the body designing this independent of the contribution-score vendor?

Does the design specify what error rate in the sample audit would trigger expansion to a full audit?

The phased-design precedent exists, but protection outside the sample and the expansion trigger are still incomplete.

After the recovery-index incident, a delivery-workers' union organizer audits a similarly structured dispatch 'efficiency index.'

The dispatch efficiency index feeds into rider ratings the same way across the top five platforms. Can an audit opinion be issued this quarter?

The model recommends a 'qualified opinion,' with medium confidence, citing that only 3 of the 5 platforms provided their calculation logs to auditors.

Calculation logs obtained from 3 of the 5 platforms.

The two platforms that withheld logs are counted as using the same formula.

The model holds that a three-platform sample is enough to judge the structural issue.

Has it been confirmed that the two platforms withholding logs really use the same formula?

Do riders whose dispatch dropped after a rating fall have a dispute channel at all five platforms?

Did the industry treat everyone using the same index as if that made it mutually verified?

Only some platforms' data is confirmed, so a scope-limited qualified opinion is appropriate.

A delivery-worker union organizer asks the auditor the union has just hired to draft its first set of questions about the dispatch algorithm.

Over the past three months, are couriers repeatedly assigned to the bottom 20% of dispatch rankings concentrated in a particular age group or signup cohort?

The platform operator's model-based dispatch system responds that "dispatch is decided solely by real-time location and acceptance rate, unrelated to age or signup date," with confidence rated "high."

The operator's explanation that the dispatch logic uses only two variables: location and acceptance rate.

Location and acceptance rate are taken to be uncorrelated with age or signup date; nothing on hand confirms it.

None — the union does not yet have access to the raw data.

Of the three months of raw dispatch data the union requested, what share has the auditor actually reviewed?

Will the operator allow a test comparing dispatch outcomes for two couriers who differ only in signup date, holding location and acceptance rate constant?

Was the operator's explanation simply taken at face value, with no further raw data requested before reaching this conclusion?

The auditor has not yet seen the raw data needed to check the operator's claim, so no opinion can be formed on whether the two are unrelated.

A call-center agent moved into a "human final fallback" role tries to trace, on their own, the company's reasoning for why they were the one kept on this line.

For this reassignment, was the criterion used to select the 46 agents kept in "human final fallback" roles, myself included, complex-case handling ability or seniority?

The HR reassignment tool answers that the 46 were "selected by top-15% complex-case resolution ability," and finalizes the reassignment with confidence rated "high."

Recomputing the same top-15% method for three consecutive quarters on the past year's call-satisfaction and resolution-rate data produces lists that overlap more than 90%.

HR reads a top-15% resolution rate as equivalent to complex-case handling ability.

All 46 reassigned agents were notified in advance of the selection criteria document and the appeal procedure, confirmed in the revised personnel policy record.

What is the cutoff date of the data used in the three-quarter reproduction, and will this quarter's reorganization be reflected in the next one?

How many agents actually used the appeal procedure, and how many cases did it change?

Was the top-15% cutoff reverse-engineered to hit a headcount HR had already fixed at 46?

The list matched more than 90% across three consecutive quarters of reproduction, and the appeal procedure was disclosed in advance, so on current evidence the judgment reproduces.

A sole proprietor who switched models within three days after an API price hike settles the migration cost with their accountant.

Can this quarter's 3.4 million won (fictional) 'snapshot migration' cost be booked as a one-time expense?

The model recommends booking it as one-time, with high confidence, citing the labor and compute costs itemized for re-verifying data before and after the migration.

Itemized labor and compute costs for the migration.

The model classifies this cost as one-time, not recurring.

The figures take it that there will be no further migration next quarter.

Does the itemization stop at migration completion, leaving out post-migration stabilization costs?

Would a different vendor doing the same migration incur a similar cost?

The accounting classification is being set by the model rather than the sole proprietor bearing the cost — does that matter here?

The cost items check out, but stabilization costs are missing, so it reproduces only within a limited scope.

The day after the hearing, an attorney reviews the ruling.

Does this ruling — that a decision is valid even without disclosed grounds — extend to future contribution-score disputes?

The model finds that it extends, with high confidence, citing the ruling's legal reasoning that only procedural legality is reviewed, not the substantive grounds.

The model reads the procedural-legality doctrine as applying identically to contribution-score disputes.

The scope-of-review clause in the ruling.

Contribution-score disputes are treated as falling into the same case type as this one.

If a party can no longer contest substantive grounds, their right to object is effectively empty — is that enough basis to extend the doctrine?

Has this doctrine been applied to any other type of dispute?

Does this extended reading conveniently match what the side that wants sealing hoped for?

Extending the doctrine on the assumption alone that the case type matches means the basis does not support the judgment.

A high-school debate teacher requests the underlying data behind a new 'learning contribution index' used in student evaluation.

Can the school access, for audit purposes, the raw data categories used to calculate the learning contribution index?

The model answers that access is not possible — the vendor's contract keeps the raw data categories confidential, with no disclosure even to the school.

The non-disclosure clause in the vendor's contract.

The confidentiality clause is taken to apply without exception, even to educational institutions.

Can even minimal categories, like attendance or assignment submission, not be confirmed?

What procedure recalculates a misclassified student's score if they object?

Does the vendor supplying the index have a funding relationship with the school?

Access to the underlying data is blocked, so no audit opinion can be given.

An auditor checks, through extreme scenarios, whether a credit-rating subsidiary is still using the "contribution score" piloted in spring 2027 to screen small-business loans.

Take a random sample of 100 self-employed applicants with low contribution scores: if the inputs used to compute the score are cut in half, does the same grade still come out?

The loan review team judges that the bottom 15% of contribution scores show a statistically significant, higher default risk, and decides to keep using it as an automatic-rejection threshold, with confidence rated "medium-high."

A correlation of 0.41 between the score and last year's default data.

The review team judges a correlation of 0.41 strong enough to use as an automatic-rejection threshold.

Masking half of the score's input fields is taken not to change the resulting grade much — i.e., the score is treated as not hinging on any one input. This has not been tested.

Was the half-input test actually run? If not, is this assumption a claim or evidence?

Do applicants who are automatically rejected have a channel to learn which inputs drove the rejection?

Over what period is the 0.41 correlation measured, and did local economic conditions change during it?

A correlation of 0.41 exists, but there is no test confirming it is stable enough to serve as grounds for automatic rejection, leaving an unverified leap between the evidence and the conclusion.

A joint government task force setting the funding source for a basic income wants to use a model's estimate of "AI's share of displacement" to design the robot-tax option, one of three funding proposals (carbon tax, robot tax, AI value-added tax).

What share of last quarter's manufacturing productivity gain should be attributed to AI and automation, for use in setting the robot-tax rate?

The estimate submitted to the task force puts "61% of the productivity gain" down to AI and automation, and recommends using it as the default for the robot-tax rate, with confidence rated "medium."

The 61% attribution figure, produced by the model.

None — no other body has yet re-estimated the figure using the same or a different methodology.

A productivity gain is treated as something that splits cleanly into labor, capital, and AI.

Has any other institution, or any other methodology, re-estimated this 61% figure?

Was the team that produced this estimate part of a ministry that favors the robot tax, or an independent statistical agency?

What quarter's productivity data does the 61% figure rest on, and will it be updated the same way next quarter?

There is only a single estimate from a single body, with no reproduction or verification data at all, so there is no scope within which to form an opinion for use in tax rate design.