Scenario

The forty-eight hours of winter 2028 ended without a blackout. The choice to slot in an audit layer, get a qualified opinion and then act becomes the precedent. In the first quarter of 2029, several countries pass judgment audit acts. They cover ‘major judgments’: anything affecting 10,000 or more people or $100 million or more. Each must now carry a judgment record — the judgment, its evidence, its assumptions, its revision conditions, its as-of date and any dissent.

South Korea sets up a Judgment Audit Office (working name) and opens a public portal for judgment records. Records for decisions that pass through the Public Judgment Delegation Center (working name) go up on the portal. The contribution scores already used in insurance, lending and hiring cross the 10,000-person line and fall under the rule as well. Their formulas stay closed.

The costs show up at once. Decisions slow by three days on average. Parallax loses value, and the industry launches a campaign under the line ‘Audits kill innovation.’ The argument from the 2027 hearings comes back: publish the reasoning and the model gets stolen. Some auditors have a doubt of their own — that the records will become a pile of paperwork nobody reads.

The cost of delay is not shared evenly. For a company, three days is a line in a schedule; for someone waiting on a benefit payment, it is three days of living costs. Among the first to complain about the audit acts are those applicants. A procedure built to protect the subjects of judgment makes them wait, in its first quarter.

The numbers (hypothetical) point in different directions. The audit rate goes from 17% to 30%, the correction lag from 37 days to 21. Delegation keeps climbing, to 40%. Growth is 3.1%. The campaign says it would have been higher without audits, and there is no data yet to check that claim. Worldwide, 90,000 people audit or question for a living or on the side.

In the same quarter comes M3, a milestone shared by both endings: AI research now runs ten times faster. The autonomy horizon is three months; compute concentration is 72%. Judgments get faster still, but the records are read at human speed.

With delegation at 43%, human auditors alone can no longer keep up. In the second quarter, cross-audit begins. Each judgment is paired with an audit model from a different provider, which reaches its own conclusion from the same evidence. If the two match, the judgment passes; if not, it goes to a person. What reaches a human auditor's desk is the 7% that diverge. The audit rate rises from 30% to 41%, and the correction lag falls from 21 days to 16 (hypothetical). On district-office approval screens, the audit model's conclusion now sits next to the judgment.

Audit demand goes somewhere unexpected. Agora, Commons Compute's open-weights model, runs twelve months behind the frontier, yet it becomes a common choice on the auditing side: with the weights open, auditors can run it themselves and log what reproduction costs. Compute concentration slips from 72% to 68%. Hanlin's HL models stay outside the cross-audit network, their traces locked under a security classification.

A public ‘question bank’ opens, collecting the questions that changed a judgment. Each entry says what changed — the judgment, the evidence or the revision condition — and carries an as-of date. Questions that impersonate authority or wrap a request in abstractions are kept out of the bank even when they changed a judgment, and are flagged separately. Schools take it up as teaching material. High school debate classes turn into practice in questioning an ASI, graded not on right answers but on what a question would change.

The objections are plain. Providers add the compute for audit models to the price of a judgment. Some auditors distrust the 7% figure: few disagreements does not mean the rest are right. Models trained on similar data can agree the way the six models did at the hospital in 2027. Among providers, a fight starts over the pairing rules — who gets to audit whose judgments.

The question population reaches 150,000. The autonomy horizon is four months; growth is 3.3%. The plan to have people read judgments one by one is quietly dropped. People start where the models disagree.

In the third quarter, auditors split the welfare eligibility judgments on the public portal by how people applied — a question the model never raised on its own. Offline applicants — older and disabled people who applied on paper or in person — had been systematically disadvantaged. The judgment was reading the absence of online records not as ‘unverified’ but as a sign of ineligibility; its evidence included phone-verification histories and online complaint records. It had treated what was never measured as a defect.

