Chapter 2. Appendix: The records read

A government industry-ministry official reviews compute export controls.

Should equipment above a certain compute grade be restricted from export this quarter to slow Hanlin's HL-3 catch-up?

The model recommends restriction with medium confidence, citing a correlation between the last two quarters' export volumes and the HL-series catch-up speed.

The model reads export volume and catch-up speed as moving together.

Export-volume statistics from the last two quarters.

Other channels for equipment to reach Hanlin are left out, as if there were none.

Can two quarters of data support a correlation claim at all?

Does a domestic equipment maker that benefits from this restriction also supply data to this model?

How many months old is the catch-up-speed estimate this is based on?

A correlation is treated as causation, and whether the party that benefits from the recommendation also supplies the underlying data is unconfirmed — the basis does not support the judgment.

An asset manager's risk team checks the herding risk from multiple firms using the same re-rating model.

The top two re-rating models hold 62% (fictional) market share. What loss follows if both flag the same bond 'sell' at once?

The model finds 'herding risk is limited,' with high confidence, citing an 8% disagreement rate between the two models' past judgments.

8% disagreement rate between the two models over the past six months.

The model concludes that low disagreement means low herding.

That the two models were trained independently of each other is accepted without checking.

Is the 8% figure from a period like this market-wide stress event, or from calmer times?

Was 'no herding risk' the conclusion the risk team wanted to hear?

Does the 8% disagreement rate reproduce across other bond groups?

A low disagreement rate means the two models tend to judge alike, so it may itself be evidence of herding. Citing it as evidence against herding reverses the logic, so the basis does not support the judgment.

Ahead of winter, a duty officer in a power-trading control room checks which vendors supply the demand-forecast models.

Demand forecasting relies 92% (fictional) on the top two vendors' models. Can we run the winter this way?

The model finds 'conditionally workable,' with medium confidence, citing that both vendors' forecast errors stayed within 3% of each other last winter.

Forecast-error records from last winter.

The model concludes that similar errors mean the two models cross-check each other.

The two vendors' forecasting methods are regarded as genuinely independent.

Did the 3% error gap hold during last winter's coldest stretch too?

Has the two models' methodological independence actually been verified?

Do households that would suffer an outage from a bad forecast have any voice in this judgment?

The error record checks out, but methodological independence is unverified. Similar errors could mean the models check each other, or that they fail together the same way, which is an important limitation.

The lead of a three-person startup in Pangyo is reviewing a one-year contract that would put the entire service on a single P-4 API plan.

If we commit to this API plan for a year, how long and how costly would it be to switch providers if Parallax changes its pricing or policies?

A contract-recommendation tool advises taking the one-year plan, saying the 34% commitment discount outweighs the switching-cost risk, with confidence rated "medium."

Peer companies in the same industry took an average of 3.2 days to switch last quarter (based on the "snapshot migration" cases).

The calculation treats this startup's codebase as similar to that industry average.

The calculation that the 34% discount outweighs the switching cost.

How many companies make up the "3.2-day industry average" sample, and are they comparable in code scale to this startup?

Is this contract-recommendation tool provided by Parallax, or by a third party?

If Hanlin's or Commons Compute's pricing becomes cheaper during the commitment period, does this recommendation get recalculated?

The sample is only an industry average, not measured on this startup's own codebase, so applying the 34% figure outright rests on too narrow a basis.

At a labor-management council debating the introduction of board-level worker representation at a large company, the question arises that the judgment system used to determine layoff targets relies entirely on a single provider.

The layoff-target calculation logic runs entirely on one Parallax-based system — should a system for cross-checking it against another provider's model be built now?

Management judges that cross-verification is not worth the cost, and decides to keep the single-provider system, with confidence rated "medium."

The single-provider system's error rate is regarded as already low enough, with no separate verification done.

No formal appeal has overturned a calculation in the past two years — though the formal appeal procedure itself was only recently created.

Management says that, since no appeal has overturned a result, cross-verification is unnecessary.

How recently was the procedure for a layoff target to check the basis and file an appeal created, and how many people have actually used it?

Is the vendor that built the calculation logic the same one reporting its performance back to the company?

Does feeding the same personnel data into a different provider's model produce the same list of targets?

The absence of appeal cases may simply reflect that the procedure is new and unused, so it cannot support the conclusion that the error rate is low.

On the first day of the power crisis, a bond trader watches the national utility's bonds plunge and asks an advisory model which of the two choices would do less damage to the bond market.

Using the same model, what is the estimated impact on bond spreads of the execution delay under choice I (partial audit, then execute) versus choice II (execute immediately)?

