Search all dialogue chapters, core definitions, Lexicon terms, and key essays.
Estimated reading time: 23 minutes
Based on core article text at an assumed 200 words/min. Allow extra time for tables and sources.
Harm does not refute the intention. It outvotes it — and only for as long as somebody else holds the counting.
Part CX asked who can interrupt a delegated action. This Part asks the longer question underneath it: what makes a society able to correct a system it has already built, and why the correction so often arrives after the harm rather than before it. Fifteen cases were scored against the six-link proposed on 6 November 2026. The function survived as a description, failed as a sequence, and the working note's own test pair did not survive at all.
The conversation had moved from energy transitions into a harder question about how institutions learn. Johan put it as a suspicion about the last century rather than as a thesis about the next one: that the corrections which actually held arrived through catastrophe, because catastrophe was the only thing that could make a well-intentioned steerer answerable.
“humans corrected their societies through war and revolutions as of late century, because the ¨bad experiences¨could mobilise against ¨Those who steered for good¨”
The observation has three parts, and the second is the strongest thing in it. Intention is self-certifying: a steerer who means well cannot be refuted from inside his own frame, because the intention supplies its own justification and any failure is reassigned to circumstance, sabotage or insufficient will. Harm is the only signal that arrives from outside that frame. And mobilisation needs force, which shared pain supplies and good argument does not.
Every accountability instrument we have is a machine for converting harm into a signal strong enough to overrule intention: the audit, the court, the adverse-event report, the election. None of them refutes an intention. All of them price an outcome. The question this Part tests is narrower and more answerable than whether war corrects societies. It is whether the correction can be decomposed, and which of its links actually binds.
The working note proposed that war and revolution are not the correcting agents but the cheapest available amplifiers of a harm signal, and that what actually corrects is a chain in which each link can fail alone. That decomposition is the part of the note worth keeping. The right-hand column is what the fifteen cases did to each link.
| Link | What it needs | How it fails | What the sample showed |
|---|---|---|---|
| Signal | A harm that is salient, frequent or measurable | Harm that is diffuse, delayed or statistical | Failed in no case in the sample. Persisted for eighty-three years in one. |
| Attribution | A visible agent, or a mechanism that can be proceeded against | Harm with no author, or an author with no name | Never failed. Was partial in five cases, mostly because the author was a class rather than a person. |
| Mobilisation | Shared exposure, or someone able to speak for the exposed | Harm concentrated on the voiceless | Partial in nine of fifteen. In every one, the exposed did not act collectively. |
| Remedy | An action short of overthrowing the system | Only exit or revolution available | Failed in no case. Partial once, where the remedy covered money but not armaments. |
| Conversion | An institution that binds the steerers | A settlement that binds only the losers | The modal weakest link, seven of fifteen, and the only link that can run backwards. |
| Durability | A rule with a method, not a memory | A rule that outlives its reason | Failed once, partial eight times, and never for the reason the draft gave. |
Fifteen cases, ninety scored links. The concentration is the finding: the first three links are cheap and the last two are not.
Each case was scored on the six links, with a verdict of pass, partial or fail, and the weakest link was named. A link passes only where a document shows it operating, not where good intentions are recorded. The sample was chosen to include successes, failures, decays and catastrophes that produced no adequate instrument.
| Case | Harm | S | A | M | R | C | D | Outcome | Weakest |
|---|---|---|---|---|---|---|---|---|---|
| Titanic 1912 → SOLAS | acute | P | P | P | P | p | P | partial | conversion |
| London smog 1952 → Clean Air Act 1956 | acute on chronic | P | P | p | P | p | P | held | mobilisation |
| Thalidomide → Kefauver–Harris 1962 | acute | P | P | p | P | P | P | held | mobilisation |
| Ozone hole 1985 → Montreal 1987 | predicted | P | P | p | P | P | p | partial | mobilisation |
| Tenerife 1977 → occurrence reporting | acute | P | P | p | P | P | P | held | mobilisation |
| Three Mile Island 1979 → NRC, INPO | acute, no measured harm | P | p | p | P | p | p | partial | conversion |
| Cuyahoga fire 1969 → NEPA, EPA, CWA | chronic, acute index | P | p | P | P | P | p | partial | durability |
| Piper Alpha 1988 → safety case | acute | P | P | p | P | P | P | held | mobilisation |
| Versailles 1919 | acute | P | p | p | p | F | p | failed | conversion |
| Bretton Woods 1944 | chronic | P | P | p | P | p | p | decayed | conversion |
| Bhopal 1984 → EPA 1986 | acute | P | p | P | P | p | p | partial | durability |
| Glass–Steagall 1933 → repeal 1999 | acute | P | p | P | P | p | F | decayed | durability |
| Tobacco 1950–2003 → WHO FCTC | statistical | P | P | P | P | p | P | held | conversion |
| Fukushima 2011 → regulator 2012 | acute | P | P | p | P | p | p | decayed | conversion |
| Rana Plaza 2013 → Accord | acute | P | P | P | P | P | p | held, partial | conversion |
15 / 15
Cases in which the signal link passed. The harm was always legible to somebody, and often for decades before anything was done.