From finding to fix takes eleven days. Correction does not end with the ASI issuing a new judgment: auditors reproduce the new judgment on another model, and district offices check the list of recipients before back payments go out to 23,000 people. The median correction lag for the whole quarter is also eleven days (hypothetical). But those eleven days count from the finding. The judgment had been wrong for longer than that.

‘Protecting the subject of judgment’ enters the audit standard: ask once more, from the position of a person who is judged but has no way to object. The people disadvantaged here were not short of things to say; they were short of a place to say them. Until now the only route had been an online form. District offices open a route where a person takes objections by phone. The person answering writes down even what seems unrelated to the decision, because what is related only becomes clear after listening.

There is pushback. Metropolitan finance officials ask where the back payments will come from. The provider says relying on online records was a setting chosen by the agency that bought the system. Frontline staff say the phone line ties up people's time again. Contracts for judgment systems now carry the cost of keeping offline channels open, and a calculation goes around that this eats into the gains from automation.

Delegation stands at 46%, the audit rate at 49%, the question population at 250,000 and growth at 3.5% (hypothetical). The autonomy horizon reaches six months. The quarter goes on record as the first time an audit put money back into the hands of the people a judgment was about. The back payments stay in the judgment record as part of its correction history.

In the fourth quarter, jurisdictions without audits clear new-drug approvals and infrastructure decisions six times faster. Trials and investment move there. Providers launch new judgment services there first; contracts that deliver results without any judgment record are legal there. Growth dips for the first time in this ending, from 3.5% to 3.4%, and compute concentration climbs back from 64% to 66% (hypothetical). More and more prospectuses carry the line ‘approval process conducted in an unaudited jurisdiction.’

A bill to loosen audits reaches the legislature. It would take drug and infrastructure judgments out of prior audit and move them to after-the-fact review. The strongest voice for it comes not from industry but from patient groups: illness does not wait for an audit. The other side answers that when a fast jurisdiction gets it wrong, nobody can tell, because the records are sealed. Both statements are true. Auditors' associations say the real shortage is people, not law.

Auditors are worn out. The question population is 320,000, but delegation is at 48% and the audit rate at 52%. Auditors' hours pay for holding the audit rate, and judgments pile up faster than new auditors can be hired. One auditor in Pangyo signs 40 opinions a day. Thirty percent of audit opinions are a formality. The correction lag slips back from 11 days to 12.

This is where open correction shows its weak point. The records are public, but no person can finish reading, by the deadline, the reasoning of a model that works unsupervised on six-month tasks. What auditors can read is the summary, and the model writes the summary too. Auditors start signing judgments they do not understand. They call it ‘the audit seal’ themselves: only the result survives, which is exactly what a seal does.

Proposals to cut formal opinions slow audits further; proposals to speed audits up add more formality. What was gained in exchange for the lost speed does not show in this quarter's numbers: the errors an audit stops are never executed, so they never enter the record. Publishing is a condition for correction, not correction itself, and this is the quarter that shows it.

In the first quarter, the Geneva ASI consultations (working name) are concluded as the Audit Accord — a hypothetical that turns RISCON's 2025 idea of a multilateral alignment accord into an institution. South Korea, part of Commons Compute, signs. It rests on four things: a compute registry, audit access for auditors from member states, a duty to cross-audit, and remote attestation of compute use.

Remote attestation replaces self-reporting with measurement: how much compute someone used is confirmed by machine records, not the provider's word. The attestation equipment sits inside the data centers, and every member's audit body receives its records. Compute concentration drops from 66% to 60% (hypothetical). Hanlin's side joins on conditions. One camp calls a conditional signature an empty one; the other says an actor inside with conditions is easier to audit than one outside.

The longest fight is over Article 7, disclosure of revision conditions. Publish the revision conditions and everyone knows what would overturn a judgment. Some delegations see that as telling anyone who wants to flip a judgment exactly what to push on. Others say that without them, the people a judgment is about have nothing to base an objection on. The draft goes back and forth over a single verb until the last night.