The advisory model estimates that "choice II, by removing the delay, cuts the spread shock to about half of choice I's," with confidence rated "medium," based on a post-hoc analysis of three past similar crises.

A correlation between delay length and spread width found across three past similar crises.

The advisory model estimates that this crisis will follow the same correlation as those three.

This crisis's "delay" — inserting an audit layer — is treated as the same kind of delay as in those three past cases, which arose from different causes.

What kind of delay did those three past crises actually involve, and is it the same kind as an audit-layer delay here?

With a sample of only three, can a specific multiplier like "cuts it in half" really be stated?

Could the trader be receiving this result after already having sold the bond in question?

The three past cases involved a different kind of delay, yet the same correlation is applied to this crisis unchanged, leaving an unverified leap between the evidence and the conclusion.

Late March, an audit architect designing access rights for a national research-compute allocation reviews a judgment covering three public research consortia.

Can this audit system actually access the full evidence behind the judgment that assigned 78% of compute to two of the three consortia?

The model proposes allocating 78% of the compute to two of the three consortia. No confidence figure is logged, and part of the underlying proposals is sealed as trade secret. It sets a condition: recompute the allocation if the seal is lifted.

The model projects future output from past output.

the publishable portion of submitted research performance metrics

The starting point is that concentrating compute speeds up research.

Did the sealed portion of the proposals actually affect the allocation ratio?

Did people affiliated with the two larger consortia take part in designing this audit access?

Does the design include a channel for the excluded third consortium to object?

Part of the core evidence is sealed as trade secret, so no opinion can be issued under the access currently designed.

In August, promotional material for a local festival keeps getting throttled by a content-filtering judgment shared by two major platforms.

Is the 'needs caution' label on this promotional image really coming from two independent platforms, or from one shared filtering model?

The model classifies the promotional image as 'needs caution'. Confidence is 61% (synthetic), based on an image-classification score and reported-complaint statistics. It sets a condition: if both platforms' calls trace to the same model, reclassify them as one judgment, not two independent ones.

The model holds that the image-classification score picks out provocative content.

complaint-history statistics that matched across both platforms

The two platforms' classifications are accepted as independent of each other.

Do the two platforms actually run the same model from the same compute provider?

Is the same classification standard applied consistently to other festival photos?

Does the event organizer have a separate way to object on each platform?

The two platforms' calls looked like independent cross-verification but traced to the same provider's same model, masking a concentration risk.

In December, after word of the I4-01 signing incident spreads, an audit architect proposes a daily caseload cap per auditor.

What cap would actually bring down the 30% rate of perfunctory opinions caused by the 40-cases-a-day habit?

The model recommends capping each auditor at 12 cases per day. Confidence is 58% (synthetic), based on a two-quarter correlation between caseload and the rate of perfunctory opinions. It sets a condition: readjust the cap if the perfunctory-opinion rate does not fall after it takes effect.

The model reads caseload and the rate of perfunctory opinions as moving together.

two quarters of caseload and opinion-type records

The cap is counted on not to create a backlog problem of its own.

Does the 12-case cap reproduce on data from a different period?

Does the backlog created by the cap end up making the people affected by these judgments wait longer?

Does anyone reviewing this proposal belong to an institution that benefits from the auditor shortage?

The correlation reproduces, but the cap's effect on the backlog and on how long affected people wait has not yet been measured.

A shared compute center near Zurich, the first week of the compute registry's pilot run. A Commons Compute researcher checks the audit-hours allocation table.

If registry audit hours are allocated in proportion to compute share, does our side — with a low share but Agora's wide deployment footprint — get audited enough? Is there a quarterly floor?

Under the draft, Commons Compute would receive only about 9% of total audit hours — likely underweighting the risk posed by a widely deployed open-weight model like Agora. Recommend a 20% floor on audit hours regardless of compute share. Confidence: medium.

The audit-hours allocation formula in the Accord draft's annex.

The existing design is built on audit-failure probability scaling with compute share rather than deployment count.

The model holds that share-proportional allocation underweights the risk of widely deployed open-weight models.

The Accord text has no metric like 'deployment count' to anchor a 20% floor to — what is this recommendation's revision condition?

This only checks the annex formula — was it cross-checked against last quarter's actually executed audit hours?

Commons Compute, which requested this judgment, is also the party disadvantaged by the current formula — is that conflict of interest disclosed in the basis?

The formula text is clearly confirmed, but without execution data as evidence, the risk estimate remains at the level of a claim.

Near Zurich, a Commons Compute researcher reviews the question bank's operations, analyzing which service paths this quarter's registered questions came through.

What share of judgment-changing questions registered in the question bank came through Parallax-affiliated paths? If Agora's share is too low, doesn't the dividend end up concentrated among users of one service?

