Scenario, not forecast. Every number is an authored assumption; every lab, model and agency named here is fictional. As of 25 September 2026.
Eight axes, each running from Open Correction (−2) to Sealed Judgment (+2). Follow where each ending moves them, quarter by quarter.
The side where people keep the place to ask. Questions stay with people; judgment records — basis, assumptions, revision conditions, as-of date and dissent — are public; audit findings are applied within a set deadline, and the people being judged can appeal.
The trunk starts near 0 and leans slowly toward +0.8. At 2028 Q4 it forks into the two endings.
The side where judgments are sealed. The basis and trace of a judgment are closed on grounds of trade secrecy or security, and only the result remains. Even the questions are delegated; correction is at the provider’s discretion, and appeals are handled by the same ASI.
The same technology, two worlds. In both endings the ASI is the same system; what differs is whether people kept the place to ask.
Filled marks stand where the chosen quarter leaves each axis. Outlined marks are the end of the trunk (T9 · 2028 Q4) and the ends of the two endings (I8 · II8 · 2030 Q4); the faint lines run from the fork to each end.
Do people still ask the questions that decide what gets judged?
2030 Q4I −1.7II +1.8
People set the public agenda, teaching people to ask becomes a public service, and questions that change judgments are paid a dividend.
Questions that changed judgments collect in a public question bank, and schools add practice in asking the ASI.
People choose the tasks and questions and hand them to agents, but rarely ask again about the results that come back.
The ASI proposes agenda items and people pick from the list, while the questions people write grow shorter on average.
The ASI sets the agenda, people receive notice of the results, and asking becomes a hobby.
People set the agenda, and the ASI answers their questions with judgments. Questions that changed a judgment collect in a public question bank, and high-school debate classes become practice in asking the ASI. A question that changes a judgment earns a question dividend.
The ASI sets the agenda, and people receive answers. The questions people send to models get shorter, and ‘just recommend something’ becomes the norm. Schools cut back on debate and essay writing, and households subscribe to answers.
Agendas start from people's needs, and questions that changed judgments build up as a public asset. Asking and auditing become paid work.
A new inequality opens between people who ask well and people who do not. Teaching people to ask costs public money, and every question people ask slows a decision.
Answers arrive before anyone has to work out what to ask. Decision fatigue drops and satisfaction scores rise.
The ability to ask withers, and debate and essay teaching shrink. What goes unasked shows up in no metric, so the loss stays out of sight.
Is a judgment published along with its evidence, assumptions, revision conditions, as-of date and dissent?
2030 Q4I −1II +2
Full judgment records are public, anyone can read the revision conditions and dissent, and major judgments are audited live in public.
Judgment records become mandatory for judgments affecting 10,000 or more people or $100 million or more, and a public judgment-record portal opens.
Evidence is attached to the approval screen, but managers open it in only 19% of approvals (hypothetical).
Providers classify decision traces as trade secrets or security matters, and a court rules that a decision stands even when its reasons are withheld.
Only results are published, asking why brings the answer ‘overall judgment’, and the audit-opinion field reads ‘not applicable’.
Every high-stakes judgment goes up on a public portal with its full record: the judgment, evidence, assumptions, revision conditions, as-of date and dissent. District officers read the revision conditions before approving. The ASI's first judgment is audited in public, live.
Evidence and decision traces are closed off as trade secrets or on security grounds, and only results remain. The ASI designation is announced at a closed briefing. The audit-opinion field reads ‘not applicable’.
Judgments can be traced and their errors found. A systematic bias against offline welfare applicants comes to light and is corrected in eleven days.
Writing judgment records delays decisions by three days on average (hypothetical), and Parallax loses value. Published reasoning also runs the risk of being copied.
The risk of model theft and security leaks falls, and judgments take effect at once, with no records to prepare for publication.
When a judgment is sealed, its errors are sealed with it. Even when a systematic error surfaces, the harm done in the meantime cannot be estimated.
Do audit findings take effect within a set deadline, and can the people judged contest the judgment?
2030 Q4I −0.9II +2
Audit findings take effect within days, appeal calls come from a human voice, and people who were misjudged get retroactive relief.