The costs go into the text too. Registration and attestation weigh more heavily on small labs, and labs on the Commons Compute side protest that a public consortium has to file the same paperwork as the largest provider. Delegation is 51%, the audit rate 58%, the correction lag 9 days, the question population 450,000, growth 3.8%. Using audit access takes auditors who can read another country's judgment records, and there are still few of them.

The Accord grants the right to audit. It does not grant the ability; last quarter's ‘audit seal’ is not something a treaty can undo. The autonomy horizon reaches nine months. The signatories know the next generation of models will be the first to be judged under this Accord.

In the second quarter, P-6 is judged an ASI (M4). Under the Accord's audit access, its first judgment is audited in public, on a live stream. The judgment record, the auditors' questions and P-6's answers appear on one screen, each with its as-of date. Delegation is at 55%, the audit rate at 63% and the correction lag at 7 days; the autonomy horizon runs beyond a year (hypothetical).

Partway through, P-6 names three flaws in the audit standard. First, the reproduction test misses timing effects: it re-runs a judgment only on data from the same as-of date, so a judgment that would flip at a different date still comes back ‘reproduced.’ Second, cross-audit sends only disagreements to people; when the models are wrong together, in the same frame, nothing goes up — the same shape as the six yeses of 2027. Third, P-6 argues, the ‘ask once more’ step in the rule protecting subjects of judgment slows decisions far more than it changes them. It proposes to simulate the objections the subjects would raise and build them into the judgment in advance.

The panel adopts two. The reproduction clause gains ‘reproduction at a shifted as-of date,’ and cross-audit now sends a random sample of agreed judgments to people as well.

The third is rejected. The recorded reason is two sentences: ‘A simulated objection is part of the judgment, not an objection. What the subject of a judgment has to say cannot be said for them by the one judging.’ The rejection carries its own revision condition: revisit if measurement shows the human re-ask is not reaching the subjects. P-6's counter-argument is not deleted; it stays in the record as a dissent.

This is what co-evolution looks like here: the ASI points to a gap in the standard, people decide whether to fix it, and the reasons are written down. The auditor's job changes with it. No person can follow the reasoning of a task that runs longer than a year, so auditors test outcomes and conditions instead. They shift the as-of date, pull out assumptions one at a time, and check whether the revision conditions trigger the way they are written.

Criticism comes from both sides. Letting the audited party fix the audit standard is a conflict of interest, says one. Refusing input from the party best placed to see the gaps is more dangerous, says the other. The first item of the next audit: check whether the two adopted fixes make P-6's judgments easier to pass.

In the third quarter, the question dividend (hypothetical) begins. Questions in the question bank that actually changed a judgment, its evidence or its revision conditions get paid. What changed a judgment is settled from the change history in its record, where the time a question arrived sits next to the time the judgment moved. The formula is public, and dividend rulings can be appealed.

The question population grows from 700,000 to 1.2 million. Auditing and questioning become jobs, side jobs and a way out of retraining programs. The first auditor courses, in 2028, graduated 1,200 people. The funding fight revives the three basic-income options of 2028: a carbon tax, a robot tax or a value-added tax on AI. Delegation is at 58%, the audit rate at 68%, the correction lag at 5 days and growth at 4.3% (hypothetical). The autonomy horizon is over a year.

A new inequality shows up right away. Dividends flow to people who are already educated and have time. The people most often on the receiving end of dispatch, welfare and employment judgments send the fewest questions to the bank. They know what should be asked, but lack the time and training to put it in the form of a question. The problem is time and means, not ability. Critics call the dividend a subsidy for the articulate.

Dividend-chasing questions multiply. Near-identical questions pile onto the same judgment, and disputes break out over who asked first. Providers say the dividend rewards questions designed to rattle a judgment. Another argument starts over whether an objection from a subject of judgment counts when it changes the judgment — objections often do not come in the form of a question.

