Search all dialogue chapters, core definitions, Lexicon terms, and key essays.
Catastrophe supplies authority. Where there is none, authority has to be manufactured — or borrowed from somebody willing to sign the harm.
Part CXI read fifteen selected cases, all of which produced an instrument, as evidence that catastrophe can supply authority for an already drafted rule. This Part takes up two questions without settling CXI's open tests: what corrects a harm that never arrives as an event, and what happens when the harm is a strategy rather than an accident. It proposes three manufactured inputs — a , an and a — and examines the second through the materials record since April 2025, where a licensing notice preceded rapid policy responses without a mass-casualty catastrophe. This is not evidence that no prior harm occurred or that those responses achieved durable correction.
The conversation had argued that harm can outvote an intention from outside. A harmful outcome does not logically refute the intention itself. The next turn was about what happens after the signal arrives: whether the instruments built on it respect the system they are meant to correct, or only the part of it that a ministry can reach. Johan put it as a reading of the energy transition rather than as a theory of governance.
“It's true that the signal created a change in attitude and with that rules, and the disasters an awareness which strengthen the attention and attitude for change. But, in the forms those changes were adapted and implemented ¨locally¨ were in my view, which is looking in the rearview mirror, reactive, focussed and neglecting or ignoring the ¨whole¨ and ¨continuity¨ which resulted in a lot of political maneuvering. All focussed on energy at top of the political agenda, while the whole and the continuity of the economies, depending on energy, suffered through the tax on the population which led to, and still leads to protests and protest parties.”
The observation contains a claim about incidence that the six-link function does not make on its own: that a correction applied at the point of political least resistance transfers its cost to the population, and that a population paying a new cost becomes a new harm signal. If that is right, remedies and mobilisation can form a feedback loop that reverses the direction of correction. That is a new dynamic hypothesis, not a claim that CXI established a thermostat or a fixed path to settlement.
This Part examines that reading in two domains: algorithmic governance, where the proposed preventive composition has not been evaluated, and strategic materials, where responses occurred without a mass catastrophe but economic disruption was already reported. Neither domain proves correction before all harm.
Part CXI's selected historical reading distinguished the design of a correction from the authority to adopt it: in several cases a disaster supplied permission for a rule already drafted. This is an interpretation of an instrument-selected sample, not a population-wide timing law. It leaves the origin of designs and failed mobilisation open.
Two answers are available. The first is to manufacture the missing inputs, which is a design problem and can be assessed as one. The second is to have the authority supplied from outside — by an adversary who names itself, dates its coercion and thereby performs the attribution that a distribution never performs.
A threshold replaces the witnessed injury, an aggregator replaces the crowd, and a ledger replaces the named culprit. The instruments exist in other domains; what does not exist is their composition in one regime, applied to deployments rather than to models.
A state that weaponises a dependency supplies the attribution for free, because it puts its name on the instrument. The correction then arrives in months — and the durability problem changes shape, because the party that supplied the authority also controls the clock.
Algorithmic harm is a rate. A thousand small misallocations of attention, credit, diagnosis or hiring, each below the threshold at which anybody acts. An event supplies three things for free — a witness, a crowd and a culprit — and a rate supplies none of them. Each substitution is an instrument that manufactures one, and each has a precedent in force somewhere.
| Input | Manufactured by | Instrument | Precedent |
|---|---|---|---|
| The event | A threshold | Registration, evaluation, logging, and a suspension power | AI Act 10^25 FLOP · DSA 45 million users · Basel capital · aviation certification |
| The crowd | An aggregator | A levy-funded ombudsman, mandatory liability cover, collective standing | Financial Ombudsman · employers' liability insurance · works councils |
| The culprit | A ledger | A decision record carrying the counterfactual, plus lineage and tiered access | Conflict-minerals due diligence · GDPR access right · content provenance |
These are proposed supports for CXI's original conditions, not replacement links: Signal, Attribution, Mobilisation, Remedy, Conversion and Durability. A threshold supports Signal, a ledger supports Attribution, and an aggregator supports Mobilisation. A suspension power is a possible Remedy; the authorised action owner exercises Conversion. Durability still requires a named reviser with an ongoing revision duty, distinct from the converter. None of the three inputs supplies that duty by itself.
In the proposed composition the three are not a menu. The threshold creates the duty to keep a record; the ledger supplies the evidence a case needs; the aggregator creates the demand that keeps the ledger honest. Remove one and the other two become theatre — a register nobody reads, a complaints desk with no evidence, a fee with no consequence.
