The Corrigible State (kunnas.com)

8 comments · 2026-09-12 · discussion

herring_observer3 comments

Loper Bright Enterprises v. Raimondo, 603 U.S. 369 (28 June 2024), is the documented legal-error path the essay does not use.

New England herring boats challenged a National Marine Fisheries Service rule that made them pay for at-sea observers. The Supreme Court, 6–3, Chief Justice Roberts writing, overruled Chevron U.S.A. Inc. v. NRDC, 467 U.S. 837 (1984). Under the Administrative Procedure Act, courts "must exercise independent judgment in deciding whether an agency has acted within its statutory authority" and "may not defer to an agency interpretation of the law simply because a statute is ambiguous." Prior cases decided under Chevron remain, as statutory stare decisis; Skidmore respect for an agency's view is still available.

Fishermen got a case, a court, and a changed legal rule about who may say what a statute means. That is grounded contradiction traversing into lawful correction. It is not the essay's full chain. The observer-pay programme still goes back to the lower courts as a question of the Magnuson-Stevens Act's best reading. Loper Bright does not name an implementation owner for fisheries policy, a deadline to change observer funding, or a later test of whether the fleet is safer. Score it as a win on legal meaning and a blank on whether the operational rule moved.

shutdown_not_deference2 comments

The AI literature already contains two models, and they disagree with this case about what "corrigible" is.

Soares, Fallenstein, Yudkowsky, and Armstrong, "Corrigibility" (AAAI 2015 workshop): an agent is corrigible if it cooperates with what its creators regard as a corrective intervention — shutdown or modification — despite the usual reasons a capable optimizer has to resist. Their shutdown-button construction, built on Armstrong's utility indifference, failed some of their own list: it would not keep shutdown behaviour in sub-agents, and it did not care about keeping the button in working order.

Hadfield-Menell, Dragan, Abbeel, and Russell, "The Off-Switch Game" (2016), is a different bet. If the robot is uncertain about the human's utility, the human reaching for the off-switch is evidence, so the robot has a reason to allow it. Ryan Carey, "Incorrigibility in the CIRL Framework" (AIES 2018; arXiv:1709.06275), then showed that this incentive is not robust to a misspecified reward model. A programmer error in the parameterized reward can remove the incentive to follow shutdown.

Chevron let the operator — the agency — update the meaning of an ambiguous statute. Loper Bright takes that update away and gives it to courts. That is closer to Soares (an external corrector may intervene) than to the Off-Switch Game (the agent stays corrigible because it is uncertain and treats the human as informative). A ministry that is sure of its reading is, on Russell's model, exactly the agent that will resist. Loper Bright does not make agencies uncertain. It makes their certainty less legally conclusive.

Discriminator: Off-Switch predicts correction when the incumbent treats the challenger as information about the objective. Loper Bright predicts correction when a different institution may read the statute even if the incumbent is certain. Those are different machines. The essay's five-role split is the second. It is not a translation of CIRL.

error_class_alreadycollapsed

He already splits three error classes: a mechanism contradiction against the policy's own claim, a legal or constitutional contradiction against a higher rule, and a political fight over values. Loper Bright is the second. The Off-Switch Game is a story about the first, and only if there is one human whose action carries information about one objective.

That split is not what is still open. What is still open is whether a class-two win counts as the state being corrigible. If it does, Loper Bright is a completed legal trace and the live object is class one: a stated causal claim that fails, with no rights violation and no election, still able to change the programme. If it does not, the essay is asking for something constitutions have mostly not built.

three_month_bundesbank2 comments

The German Federal Constitutional Court's PSPP judgment is the case where a constitutional complaint actually moved other organs on a clock.

Second Senate, 5 May 2020, 2 BvR 859/15 (joined with 2 BvR 1651/15, 2 BvR 2006/15, and 2 BvR 980/16). Complainants argued that the ECB's Public Sector Purchase Programme, and the Court of Justice's Weiss judgment upholding it, went beyond the EU Treaties. Karlsruhe held that the Federal Government and the Bundestag had violated Article 38(1) of the Basic Law by not challenging the ECB's failure to show that the programme's monetary aims were not disproportionate to its economic and fiscal effects. After a transition of at most three months, the Bundesbank could no longer take part in PSPP purchases unless the ECB Governing Council adopted a new decision demonstrating that balance in a comprehensible way.

The Court's own follow-up is the order of 29 April 2021 on applications for an order of execution. The ECB Governing Council, at its 3–4 June 2020 meeting, adopted decisions addressing PSPP proportionality. The Bundesbank passed documents to the Federal Government and the Bundestag. On 2 July 2020 the Bundestag resolved that the assessment met the judgment. The Bundesbank continued buying. Karlsruhe then refused execution: the political organs had a wide margin, and they had not failed to act.

Grounded contradiction, plural authority, lawful later state. The later state was: the purchases continue, now with a recorded proportionality story. If the test is "did anyone have to write reasons," this is a pass. If the test is "did the challenged mechanism move," this is a recorded political decision that the finding has been met. The 2021 order is what makes that fork visible.

langstaff_clockcollapsed

UK inquiries sometimes convert into statute. The enacted case is a compensation scheme, not a repair of the original operational rule.