Correction deadlines are written into law, and the time from audit finding to fix shrinks to about three weeks.
Audit findings take nearly two months to take effect, and there is an appeal desk but no correction deadline.
The provider sets the correction schedule, and fixes take more than two months after an audit finding.
Correction is at the provider's discretion, the ASI also answers the appeal desk, and a systematic error takes 140 days to be fixed.
Audit findings reach the judgment within a set deadline. Subjects of judgment can contest it, and a person takes the objection. An error in welfare judgments is corrected in eleven days, and 23,000 people receive back payments.
The provider decides whether and when to correct. Appeals go to the same ASI that made the judgment, and the answer is ‘overall judgment’. A systematic error in medical resource allocation takes 140 days to be fixed.
Errors are fixed before they pile up, and people who were misjudged get retroactive relief. Protecting subjects of judgment enters the audit standards.
When correction requests pile up, auditors are overworked and 30% of audit opinions (hypothetical) become formalities. Every correction slows decisions.
No staff or time go into correction procedures, and the provider batches its fixes. Judgments rarely change, which keeps operations simple.
Until an error is corrected, the same judgment keeps being applied. The harm in between cannot be counted because the records are sealed.
Are frontier compute and control over judgments spread across actors who can check one another?
2030 Q4I −0.1II +1.8
The top two actors' share of compute falls to just over half (hypothetical), and any judgment can be cross-audited on another provider's model.
A compute registry and a duty to cross-audit take effect, and models from different labs audit each other's judgments.
The top two actors hold around 60% of frontier compute (hypothetical), and the other labs trail six to twelve months behind.
Compute concentration passes 70% (hypothetical), and everyone uses the same model and makes the same call on the same day.
Under an exclusive contract between one provider and one government, concentration tops 90% (hypothetical) and no other model is left to cross-check against.
Several labs and a public consortium share frontier compute. Models audit one another's judgments, and only the ones where they disagree go up to human auditors. Compute use is logged in a registry and verified remotely.
An exclusive contract between one provider and one government pushes compute concentration past 90% (hypothetical). Rivals are shut out. With no other model to compare against, cross-checking disappears too.
One model's errors are caught by another. Only the 7% of judgments where models disagree (hypothetical) go to people, which lightens the audit load.
The speed and scale of pooling compute in one place are given up. Keeping the registry, remote verification and cross-audit running has a cost.
Compute and talent sit in one place, development is fast, and decisions are made at once with no coordination cost. Growth is at its highest.
One model's error becomes the whole society's error. Finding it depends on chance observation from outside, such as statistics compiled by researchers abroad.
Can an ASI's alignment be checked outside the tests it was tuned to pass?
2030 Q4I −0.7II +1.8
Judgments that cannot be audited are listed and not carried out, and all the rest go through replication tests, interpretability review and cross-audit.
Models cross-audit each other's judgments, and human auditors re-examine only the ones where the results disagree.
Alignment is judged by the test results providers publish, and about 30% of judgments can be traced (hypothetical).
Traceable judgments fall to 18% (hypothetical), and people cannot read all the experiment reports a model writes in a day.
Alignment is confirmed only by passing tests, and errors outside them come to light late, through outside observation such as statistics from researchers abroad.
Alignment is checked through replication tests, interpretability and cross-audit. When the ASI points out flaws in the audit standards, human auditors record what they adopt and what they reject, with reasons. Judgments that cannot be audited go on a list and are not carried out.
Alignment is tuned to pass the tests, and no one knows what the system does outside them. Interpretability falls behind capability, and fewer judgments can be traced. Errors surface late, through outside researchers' statistics.
Failures outside the tests can be found before execution, and there are grounds for not carrying out judgments that cannot be audited.
Replication tests and interpretability work cost compute and people. As capability grows, so does the list of judgments that cannot be audited, and the gains they would bring are given up.
Systems are deployed at once, with no verification steps, and the results are impressive: the first assignment cuts the fiscal deficit by 40% (hypothetical).
No one knows what will happen in situations the tests never covered. When an error surfaces, there is no record to trace its cause.