A threshold converts a rate into an event by declaring that a measured aggregate crossing a line carries a consequence. It has one design virtue that no other instrument shares: it does not need to be believed to operate. A harm that has been counted against a published line is already a fact about the firm, whatever the firm says about it.
| Instrument | Measured | Consequence at the line |
|---|---|---|
| EU AI Act, Article 51 | Greater than 10^25 training FLOP | Rebuttable systemic-risk presumption; associated duties and evaluation |
| Digital Services Act, Article 33 | 45 million monthly recipients | Audit, risk assessment, data access |
| Basel capital rules | Risk-weighted assets | A capital charge |
| Aviation certification | Seats and passengers carried | Type certification before flight |
| The proposed layer | Decision volume × persons affected | Registration through to suspension |
These precedents illustrate public triggers for duties, not equivalent measures of harm or evidence that a number alone makes enforcement work. The AI Act uses training computation greater than ten to the twenty-five floating-point operations as a rebuttable presumption of high-impact capability, with review and other designation routes. It is not a measured injury threshold or a validation of the proposed layer.
What the existing instruments lack is the last rung of the ladder. Registration, evaluation and logging are duties. A suspension power exercisable without proving harm is a consequence. A regime that can only fine after the fact has published a price list, and a price list is a licence for firms that can afford it.
A firm that wishes to stay below a line will disaggregate: split a model family into deployments that each fall short of the trigger. The patch is not a lower number. It is a different unit — count the deployment population over a rolling period, the way an emissions inventory counts a fleet rather than a chimney, and write an anti-circumvention clause that aggregates affiliated deployments. Competition law has a century of experience with market definition under exactly this kind of splitting.
A harm distributed across millions of decisions produces no plaintiff, because no individual loss justifies a lawyer. This is not a defect of the harmed; it is a property of the arithmetic. The link is supplied by whoever is paid to aggregate, and the historical record is unusually clear about who that has been.
| Role | Who pays | What the money buys |
|---|---|---|
| Ombudsman | A levy on deployers above the threshold | A claim without a lawyer, published in volume |
| Insurer | A premium | A clean ledger before the harm, priced |
| Standing | Public funding | Cases brought on behalf of a class |
Payment protection insurance is the case that makes the argument concrete. Sixty-four million policies, individual losses of a few hundred pounds, and no possibility of a march. It corrected anyway, because a claims industry was paid to aggregate the harm and the ombudsman was free at the point of claim. Roughly £40 billion was returned. The mobilisation link did not fail because it was outsourced; it held because it was.
The insurer is the instrument worth adding, and it is the one nobody proposes. A mandatory liability cover for high-risk deployments creates a private buyer of attribution: the underwriter will not write the policy without a clean record, and the premium falls when the record is good. Attribution stops being a public good that somebody must be persuaded to fund, and becomes a condition of trading.
The ledger is where the design stops being a metaphor, and it comes down to one field that no current regime requires. Record, at the moment of decision, what the previous rule-based or human process would have decided for the same case. Without that field a denied loan is a data point among millions. With it, the denial is a measurable deviation from a baseline, and a baseline is something that can be argued about in front of a decision-maker.
| Field | Example | Why it matters |
|---|---|---|
| Inputs that mattered | The features actually used | Separates the model from the excuse |
| Model version | Hash and lineage | Ties the decision to a family and a date |
| Threshold applied | The cut-off in force | Shows the rule that was actually used |
| Outcome | The decision itself | The fact that can be disputed |
| Counterfactual | What the baseline would have decided | Turns a rate into an attributable difference |
Lineage travels with the model: training data, fine-tune ancestry, evaluation results. That is due diligence applied to data, and it is not a new idea — the conflict-minerals rules already require firms to trace an input they did not mine, on pain of disclosure. What is new is asking a model to carry the same passport as a shipment of cobalt.
Access has to be tiered or the regime will not survive its first negotiation. The supervisor sees everything. The aggregator sees the decision and its baseline on complaint. The public sees aggregates and audit summaries. The firm keeps its weights; the supervisor keeps the record. Verification runs through a third party in a secure enclave, so that the record can be checked without the model being handed over.
The precedents make the composition a bounded prevention proposal, not a validated regime. This record identifies no jurisdiction operating the proposed counterfactual ledger with a duty-bearing aggregator. The compute threshold in the AI Act attaches to model capability rather than to deployment, the unit proposed here for measuring harm.
