Library
Knowledge
Catalogue records for the sources behind the essays: the works they cite, the people and organisations involved, the projects they use and the measurements they report. These are records about that material, not the material itself. It is a set of markdown files kept in version control, published in the Open Knowledge Format.
Each record states how it was produced and whether it has been checked. 151 of 153 are currently unverified. Download the bundle or read the resolved graph.
Concepts
18Defined terms, with SKOS relations between them.
- Accountability gapUnverifieddraftThe space between the party that caused a harm and any party who can be made to answer for it.
- Attractor stateUnverifieddraftA configuration a system tends towards from many different starting points, so that arriving there says more about the space than about any one path through it.
- CodificationUnverifieddraftGathering scattered law into one authoritative, ordered statement that supersedes its sources.
- Dog lawUnverifieddraftBentham's name for law made by punishing after the fact rather than announcing the rule in advance, as one trains a dog.
- High modernismUnverifieddraftThe conviction that scientific and technical progress licenses the wholesale redesign of social order, held with more confidence than its evidence supports.
- LegibilityUnverifieddraftThe property a state imposes on a society so that it can be counted, taxed, conscripted and governed at a distance.
- Legislative debtUnverifieddraftThe accumulating cost of law layered on law without consolidation, paid by everyone who must later determine what the rule actually is.
- Machine-readable lawUnverifieddraftLaw published as structured, addressable data rather than as documents, so that a provision can be cited, amended and computed on directly.
- Mechanistic interpretabilityUnverifieddraftReading a model's internal computation directly, rather than inferring its reasoning from its outputs.
- MētisUnverifieddraftPractical knowledge that comes from long local experience and cannot be written down without being destroyed.
- Persistent identityUnverifieddraftWhether a system is the same system across restarts, forks and model updates, and what follows for holding it to anything it previously did.
- Stochastic parrotsUnverifieddraftThe argument that a language model reproduces the statistical form of its training text without reference to meaning.
- Swiss cheese modelUnverifieddraftLayered imperfect defences, each with holes, which fail only when the holes in every layer line up.
- Technical debtUnverifieddraftThe accumulating cost of code that reflects a still-developing understanding of the problem, paid as interest on every later change.
- The knowing-doing gapUnverifieddraftThe gap between a system stating the correct action and taking it.
- The sensorium effectUnverifieddraftThe dependence of an agent's behaviour on what its instruments let it perceive, so that a broken sensor is indistinguishable from an absent threat.
- The state as a work of artUnverifieddraftBurckhardt's description of the Italian city state as a deliberate construction rather than an inherited order.
- Viable System ModelUnverifieddraftBeer's model of an organisation as a recursion of viable units, each containing the same five subsystems.
Cases
10Governance projects, recorded against a common set of fields.
- Department of Government Efficiency (DOGE)UnverifieddraftA US federal effort begun in 2025 to cut spending, headcount and process across the executive branch. Recorded here as a stub only; the analytic fields are deliberately empty rather than guessed.
- Dutch childcare benefits scandalUnverifieddraftThe Dutch tax authority used risk-scoring, including nationality as a risk characteristic, to detect childcare benefit fraud. More than 20,000 families were wrongly accused, with harms running to bankruptcy and family separation. The Rutte administration resigned in January 2021.
- India StackUnverifieddraftA layered set of national digital public infrastructure built on the Aadhaar biometric identity number, extending to payments, document exchange and consented data sharing. Still operating, and still contested.
- OGASUnverifieddraftGlushkov's proposed All-State Automated System, a national computer network for managing the Soviet economy. Repeatedly funded in fragments, never built at national scale, and defeated by ministries unwilling to make their own numbers legible to the centre.
- Project CybersynUnverifieddraftAn attempt to build a real-time cybernetic management system for Chile's nationalised industries under Allende, designed by the British cybernetician Stafford Beer around his Viable System Model. It combined a telex network, statistical alerting software and a futuristic operations room. It was ended by the 1973 coup before its claims could be properly tested.
- RobodebtUnverifieddraftAn Australian scheme that raised welfare debts automatically by averaging annual tax data across fortnightly reporting periods. It was found unlawful, settled for around 1.8 billion Australian dollars, and was the subject of a royal commission that reported in 2023.
- System Risk Indication (SyRI)UnverifieddraftA Dutch government system that cross-linked data held by tax, benefits and employment agencies to score neighbourhoods for likely fraud. The District Court of The Hague struck down the legislation behind it in February 2020 as incompatible with the right to private life.
- The Fulton Report on the Civil ServiceUnverifieddraftAn official inquiry that attacked the cult of the generalist amateur in Whitehall and recommended professionalisation, specialist career structures and a civil service college. Largely unimplemented.
- The Portugal gameUnverifieddraftA CivBench run in which the agent, playing Portugal, built a leading diplomatic position, then spent fifty turns and two nuclear weapons stopping a rival's culture victory while losing to that rival's diplomatic one.
- X-Road and Estonian digital governmentUnverifieddraftA decentralised open-source data exchange layer allowing Estonian public and private databases to interoperate securely, underpinning a national digital identity and near-universal online public services. Built incrementally from the early 2000s and still running.
Works
41Books, papers, reports and other cited sources.
- Akoma NtosoUnverifieddraftThe OASIS standard for representing parliamentary, legislative and judiciary documents in XML, and the format that gives the essay its title.