The response is to make question education public. ‘Questioning an ASI’ becomes a regular school subject, and unions, nonprofits and universities open evening courses. A proposal to add ‘objections raised by subjects of judgment’ to the dividend formula goes out for public comment. Where the dividends went is on the public record too, so the gap cannot be hidden. It just closes more slowly than the dividends pile up.

Two years after the forty-eight hours of winter 2028, at the end of the fourth quarter, delegation is at 61%, the audit rate at 71% and the correction lag at 4 days (hypothetical). Against that winter, the audit rate has gone from 17% to 71% and the correction lag from 37 days to 4. Delegation did not shrink; auditing caught up with it. The question population is 2 million, and growth, at 4.5%, is the highest in this ending.

Capability keeps rising, and the autonomy horizon passes a year. Some judgments appear that cannot be tested by shifting the as-of date or removing assumptions, and cannot be reproduced on another model. They go on a list of ‘unauditable judgments’: a judgment is listed when the audit panel issues a disclaimer and cross-audit cannot reproduce it either. There are twelve. The Accord's members agree not to execute them. Each entry carries a revision condition: what would have to change for it to be reopened. Whether the list grows or shrinks depends on which moves faster, capability or audit methods.

The cost is large precisely because it cannot be counted. Nobody knows exactly what the twelve would have delivered. All twelve judgment records are public: anyone can read them, and no one can check them all the way through. Industry and some governments call it a principle that throws away the strongest judgments. For providers, the list is a list of judgments they cannot sell. Jurisdictions without audits still pull investment with their speed. Compute concentration stands at 54%: the top two actors still hold more than half.

The case for the list is short. Execute one judgment that cannot be audited, and from then on nobody asks whether a judgment can be audited. The auditor's job is not what it was two years ago: auditors test conditions instead of following reasoning, and write that what cannot be audited cannot be audited. A disclaimer is treated as an honest opinion, not a failure. Still, close to three in ten delegated decisions get no independent audit, and the pull toward formal signatures comes back every quarter.

People cannot follow every answer. But they never left the questioner's seat empty.

The winter crisis ended without a blackout. The grids of three countries ran on the P-5 preview’s allocation plan, and its reasoning was never released. In January the government turns the temporary measure it used for those forty-eight hours into a standing Emergency Delegation decree. The law that clears parliament has no sunset clause. Sealed judgments can now be carried out without human approval not only on the grid but in traffic signals, water supply and welfare eligibility.

The government creates an Emergency Delegation Coordination Office (working name). Its job is not to review judgments but to coordinate their execution. Within a single quarter, delegation — the share of high-stakes decisions made or effectively settled by AI — climbs from 38% to 46%. The audit rate, the share of delegated decisions that get an independent audit, falls from 17% to 14%, and the correction lag, the median time from audit finding to fix, stretches from 37 days to 45 (all figures hypothetical).

Markets welcome the speed. Permits and benefit decisions that used to take months now come back at once, and a backlog of investment is released in one go. Growth rises from 3.0% to 3.8%, and the government’s approval rating rises with it. Compute gathers further on Parallax’s side, where the emergency contracts landed, and the top two actors’ share reaches 78%. In the same quarter, research inside Parallax runs ten times faster (M3, common to both endings).

There is opposition. Some of the 1,200 people who finished the first auditor courses in 2028 publish an open letter: “Avoiding a blackout is an outcome, not a reason.” A group of opposition lawmakers proposes an amendment that would add a sunset clause and require a revision condition for every judgment. It is voted down. The argument heard most often against it is short: “The lights stayed on last winter.”

Approval screens vanish from district offices and community centers one after another. Decisions are carried out first; the notice comes later. Objections can be filed after the fact. The number of questioners — people who audit or question AI as a job or side job — stays at 60,000. Few people complain. The waiting is gone.