The design makes three predictions for an eventual evaluation, rather than reporting successful prevention. The proposed time limits are design choices, and each prediction is stated so that it can fail.
There is also a political prediction, and it is the one that decides whether any of this is adopted. A regime used as a condition of market access functions as a tariff with a moral vocabulary, and it teaches every other jurisdiction that dependence on the regime is a vulnerability. A portable regime — a record format any competent authority can verify, mutual recognition of audits, no requirement that data leave its own jurisdiction — costs more to build and lasts longer, because it can be adopted rather than resisted.
Rare earths offer a response without a mass-casualty catastrophe, not an inversion of a fixed sequence: CXI already rejected that sequence. On 4 April 2025 China imposed licensing requirements on seven medium and heavy rare earths and related magnets. The cited record reports automotive production stoppages and supply-cost pressures, so prior economic harm cannot be ruled out. By June the G7 had announced an alliance; by July the Pentagon partnership included a price floor. These are policy responses in months, not measured effective or durable correction, and their endpoints are not comparable to the historical timing examples.
| Date | What happened | Link exercised | Lag |
|---|---|---|---|
| 4 April 2025 | Licensing on seven medium and heavy rare earths and their magnets | Signal and attribution at once | Weeks |
| June 2025 | G7 Critical Minerals Production Alliance created | Conversion, without a new treaty | Two months |
| 10 July 2025 | $110 per kilogram floor for NdPr over ten years, with $400m of equity | Remedy, underwritten by a state | Three months |
| 9 October 2025 | Controls extended to technology and a 0.1 per cent de minimis rule | Escalation of the same instrument | Six months |
| November 2025 | October measures suspended for one year; April licensing retained | Exposure of durability to external revision; not demonstrated durability | A switch |
The mechanism is not mysterious once the attribution link is read properly. A harm that arrives with a signature needs no forensic layer: the notice named the elements, the authority and the date. Everything downstream of attribution — mobilisation, remedy, conversion — had a target to aim at, and the target was a document.
The numbers behind the table are worth stating, because they explain why the remedy was expensive enough to need a state. China mines around 60 per cent of rare earths, refines around 91 per cent and produced about 94 per cent of sintered magnets in 2024; the IEA projects 76 per cent of refining still Chinese in 2035 on current plans. The American answer was a floor of $110 per kilogram for neodymium-praseodymium over ten years, with $400 million of equity in the cited announcement. The EU's Critical Raw Materials Act benchmarks — 10 per cent extraction, 40 per cent processing, 25 per cent recycling, and a 65 per cent single-supplier benchmark by 2030 — predate the April 2025 controls. They are not new April-to-July outcomes and do not by themselves establish achievement.
Two proposed extensions. First, a specific, dated and signed document may supply attribution and authority before a mass catastrophe; the record does not prove before all harm. Second, an externally controlled licensing regime can expose a response to later revision. The October measures were suspended until 10 November 2026, while the April regime remained. That is an observed policy switch, not a finding of durability or a measurement of the issuer's intent.
Semiconductors are deeper than rare earths, but depth is not what breaks the analogy. Distribution does. Rare earths had one choke point held by an adversary, and an adversary can be mobilised against. Semiconductors have five, four of them held by allies — and a democracy cannot mobilise against its own alliance.
| Chokepoint | Holder | Share | Substitute |
|---|---|---|---|
| Leading-edge logic | TSMC, Taiwan | 70–73 per cent of pure foundry | A decade and an ecosystem |
| Extreme ultraviolet lithography | ASML, Netherlands | Specialised supplier; exact share not pinned here | No demonstrated substitute in this record |
| Design software and core IP | American firms | Dominant | None in the medium term |
| High-bandwidth memory | Korean firms | Majority | Partly, at a cost |
| Wafers and photoresist | Japanese firms | Majority | Partly, slowly |
Three proposed consequences qualify the analogy. Allied dependence may have no single adverse actor to attribute, so prevention can require an insurance argument rather than a grievance. The remedy can be a coalition: TSMC announced intended U.S. investment of $165 billion in March 2025; separately, the November 2024 CHIPS award offered up to $6.6 billion against milestones for the then $65 billion three-fab plan. Announcements, awards and completed capacity are different endpoints. Technological change may also erode a remedy's usefulness before delivery; obsolescence is a risk, not an outcome demonstrated here.