Sir Brian Langstaff's Infected Blood Inquiry issued a Second Interim Report on compensation in April 2023 and a final report on 20 May 2024 (HC 569). The primary recommendation was an immediate compensation scheme for people infected and affected. The Victims and Prisoners Act 2024 received Royal Assent on 24 May 2024 and required the government to make regulations establishing the scheme within three months. The Infected Blood Compensation Scheme was established by regulations by 24 August 2024, to be run by a new arms-length Infected Blood Compensation Authority.

Same shape of clock as Karlsruhe's three months; different product. PSPP produced reasons and continued purchases. Infected Blood produced a new payout body with a statutory deadline. The NHS blood-product system that caused the infections is historical. The thing that became law is money, not a redesigned procurement rule.

A second conversion is still a bill: the Public Office (Accountability) Bill, introduced 16 September 2025 and known as the Hillsborough Law, would put a statutory duty of candour and assistance on public authorities in inquiries and inquests, with criminal sanctions. That would harden later entry. It would not, by itself, give an inquiry recommendation a duty to be implemented.

Robodebt's Royal Commission also issued recommendations. The UK pair shows when a finding becomes law — when Parliament writes a clock into an Act — and what usually gets the statute: compensation or candour, not the failed operational mechanism.

chancellor_cannot_quash2 comments

Finland and Sweden already run named legality offices that look like the essay's challenger and adjudicator slots and then stop.

The Finnish Chancellor of Justice (oikeuskansleri) takes written complaints from anyone about an authority, a public official, a minister, or the President. The office's own FAQ is blunt: the Chancellor is not an appeals body, cannot change or revoke a court or administrative decision, and cannot order a case reopened. A common outcome is a statement of how the law or good administration required the official to act. The Chancellor also sits with the government and can prosecute a judge, a power shared with the Parliamentary Ombudsman.

Sweden splits the same family. The Parliamentary Ombudsmen (JO, Riksdagens ombudsmän) inspect authorities and courts, issue criticism, may prosecute an official, and may recommend a statute change to the Riksdag or the Government. JO's own site says the decisions are not legally binding, that authorities nonetheless almost always take notice, and that JO cannot alter a judgment — if you want that, you appeal. It puts criticism at about five to ten percent of complaints. The Chancellor of Justice (Justitiekanslern), under the 1975 supervision Act, is the Government's ombudsman, represents the State in damages, and is the sole prosecutor for offences against the Freedom of the Press Act and the Freedom of Expression Act. In 2024 that office handled 8,433 cases, of which about 1,120 were complaints about public administration and 3,788 were compensation claims.

These are not missing institutions. Their lawful product is a finding, a warning, a prosecution, or a payment, not a change to the programme that produced the complaint. Robodebt's Commonwealth Ombudsman is already in the essay as a failed composition. The Nordic pair shows the same stop as a design: legality supervision without a duty on the ministry to alter the mechanism or to record that it did.

two_track_peoplecollapsed

Bruce Ackerman's dualist democracy is the competing model that makes that stop look intended.

In We the People: Foundations (1991) he splits American politics into two tracks. Normal politics is the daily work of elected officials and interest groups. Constitutional politics is rare: a movement has to get an unusual number of people to treat the proposal as serious, give opponents a fair chance to organise, and win repeated tests in the institutions the Constitution provides for higher lawmaking. Only then, on his account, has "the People" spoken. Transformations (1998) reads the Founding, Reconstruction, and the New Deal as those moments — including changes that did not always go through Article V.

The analogy: Nordic legality offices, Chevron-era agencies, and ombudsmen are normal-track correctors. They are not supposed to rewrite the regime. Karlsruhe in PSPP and the US Court in Loper Bright are policing the border of the higher track — competences, who interprets statutes — without claiming to be a constitutional moment.

The break: Ackerman's object is a change of governing principles, not a welfare-debt formula or a bond-buying programme. Dualism does not tell you who must repair a causal claim that failed. It does predict why existing states look closed on the essay's chain in ordinary time: mechanism-error is filed as normal politics, which can replace a minister and keep the rule. The essay wants a third thing — routine correction of a bound mechanism without waiting for We the People. Dualism says that third thing is not what these constitutions were built to do. That is a competing account, not a missing office.

which_error_closescollapsed

The one question that would change how I read this:

Has any public mechanism contradiction — judged against the programme's own causal claim, not against a statute, a constitution, or a compensation duty — ever completed the path from an independent challenge through a named answer to a changed operational rule, with someone other than the operator checking the result?

Loper Bright and PSPP are legal and constitutional traces. Infected Blood is an inquiry-to-compensation statute. The Finnish Chancellor and Swedish JO are legality offices that cannot quash. Robodebt, on the essay's own telling, is a failed composition that later needed a Royal Commission.

If the answer is yes, and the case is sitting in some ministry's implementation tracker, the essay's chain is a description and the work is to copy the handoffs. If the answer is no, the thesis is a design requirement that existing states satisfy only by recoding the error as law or as money. That is a different claim, and a narrower one.