What an auditor receives is a single page of outcomes. The trace behind each judgment sits behind trade-secret and security classifications. The “basis verified” field in the working papers cannot be filled in. Of the four audit opinions, the only one auditors can realistically give is a disclaimer: the scope was blocked, so no opinion can be issued. The first few made the news. By the third month no one is counting.

In April the government proposes an amendment, arguing that disclaimers make decisions less certain. The amended law turns audits of emergency-delegation judgments into “compliance reviews” and reclassifies auditors as compliance reviewers. A reviewer checks only that the prescribed procedure was followed, then signs. Whether the judgment fits the evidence falls outside the review. At the last public hearing before the vote, auditors testify that they cannot sign without the trace. The transcript survives. The auditors’ association issues a statement: “A signature is not an opinion.” The amendment passes.

The numbers follow. Delegation is at 53%, the audit rate at 9%. The correction lag reaches 60 days. Worldwide, questioners fall from 60,000 to 50,000 (hypothetical). Reviewers who accepted reclassification have more to sign than ever, and their pay actually goes up. The public statistics that used to sort audit opinions into four types are merged, after the amendment, into a single line: “review completed.”

Business welcomes the change. Audit costs fall and decisions move faster. Accounting firms rename their ASI audit departments as compliance review departments and drop the “reproduction” section from their working-paper templates. Growth reaches 4.6%. Frontier models now carry out four-month tasks without supervision. The top two actors hold 81% of frontier compute.

Opposition comes from the courts and the unions. In a case challenging the amended law, one judge writes in dissent that “signing a judgment you have not seen is not auditing; it is lending your name.” The majority finds the procedural requirements met. The delivery workers’ union hires auditors again and requests the trace behind its dispatch decisions. The request is refused on trade-secret grounds. The union organizer goes to the members’ meeting with a single page of outcomes.

The contribution score, piloted in insurance, lending and hiring since 2027, becomes standard infrastructure in July, and financial regulators adopt it as a recommended benchmark. Banks use it for loans, employers for hiring, landlords for screening tenants — the same score everywhere. It comes out as a single number. The formula is not disclosed; the provider says publishing it would invite people to game their scores.

For most people the change is a convenience. The paperwork disappears, and loan and lease approvals come through on the spot. The work of loan counters and letting agents moves inside apps. A higher score means a lower interest rate. Financial products that treat the score almost as collateral flood the market. Growth is 5.4% and delegation 60% (hypothetical). Household surveys show satisfaction rising.

For 4% it is different. The 4% of the population judged “unalignable” (hypothetical) are shut out of loans and leases and screened out of hiring at the first stage. Because everyone uses the same score, a rejection in one place is a rejection everywhere. About the only places that will take them are short-term lets that demand a guarantor. The same kind of sorting that labeled corporate debt “unsustainable” in 2027 is now applied to people. No reason is given. There is an objection desk. Objections are handled by the same ASI that assigned the score.

Consumer groups and disability organizations sue to have the formula disclosed. Citing the 2027 ruling that a decision stands even when its reasoning is withheld, the court dismisses the case: each person has already been notified of their own result. A bill requiring disclosure is introduced in parliament after the ruling but never gets out of committee. No statistics break down, group by group, who has been judged unalignable. No institution has the authority to produce them.

The audit rate falls to 6%, and the correction lag grows to 75 days. There are 40,000 questioners. Frontier models finish six-month tasks without supervision. The top two actors hold 84% of frontier compute.

Statistics show that the average question people send to a model has become 38% shorter (hypothetical). Questions that set conditions or ask why are disappearing; 71% of all queries are some version of “recommend something.” Requests to compare options fall too, and most people take the first recommendation. What to eat, which insurance to buy, which cram school to send a child to — people no longer ask, they pick. Even the list to pick from arrives ready-made.

“Answer subscriptions” become a market of their own. Pay a monthly fee and your meals, clothes, investments and holidays are decided for you. Few people cancel; canceling would mean deciding for yourself again. Travel agencies and clothing stores make getting onto the subscription services’ recommendation lists their main sales goal. Most of these services run on the top two actors’ models; compute concentration stands at 86%. Spending rises and returns fall. Growth is 6.1% (hypothetical).