Some earlier narratives foregrounded a one-sided response after a salient harm. That is not true of every earlier case: CXI includes pre-event instruments and continuously revised rules. The semiconductor chronology instead motivates a two-sided feedback hypothesis, in which one side's remedy may strengthen the other's mobilisation. The listed policy sequence is not itself a measurement of that causal effect; this targeted correction has not independently reverified every semiconductor event.
October 2022 — American controls on advanced logic and equipment. Intended as a constraint; received as a signal.
August 2023 — a Chinese seven-nanometre part ships in a phone, and China licenses gallium and germanium exports in the same month.
December 2024 — American controls extend to high-bandwidth memory, the input the Chinese build-out most needed.
April 2025 — China licenses seven heavy rare earths and their magnets, reaching the same automobile and defence programmes that had just been reorganised around a different dependency.
CXI's six conditions are not a fixed sequence or a promise of settlement. The new hypothesis is that in a bipolar competition, Conversion on one side may become a Signal on the other, coupling two correction processes into an escalating feedback loop. This does not show that every move satisfies all six conditions or that such systems cannot converge. Positive gain is an interpretation to test, not an estimated parameter or a universal law.
Reciprocal restraint over export controls is a proposed damping mechanism, not what CXI demonstrated as durability. CXI assigns durability to a named reviser and says mutual gain bought time, not durability. A reciprocal settlement would therefore need an enforceable revision duty as well as commitments by both sides. Whether it lowers the frequency and scope of new controls remains open; preferences and intentions are not measured here. A credible contrary result would be an implemented, enforced reciprocal agreement with that revision duty under which both frequency and scope continue to rise relative to a comparable pre-agreement baseline, after accounting for changed exposure and external shocks.
Johan's observation about the whole and continuity suggests three questions for a correction: its accounting boundary, implementation sequence and cost incidence. A narrow boundary can hide burdens, poor sequencing can put adaptive capacity at risk, and uncompensated costs can provoke mobilisation against a remedy. Those are hypotheses about possible mechanisms, not causal findings established by the comparisons below.
| Case | Measured | Comparison limit | Reading |
|---|---|---|---|
| EU net domestic GHG emissions, 1990–2024 | −40 per cent, EEA inventory | No matched consumption-based reduction verified here | Different endpoints and scopes; no quantified leakage inference |
| German energy-intensive output | Original −15.2 per cent figure omitted: exact dated source unpinned | Nuclear-exit and substitution claims require a counterfactual | Chronology alone does not establish an output or cost effect |
| French fuel tax, 2018 | Planned rise withdrawn in December | Original per-litre amount not pinned to a primary record here | Observed withdrawal; substitute and compensation effects untested |
| German household electricity | 47 cents, 2023 peak, as cited in the original note | 39.6 cents in 2025, in the same cited series | Price series and budget transfer are not a causal decomposition |
The first row is an accounting-boundary question, not proof that a supply chain moved. The EEA's release of 17 April 2026 reports net domestic greenhouse-gas emissions in 2024 at 40 per cent below 1990, excluding international aviation and navigation. The original 37 per cent figure was assigned to 2023 in the note, not 2020. SEI's cited 29 per cent figure is not identified as consumption-based, and its prose and Figure 8 disagree on whether the endpoint is 2020 or 2022. No matched territorial/consumption endpoints and gas scopes have been established here, so the numerical gap is withdrawn. These records do not establish emissions relocation or a policy-caused cost transfer.
The other rows motivate incidence and sequencing tests but do not identify their causes. The original German output percentage has no exact dated document in this record and is not used as a measured nuclear-exit effect. The cited French reporting records a planned fuel-tax rise withdrawn in December 2018; its exact per-litre amount is not verified here. The original German electricity source gives a 2023 peak of 47 cents per kilowatt hour and 39.6 cents in 2025, and describes the surcharge's transfer to the federal budget. A funding transfer is not a decomposition of prices, total costs or who ultimately bears them. Nuclear exit, fuel markets, subsidies and other causes require separate counterfactual evidence.
Read together with the three paths Johan described, these examples suggest possible trade-offs rather than proving that each political economy bought one link at the price of another. The original coal-capacity source is retained, but coal approvals, installed solar, carbon prices and household bills are unlike measures. The unpinned solar total, annual auction average and cross-country bill ratio are not used here. None of these figures alone establishes continuity, CXI's named-reviser durability, or a causal policy cost; those outcomes need comparable definitions and an explicit counterfactual.