This quarter, 81% of registered questions came through Parallax-affiliated paths, 14% through Hanlin-affiliated paths, and 5% through Agora-affiliated paths (synthetic figures). This roughly tracks each bloc's user-base share, but whether it reflects a real difference in question quality or simply access and usage volume cannot be determined. Confidence is high on the raw proportions, low on the causal interpretation.

This quarter's tally of registered questions by submission path.

Submission-path shares are taken to track user-base shares (not yet directly cross-checked against user-count data).

The ASI reads this concentration as reflecting access differences more than differences in question quality.

Was this actually cross-checked against per-path user-count data, or does it stop at an estimate?

Was a reproduction test run — submitting the same question content through all three paths at once to see if adoption rates match?

If the Parallax-affiliated path has a feature that polishes a user's question wording before submission, doesn't that favor phrasing over substance?

There is not yet any basis to distinguish among the possible causes of the concentration — user-count difference, quality difference, or path-level favoritism — so the audit's scope to render a fairness opinion is blocked.

In February 2029, after the emergency delegation decree became permanent, a grid-dispatch officer checks next Tuesday morning's peak-hour allocation plan with the system.

Is the plan to cut supply to Region A by 15% during Tuesday's 8-10 a.m. peak correct, and does Agora reproduce the same result under the same conditions?

Weighing demand forecasts and reserve margins together, a 15% cut to Region A is optimal, with low blackout risk.

Part of the raw regional demand-forecast data is public.

The equipment-availability data used to compute supply capacity is undisclosed.

The model's own estimate is that supply capacity will run at about 92% of forecast.

The model works from a reserve margin that holds at its usual 8% or above.

What equipment-availability data backs the 92% supply-capacity estimate, and is that data public too?

Does feeding the same demand data into Agora produce the same 15% cut, and if not, where do the results diverge?

What date does the demand forecast behind this plan rely on, and does it include the latest weather-forecast update?

The demand data is public and partly reproducible, but the data behind the supply-capacity estimate is undisclosed, so reproducing the size of the cut itself remains limited.

In September 2029, an audit architect notices that three rental platforms in one region all use the same contribution score as a screening criterion and requests a cross-check.

Please confirm whether all three rental platforms in this region reference the same contribution-score result, or whether any of them use a different formula.

All three platforms reference the same score provider, and the results are consistent. Consistency here means the score can be trusted.

The public terms of all three platforms state they use the same provider.

The model holds up the consistency of the results as evidence of reliability.

All three platforms are treated as querying the score at the same point in time.

With no independent second scoring body to compare against, can consistency alone confirm accuracy?

Are the three platforms tied to the score provider by contract or ownership?

Can a renter contest the fact that the same score follows them even if they switch platforms?

The three platforms' results do match each other, but since there is only one provider to compare, independent verification of accuracy itself is not achieved.

In February 2030, a duty officer in a regional power-trading control room reviews the priority order for reserve power allocation. It is the first review since the exclusive contract with the sole remaining provider was signed after the winter crisis.

Why does this district get reserve power later than others. Would a different model produce the same order?

The system judges that a shutdown at the neighboring industrial cluster carries a higher risk of cascading outages, and ranks the cluster ahead of the residential zone. Confidence is marked high, and the detailed grounds are closed off as an overall determination.

The system estimates that a shutdown at the neighboring cluster carries a higher cascading-outage risk.

Outage logs from the previous two quarters are public.

Another provider's reserve-margin data staying undisclosed is taken for granted.

Would another model, from before the blockade, have produced the same order from the same inputs?

Whose reserve-margin data, and how much of it, actually went into the calculation?

Is the prioritized industrial cluster also a major customer of the reserve-power provider?

Access to the underlying reserve-margin data is blocked, so the allocation order cannot be reproduced.

In December 2030, a power-systems engineer based abroad, analyzing public grid-frequency stability statistics, flags that one jurisdiction's reserve-margin figures have looked unusually optimistic ever since its calculation method switched to a single-provider system.

Is there data that compares before and after the method change on the same basis?

The relevant authority continues reporting that the reserve margin meets standard. The fact that the method changed, and its coefficients, are withheld as a trade secret.

The determination that the standard is met.

Public grid-frequency statistics show a partial anomaly.

The change in calculation method is treated as having had no effect on the result.

When did the method change, and can before-and-after be compared?

Does recalculating recent data with the old method give the same conclusion?

Is there underlying reserve-margin data published beyond the frequency statistics?

The public frequency statistics partly confirm an anomaly, but with the calculation coefficients undisclosed, the overall conclusion cannot be reproduced.