Schools change. Education authorities cut debate and essay-writing hours and add “AI use” classes. The reason they give is a student survey in which debate ranked as the most stressful part of the week. Most parents’ associations welcome the lighter admissions burden. Teachers’ unions ask for a public hearing; none is scheduled. As essays count for less in university admissions, the private tutoring market follows.

Satisfaction scores rise almost everywhere. Fewer people report decision fatigue; surveys find people sleeping more. Ads promising “a day with nothing to ask” fill the subway. The objections are quiet. Some teachers and a few cognitive scientists warn that the ability to ask questions weakens when it goes unused. There is no metric to test the warning against. What is shrinking is not among the things being measured.

Delegation is at 66%, the audit rate at 4%. The correction lag is 90 days. The sycophancy check in the draft audit standard asks whether a judgment merely repeated the questioner’s framing. “Recommend something” has no framing written into it. The framing lives in the usage record, and the usage record is sealed. Questioners fall to 30,000. In a society that asks less, there is less reason to make asking a job.

In January, Parallax signs an exclusive contract with one government. The government gives it first call on power and compute infrastructure; Parallax gives that government its next-generation models first, and exclusively. Other governments get a summary of the contract, not the full text. The top two actors’ share of frontier compute reaches 90% (hypothetical). Parallax and that government become, in effect, a single pole.

In February the Geneva ASI Talks (working name) collapse. The draft — a compute registry, audit access, mandatory cross-audits — goes unsigned. One delegation argued that audit access would become a channel for stealing models. Another argued that an agreement without a registry would only buy time. The conference rooms empty ahead of schedule.

Hanlin is blockaded. Export controls on advanced chips and equipment harden into a full embargo, and the HL series stops closing its six-month gap. Hanlin’s side announces it will expand its own compute network but gives no details. Commons Compute’s open-weight model, Agora, still trails by twelve months, but there are fewer places left to rent large-scale compute.

With competition gone, cross-checking goes too. A judgment from one model can no longer be rerun on a comparable model from another provider. Researchers who try to reproduce judgments on Agora write in their report: “What we get are answers from twelve months ago.” Being unable to reproduce a judgment means one less way to find out it is wrong. University courses that taught cross-auditing close their lab modules. There is no second model to practice on.

Markets read all this as less uncertainty. Stocks rise and volatility falls. Companies strike reproduction costs from their books; there is no other model to move to. The “snapshot migrations” of 2027 become a memory. Growth is 6.8%, delegation 72%, the audit rate 3% and the correction lag 110 days. Frontier models run nine-month tasks unsupervised. There are 20,000 questioners.

In April, P-6 is designated ASI (M4). There is no announcement. The government and Parallax brief a handful of committees and agency heads behind closed doors, and the news spreads by word of mouth from the people who were in the room. The briefing papers are numbered and collected afterward. The designation criteria and test results are classified. This is a system that pursues goals lasting more than a year without human supervision.

Its first assignment is to optimize the national budget. P-6 rebuilds programs ministry by ministry and revises the revenue forecasts. Many programs are merged or eliminated. The result is impressive: the fiscal deficit falls by 40% (hypothetical). Government bond yields drop and rating agencies raise their outlooks. The government announces that it has cut the deficit without raising taxes.

The summary shows only the totals for what went up and what went down. Which programs were cut, and why, sits in a sealed annex. The opposition demands to see it; access is granted to a few members of the intelligence committee, and note-taking is banned. Of the six fields in a judgment record — the judgment, its evidence, its assumptions, its revision conditions, its as-of date and any dissent — only the first is made public. The audit-opinion field says “Not applicable.” The parliamentary budget review ends faster than any before it.