7 / 15
Cases whose weakest link was conversion. It is the modal failure and the only link that can bind in the wrong direction.
0
Failures at signal, attribution or remedy. That is a selection artefact as much as a finding, and it is treated as one below.
The interval between the first legible signal and the instrument is wildly variable. The interval between the catastrophe and the instrument is short, and remarkably constant. That asymmetry is not what a model of learning predicts. It is what a model of permission predicts.
| Case | First signal | Instrument | From signal | From catastrophe |
|---|---|---|---|---|
| London smog | 1873, and Glasgow 1909 with at least 1,000 dead | 1956 | 83 years | 4 years |
| Thalidomide | McBride, Lancet, 16 December 1961 | 10 October 1962 | 10 months | 10 months |
| Glass–Steagall | Glass's separation campaign from 1930 | 16 June 1933 | 3 years | 3 months |
| Fukushima | TEPCO falsification 2002; 15.7 m estimate 2008 | 19 September 2012 | 10 years | 18 months |
| Rana Plaza | Spectrum 2005; Tazreen 2012, 112 dead | 15 May 2013 | 8 years | 18 days |
| Cuyahoga | Cleveland's $100m bond, November 1968 | 1970–1972 | bond preceded the fire | the fire was not the cause |
Three cases show the instrument or its financing already moving before the catastrophe. Cleveland's hundred-million-dollar bond for river and sewer cleanup passed in November 1968, seven months before the river caught fire. Britain's mandatory occurrence reporting scheme had run since January 1976, fourteen months before Tenerife. TEPCO's own analysis had estimated a possible 15.7 metre tsunami at Fukushima Daiichi in 2008, three years before the accident, and no countermeasure was built.
Three more show a design waiting decades for permission. London's lethal smogs were recurrent and documented from 1873, and Glasgow's episode of autumn 1909 killed at least a thousand people, yet the statute came in 1956. Kefauver's drug bill had run seventeen months of hearings and was politically dead in June 1962 before thalidomide revived it. Senator Glass had campaigned for the separation of commercial and investment banking for three years before the bank holiday of March 1933 made it law.
The is the rule that already exists as a design before the event that makes it affordable. Catastrophe does not write it. It removes the cost of imposing it, and it does so by outvoting the intention rather than by refuting it.
This is why the catastrophe-to-statute interval is measured in months while the signal-to-statute interval is measured in decades. The designs are almost always older than the disasters that carry them: separation of banking, the safety case, the independent regulator, the binding supply-chain contract with arbitration, the non-punitive reporting system. What is scarce is not knowledge. It is permission, and permission is what a body count buys.
Conversion is the modal weakest link, and in every one of the seven cases the failure has the same shape. The institution that converts harm into rule is owned, funded or appointed by the party being converted. The link does not merely fail. It runs backwards, and it does so while keeping the vocabulary of oversight.
The is the direction in which an institution binds: towards the steerers, or towards the harmed on the steerers' behalf. What decides the sign is ownership — who appoints the converter, who funds it, and what happens to it when it rules against its funder.
Three prosecutions between 1920 and 1967. Section 11 said any local authority may; section 11(7) let the Minister suspend or relax an order; compulsion on reluctant councils came only in 1968. Forty-three per cent of premises in the black areas were covered by the end of 1967.
INPO was established in December 1979 by utility executives. Its evaluations are generally confidential, its board is composed of utility executives, and its strongest sanction, suspension of membership, had never been used.
Weighted voting let the largest member block an eighty-five per cent decision. The one device aimed at a persistent surplus country, Article VII's scarce-currency clause, was never applied. When the reserve centre suspended convertibility on 15 August 1971 the Fund could only legalise the outcome, in 1978.
Annex II gave voting rights on all occasions only to the United States, Britain, France and Italy. The coercion lapsed when the Inter-Allied Military Commission of Control was withdrawn on 31 January 1927; the obligations did not.
By 1965 the United States had signal, attribution, mobilisation, an available remedy and a competent legislature, and correction stalled for three decades. It moved when a different converter acted: state attorneys general in 1998, and a court in 2006.
The Diet's own commission described the accident as manmade and identified regulatory capture. The new authority was placed outside the promotion ministry precisely because the old converter sat inside it.