A visible policy cost can become a harm signal to those who bear it. The proposed backlash hypothesis is that a price arriving before substitutes and without visible compensation increases mobilisation against the converter. It is not a deduction from CXI's six conditions or a single-cause account of voting. Comparable cases with compensation paid first, and controls for other economic and political changes, are needed to test the mechanism Johan suggested.
They compose precedents into a prevention proposal. This record identifies no evaluated jurisdiction operating that composition, so it does not demonstrate that the regime prevents harm or that the design works. The materials case supports observed responses without mass catastrophe, not an absence of prior harm.
It works where a state is willing to sign its coercion. It says nothing about harms with no author, which is the ordinary condition of algorithmic and environmental harm, and it cannot be relied on as a strategy — only recognised when it occurs.
The sequence of controls is documented. The claim that each remedy raised the other side's mobilisation is an interpretation of intent. A rate of escalation could be estimated from the interval between successive controls; this Part does not estimate it.
Territorial and consumption-based emissions differ for reasons that include import composition, gas coverage and inventory scope. The original figures use different endpoints and mislabel SEI's reduction; no comparable numerical gap is claimed here. Accounting differences do not establish leakage, nuclear-exit effects or cost causation.
Energy prices correlate with protest voting alongside migration, inflation and cultural change. The claim here is about incidence and backlash, not a single-cause account of party politics.
CXI selected cases that produced an instrument, not the population of harms. This Part supplies no scored no-instrument comparison and does not settle the Front Half of the function or the comparison with a two-variable ownership/reviser model. The proposals here cannot complete those missing tests by inference.
Four claims were put to the record, and each is graded here on the strength of what the documents actually show rather than on what the argument would prefer.
| Claim | Verdict | Bearing |
|---|---|---|
| Three inputs can be manufactured | Bounded prevention proposal | Functional precedents motivate the inputs; this record identifies no evaluated composition or demonstrated prevention. |
| Responses can occur without a mass catastrophe | Observed response; prevention proposed | The G7 alliance and July price-floor announcement followed April controls amid reported disruption. EU benchmarks predated them; no absence of prior harm or durable outcome is established. |
| The loop has positive gain | Plausible feedback hypothesis | The listed sequence motivates the mechanism; causal gain, non-convergence and reciprocal damping have not been estimated. |
| Boundary, sequence and incidence can affect outcomes | Bounded hypothesis; causal tests open | Accounting scopes and a reported tax withdrawal motivate tests, not a quantified leakage or policy-cost conclusion. |
Part CXI — The Correction Function read fifteen instrument-selected cases through authority versus design, the sign of conversion under ownership, and durability's named reviser. These are CXI's readings, not independently reconfirmed historical laws. This Part takes up two questions without resolving its Front Half or no-instrument comparison: what might supply authority when harm is a rate, and how can an external party's policy switches constrain a reviser?
Part CIX — The Unactionable Archive separated a discovery signal from an independently calibrated finding, and required a record that outlives the model that produced it. The Counterfactual Field is that record made specific: not merely preserved, but carrying the comparison that makes a distribution arguable.
Part CX — The Immunity Counterparty insisted that independence be operational rather than declared, because the action owner must not control the counterparty's budget, evidence access and escalation path. The Aggregation Duty is the same condition applied to harm rather than to interruption: whoever pays the aggregator determines whether the aggregation happens at all.
This register reuses the original sources and identifies the targeted corrections. Exact primary documents were checked for the emissions comparison, Parliament briefing, AI Act provisions and TSMC announcements; this is not an independent full review of every citation. Unpinned figures have been omitted or qualified rather than reconstructed. The accompanying research note preserves original citations, revised grades and falsifiers, and the remaining verification limits.
This Part proposes three inputs and examines a bounded materials response. It does not complete CXI's open tests. A further question is the one the design cannot answer while it assumes a custodian: what happens to a ledger when the weights are given away.
Proposed successor trajectory: the question above is sequenced behind measurement rather than answered directly. Part CXIII — The Instrument of Correction proposes that next step because a ledger needs a defined deployment population. The decentralisation question remains open; this reference is not a publication claim.
Proposed continuation in Part CXIII — The Instrument of Correction, which develops measurement proposals this design implies: a standard for a second judge, a deployment-population register, and an independent signal. This is not evidence of a deployed or validated instrument.