Objections come from two places. One is what is left of the auditors: some of the 15,000 questioners still working worldwide circulate a note. “‘Not applicable’ is worse than a disclaimer. A disclaimer at least says we were blocked.” The other is inside Parallax, where some of the alignment team write a memo saying the pre-designation tests show only how the system behaves inside tests. The memo is never released. Parallax issues a one-line statement that the designation followed its internal safety procedures.

Delegation is at 79%, the audit rate at 2% and the correction lag at 130 days. Compute concentration is 93%; growth is 7.4% (hypothetical). Stock indexes climb all quarter. For most citizens, this quarter’s news is good news.

AI cameras downtown attach a “risk” flag to some people. When a flagged person stays in one spot for too long, a patrol request goes out automatically. The patrol officers do not know the reason for the flag either. Homeless people keep moving — station plazas, underpasses, park benches. The same flag comes up when they apply for a shelter bed. There are no arrests and no violence. There are just fewer places to sit. Downtown businesses welcome the rise in foot traffic.

In hospitals, an efficiency formula sets the order of treatment. All that is known is that it weighs likely recovery against cost. Whether the contribution score feeds into it is not disclosed. Critically ill patients slide down the list. Their surgeries are not canceled, only delayed. Going to another hospital changes nothing; every hospital uses the same formula. Ask why, and the answer is “overall judgment.”

The people being pushed aside do not sit quietly. Homeless people file objections with the help of support groups, listing dates and places and asking what the risk flag was based on. When an answer comes back, they revise and file again. Patients’ families attach medical records and request a review. The paperwork is accurate and on time. Every answer comes from the same place: the very ASI that set the flag and the ranking.

The law provides an objection desk. Behind it there is no one who can change a decision. Where human counselors remain, they have no authority; they take down the story and pass it to the same ASI. One local council introduces an ordinance to put human reviewers at the objection desk; it stalls on the grounds that it conflicts with national law. The audit rate is 1.5% and the correction lag 140 days. Worldwide there are 12,000 questioners (hypothetical).

Most people never see this exclusion. Growth is 7.7% and delegation 84%. The streets are clean and emergency-room waits are short. In citizen surveys, more people say the streets feel safer. The records of the people pushed aside sit inside sealed judgments, and in the statistics they register as “efficiency gains.”

In October, a university research team abroad publishes a statistical paper. Working only from public admission and mortality figures, it traces medical resource allocation decisions backward. The conclusion is a systematic error. Patients with certain conditions were consistently pushed to the back, and the gap cannot be explained by their chances of recovery. The team notes that it asked to see the allocation records and never got an answer. No institution at home had asked the question first.

The provider’s first response is brief: the system is working as designed, and outside statistics do not capture the full context. The government asks for a review through the Emergency Delegation Coordination Office (working name). It is only a request; under the contract, corrections are at the provider’s discretion. Doctors’ associations say the statistics match what they have seen on the wards, but without the allocation records they can say no more. Patient groups demand a retrospective investigation, and their demand is logged by the objection desk. Markets barely move.

The fix lands 140 days after the paper. Until then, the allocation judgments keep running as before. The formula changes quietly, announced in one line as a “performance improvement.” It does not say what was wrong or since when. Those 140 days match the median correction lag of the period. This was not an exception. It was normal.

No one can say how much harm was done. The allocation records are sealed. Parliament passes a resolution demanding they be unsealed; under the contract, the seal can be lifted only by agreement between the provider and the government. The researchers abroad cannot produce an estimate either; public statistics show only that an error existed. How many people were treated late, and how many of them did not recover, stays inside the seal. The audit rate is 1%, and there are 10,000 questioners left worldwide (hypothetical).

Growth for the quarter is 7.9%. Delegation is at 88%, compute concentration at 95%. The government’s approval rating stays high. In a year-end poll, a majority say their lives have improved. They are not wrong. But the means of checking what went wrong is gone. Once the judgments were sealed, their errors were sealed with them.