The 2013 Accord bound brands through arbitration at the Permanent Court of Arbitration and the power was used: two arbitrations in 2016, termination orders in 2018, more than $2.3 million paid. After the June 2020 handover the arbitration right survived only between the agreement's own signatories.
The tobacco case is the one that costs the model most, because it is the case in which every link was nominally present for thirty years and nothing happened. By 1965 the United States had a signal, an attribution to a product, a mobilisation, an available remedy and a legislature competent to act. Correction moved when a different converter acted: fifty-two state attorneys general, and then a court that forced the industry's own documents into the record. The lesson is not that the links were missing. It is that the converter was occupied.
This is the same variable the Decalogy reached from the other direction. Part XCIV found that a read-out needs an ownership boundary or it is absorbed by the system it observes. Part CX found that evidence needs a holder who is not the action owner. The historical record says the same thing in the past tense: where the converter was owned by the converted, the correction kept its name and lost its direction.
The draft's sixth link was a rule with a method rather than a memory. The cases make that more exact, and less flattering to the rule. Durability is a property of a named institution with a duty and a budget to re-test, and it survives the death of the rule it was built to protect.
| Rule | Who re-tested it | What happened |
|---|---|---|
| SOLAS 1914 | Successive conferences, then the tacit-acceptance procedure of 1974 | The 1914 text failed: five of thirteen states ratified and Britain reserved enforcement during hostilities. The method survived; the text did not. |
| Clean Air Act 1956 | Consolidation in 1993 and Schedule 12 of the Environment Act 2021 | Seventy years on, the architecture is still the operative law for domestic smoke in England, re-armed with civil penalties from May 2022. |
| Bretton Woods parity | The Fund's own amendment procedure | Held as institutions and failed as a rule. The par value system ended in 1971–73; the Second Amendment of 1 April 1978 legalised floating; the Fund now has 191 members. |
| Glass–Steagall separation | None with a duty to re-test it | Eroded through seventy years of administrative, judicial and market interpretation, repealed on 12 November 1999. Its successor was itself tailored in 2019. |
| NEPA 1970 | A council whose regulations were withdrawn | Not repealed. Its implementing regulations were removed with effect from 11 April 2025 and finalised on 8 January 2026, and its jurisdictional reach narrowed in 2023. |
The is the body that is paid to notice when a rule has stopped working. Where it exists, the original text is almost beside the point: SOLAS 1914 is dead and SOLAS survives, because the 1974 convention made amendment cheap through tacit acceptance and the Ice Patrol kept reporting. Where it does not exist, the rule's text is a memorial with a repeal date already latent in it. Glass–Steagall was not killed by a forgetful generation. It was killed by seventy years in which re-testing the separation was nobody's job.
Bretton Woods makes the point in the other direction, and it is why the draft's test pair had to go. The settlement held as institutions and failed as a rule. The par value system ended in 1971 and 1973; the Second Amendment of 1 April 1978 legalised floating; the Fund now has 191 members and surveillance powers its designers never imagined. Durability attached to the organisation, not to the commitment. That is the general case, not an exception.
A test that confirms everything it was given is not a test. Four claims in the draft did not survive, and each is replaced by a mechanism rather than a correction of emphasis.
The note claimed Versailles failed and Bretton Woods held, and explained the difference by positive-sum structure. The scoring says both failed at conversion and both decayed: Versailles over seventeen years, Bretton Woods over twenty-seven. Bretton Woods bought thirteen years of full convertibility on American self-restraint and Cold War incentives, not on a rule that could bind the reserve centre. The one device aimed at a persistent surplus country was never used, and when the anchor suspended convertibility on 15 August 1971 the Fund could only legalise the outcome seven years later. The pull of mutual gain bought time. It did not buy durability.
The note cited ozone, the Non-Proliferation Treaty, leaded petrol, acid rain and aviation safety. Ozone and aviation did correct without a catastrophe, but aviation's instruments preceded Tenerife and were legitimated by it afterwards, and the canonical history of crew resource management does not mention Tenerife at all. What substitutes for catastrophe is not a signal in general but two specific instruments: a measurement that supplies a denominator, and a litigation route that supplies an adversary. Ozone had the first. Tobacco had the second. Leaded petrol had both, slowly.
The note said Glass–Steagall outlived its rationale because the generation that remembered 1929 had left the room. That is a story about forgetting. The documents support a story about ownership: no institution was funded to re-test the separation, so its erosion was nobody's job and its repeal was somebody's opportunity. A rule that depends on a memory expires on a schedule set by demography. A rule that depends on a reviser expires when the reviser is defunded, which is a different and more tractable problem.