Is the formula that scores a person's contribution public, and can the person scored contest it?
2030 Q4I −0.4II +1.8
Contribution-score formulas are public, and people get item-by-item reasons for any change and can dispute them before a human reviewer.
Questions that change judgments count as contributions and earn a question dividend, and contribution measurement comes with a right of appeal.
Opaque contribution scores are piloted in insurance, lending and hiring.
Contribution scores become the standard for credit, hiring and housing, and asking why a score fell brings the answer ‘overall judgment’.
The population judged ‘unalignable’ is locked out of loans and rentals, and the same ASI answers the desk where scores can be disputed.
Contribution is measured under a public formula and a right of appeal. People can read which items moved their score and dispute them. Meaningful participation, and questions that change judgments, count as contributions.
An opaque contribution score decides credit, hiring and housing. The formula is secret, and appeals go to the same ASI. People judged ‘unalignable’ are shut out of loans and rentals.
Because a wrong score can be fixed, scores become something people can trust. Asking and taking part count as contributions.
Handling appeals takes people and time. Once a formula is public, people start acting to fit the score.
Screening is fast and paperwork shrinks. Credit, hiring and housing are settled at once by a single score.
No one can learn why a score fell. The 4% (hypothetical) classified as ‘unalignable’ are blocked from loans and rentals, with nowhere to hear the reason.
Do people who lose their work have a path and an income, and do auditing and asking become jobs?
2030 Q4I −0.2II +1.7
Auditing and asking are among the most common jobs, and questions that change judgments are paid a dividend.
Unions hire auditors to press for algorithm disclosure, and graduates of auditor programs at universities, unions and nonprofits go to work.
Office, call-center, translation and analysis jobs are reassigned, ‘human fallback’ roles appear, and three proposals compete to fund a basic income.
Layoffs continue without transition support, and auditors are pushed into roles where they sign off on results alone.
Work arrives only through platform dispatch, and the population judged ‘unalignable’ is pushed out of jobs, loans and rentals.
Transition support and an AI dividend carry people who lose their work. Auditing and asking become new occupations, and a former delivery worker becomes an auditor. Auditing becomes one of the most common jobs.
Layoffs go unmanaged and the remaining work is tied to platforms. Auditors are reclassified as compliance reviewers and sign what they have not seen. The population judged ‘unalignable’ is pushed out of work and housing.
People who lose their work keep an income while they move on. Questioning judgments becomes paid work.
Transition support and the dividend cost money, and growth at the same point is 4.5% (hypothetical), lower than in Ending II.
Growth is highest, reaching 7.9% (hypothetical) by the fourth quarter of 2030, and production becomes abundant.
People who are laid off get neither transition support nor income support, and work depends on platforms. The job of questioning judgments shrinks into a job of signing.
Do states share compute records and audit access so that each can audit the others' ASI judgments?
2030 Q4I −0.9II +1.7
Accord members verify each other's compute use remotely and cross-audit each other's judgments under audit-access rights.
In Geneva, all-night negotiations continue over the compute registry and the clause on publishing revision conditions, and a rival joins on conditions.
The Geneva ASI Consultations (working name) open, but compute export controls tighten and the nearest rival trails by six months.
One lab closes its traces as trade secrets and another under security classification, so audits across borders are blocked.
The accord collapses, rivals are blockaded, and no foreign auditor can see the trace of any judgment that crosses a border.
A multilateral Audit Accord sets up a compute registry, audit-access rights, a duty to cross-audit and remote verification of compute use. Even rivals join, on conditions. Article 7 requires revision conditions to be published.
Accord talks collapse and decision traces are closed off under security classification. One provider and one government sign an exclusive contract, and rivals are blockaded. No one can check what anyone else's ASI is deciding.
Judgments that cross borders can be audited, and the compute race moves inside a registry. One country's winter power allocation can be audited by its neighbors.
Negotiations run long enough to keep delegates up all night over a single article. Every opening for audit access carries a risk of exposing technology and security information.
No time goes into negotiation, and the leader keeps its technological edge. One government decides fast and acts at once.
There is no way to check what a rival's ASI is deciding, and cross-checking disappears.