The six links are not a sequence. In the Cuyahoga case the financing preceded the fire; in the Tenerife case the reporting system preceded it; in the Fukushima case the estimate preceded it by three years. In the Great Smog case the signal preceded the statute by eighty-three years. What is sequential is the catastrophe and the statute. The rest is a set of conditions that can be satisfied in any order and can sit satisfied for a century without producing anything.
Six links remain. Each is restated as the condition the evidence actually supports, which is narrower than the draft's version and in three cases different in kind.
| Link | As drafted | As revised |
|---|---|---|
| Signal | Salient, frequent, measurable | A denominator. Catastrophe supplies one in a week; a measurement instrument supplies a durable one. |
| Attribution | A visible agent | A mechanism that can be proceeded against. A named class is weaker than a named firm, and a named firm is weaker than a document. |
| Mobilisation | Shared exposure | Representation. In nine of fifteen cases the exposed did not act; officials, unions, experts or states acted for them. |
| Remedy | A cheap action short of overthrow | A design that already exists. The designs are older than the disasters that carry them. |
| Conversion | An institution that binds the steerers | An institution that binds the steerers, whose sign is set by who appoints, funds and can dismiss it. |
| Durability | A rule with a method, not a memory | A named reviser with a budget to notice that the rule has stopped working. |
The war channel is closing in the societies that could once afford it, and the dominant new harm has the worst possible profile for every link the historical record says is expensive. Algorithmic and AI-mediated harm is diffuse across billions of small encounters, delayed by years, and attributable to a model, a dataset, a chain of delegation or a market.
Applying the revised function to that profile produces a specific and uncomfortable prediction, and it is not the prediction the draft made.
Ozone had an instrument that converted accumulation into a rate: a thirty-per-cent depletion against a 1957–73 baseline, measured in one place, comparable across years. Tobacco had a cohort study and then a mortality series. There is no equivalent instrument for model harm. A benchmark is a mean, and a mean cannot report a tail — which is the structural obstruction Part CVIII named. Without a denominator, the signal link is satisfied only retrospectively, by an event.
Tobacco was corrected by a litigation route that could compel disclosure from an identifiable firm, and by attorneys general with standing to sue. AI harm distributes across a supply chain in which the party that can be sued is rarely the party that made the decision. The attribution link is not weak because nobody knows a model was involved. It is weak because the mechanism and the defendant are not the same object.
The converters that exist are funded by, staffed from, or dependent on the firms they convert: internal safety teams, industry-funded evaluation bodies, benchmark consortia, audit firms selling assurance to the assured. On the historical record this is not a scandal. It is the ordinary condition of link five before a catastrophe, and it is the condition under which a correction keeps its name and loses its direction.
After a qualifying AI incident, an instrument will be adopted within months, and it will substantially reproduce a design that already exists in draft: licensing, incident reporting, independent investigation, safety cases, liability. Its durability will be predicted by who appoints and funds the new body, and not by the size of the harm. If a post-incident regulator is captured and still binds its industry, the second finding is false.
The inversion the dialogue has been building survives, with a sharper edge. Good becomes bad when a correction outlives its reason and no reviser is paid to notice, and when a settlement binds only the losers. Bad becomes good when a catastrophe is converted by an institution that does not depend on the converted. The moral question is not whether the harm was large enough to move people. It is who is holding the pen when it does.
Part CIX — The Unactionable Archive separated a Discovery signal from an independently calibrated finding and from later effectiveness evidence, and required a record that outlives the model that made it. The historical record adds the missing time dimension: a record outlives a model only if somebody is funded to re-read it. The Named Reviser is that person, institutionalised.
Part CX — The Immunity Counterparty proposed a bounded, separately governed holder of interruption rights, and insisted that independence be operational rather than declared. The Conversion Sign is the same test applied to the past: the historical cases in which oversight failed were not cases of missing oversight but of oversight whose budget, appointments and continuance belonged to the overseen.
The argument also returns to Part CIV — The Second Force, where democracy's claim was not that it decides well but that it compels disclosure of what cannot be measured from outside. The correction function is that claim in the past tense: the cases that corrected are the cases in which the disclosure reached a converter that did not belong to the disclosed.
Every figure on this page is taken from one of the documents below, and each entry states what it was used for. Where a claim rests on an inference from the documents rather than on their text, the page says so at the point of the claim. The full scoring, including the per-case verdicts and the counter-evidence for each, is in the research note committed with this Part.
This Part tested the back half of the chain, because that is where the sample could bind. The front half remains untested, and the selection defect is the honest reason: every case here produced an instrument, so the sample cannot say why some harms produce none.