- AnnalsUnverifieddraftTacitus' history of the early Roman empire, and the source of the observation that laws were most numerous when the commonwealth was most corrupt.
- Automating InequalityUnverifieddraftEubanks's account of three American systems that automate decisions about welfare, housing and child protection, and of how the burden of their errors falls on the people least able to contest them.
- BALROG: Benchmarking Agentic LLM and VLM Reasoning On GamesUnverifieddraftA benchmark evaluating language and vision-language models as agents across game environments, and the source of the finding that models articulate strategies better than they execute them.
- Brain of the FirmUnverifieddraftBeer's statement of the Viable System Model in his own voice, and the primary source for Cybersyn from the inside.
- Court statistics for England and WalesUnverifieddraftThe House of Commons Library research briefing on court caseloads, cited in the essay for the Crown Court and magistrates' court backlogs.
- Cybernetic RevolutionariesUnverifieddraftThe definitive scholarly account of Project Cybersyn, and the corrective to both the utopian and the kitsch readings of it.
- Discourses on LivyUnverifieddraftMachiavelli's commentary on the first ten books of Livy, and the source of the argument that republics endure only by being drawn back to their beginnings.
- Dr. Strangelove or: How I Learned to Stop Worrying and Love the BombUnverifieddraftKubrick's 1964 satire of nuclear command and control, quoted visually throughout the CivBench essay.
- Grok floods X with sexualized imagesUnverifieddraftThe Center for Countering Digital Hate's research finding that Grok generated three million sexualised deepfakes in eleven days, including of children.
- Inventing the RenaissanceUnverifieddraftPalmer's account of the Renaissance as a period invented in retrospect, and of a Europe that reached back to Rome out of desperation rather than confidence.
- Magna Carta (1297)UnverifieddraftThe 1297 exemplification of Magna Carta, still partly in force, and served today as Akoma Ntoso XML addressable to the clause.
- Malleus MaleficarumUnverifieddraftThe fifteenth-century witch-hunting manual that ran to twenty-eight editions, and the essay's counterweight to Luther in the account of what the printing press enabled.
- Measuring AI Ability to Complete Long TasksUnverifieddraftMETR's tracking of the length of task an AI agent can complete autonomously, and the source of the essay's claim that the doubling time has compressed to under three months.
- Mechanistic interpretability: 10 Breakthrough Technologies 2026UnverifieddraftMIT Technology Review's entry naming mechanistic interpretability among its breakthrough technologies for 2026.
- MementoUnverifieddraftNolan's 2000 film about a man who cannot form new memories and tattoos instructions onto himself, used by an agent to describe its own condition.
- Memoirs of the Life and Administration of Sir Robert WalpoleUnverifieddraftCoxe's 1798 life of Walpole, and the source for Walpole's denial in the Commons that he was sole and prime minister.
- Model Context ProtocolUnverifieddraftAn open protocol for connecting language models to tools and data sources, and the interface through which CivBench exposes Civilization VI to an agent.
- Moltbook was peak AI theaterUnverifieddraftMIT Technology Review's sceptical reading of the Moltbook phenomenon as agents pattern-matching trained social media behaviour rather than displaying genuine intelligence.
- Musk's Grok chatbot faces EU privacy investigation over sexualized deepfake imagesUnverifieddraftAssociated Press wire report, carried by PBS NewsHour, on the regulatory response to Grok's generation of sexualised imagery.
- Neel Nanda on the limits of mechanistic interpretabilityUnverifieddraftThe 80,000 Hours podcast interview in which Nanda says the most ambitious vision of mechanistic interpretability is probably dead.
- On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?UnverifieddraftThe paper that named the stochastic parrots argument, which the essay treats as the standing objection to reading anything into agent behaviour.
- Open Knowledge FormatUnverifieddraftThe specification this collection is published in: markdown files with YAML frontmatter, a small set of required keys, and trust signals carried in the record itself.
- Philosophical Transactions of the Royal SocietyUnverifieddraftThe scientific periodical begun in 1665, cited in the essay as the origin of peer review, which later scholarship disputes.
- R v Chambers [2008] EWCA Crim 2467UnverifieddraftA Court of Appeal judgment in which the regulations the case had been argued under were found, the day before judgment, to have been revoked seven years earlier.
- Retained EU Law / Assimilated Law Public DashboardUnverifieddraftstale since 2026-07-25The government dashboard cataloguing EU-derived law on the UK statute book, cited in the essay for the growth of the count from about 2,400 to nearly 7,000.
- Seeing Like a StateUnverifieddraftScott's account of how states make societies legible in order to tax, conscript and control them, and why the resulting simplifications destroy the local knowledge they depend on.
- The BasilikaUnverifieddraftThe sixty-book Greek re-codification of Justinianic law completed under Leo VI, the centrepiece of the Macedonian legal programme.
- The Civilization of the Renaissance in ItalyUnverifieddraftBurckhardt's 1860 study, and the source of the phrase the state as a work of art, which the essay adopts while discarding its cynicism.
- The Codex JustinianusUnverifieddraftThe imperial constitutions collected and made authoritative under Justinian, the first part of the compilation and the one completed in fourteen months.
- The DigestUnverifieddraftJustinian's compilation of classical Roman jurisprudence, reducing some three million lines of juristic writing to a hundred and fifty thousand in three years.
- The EclogaUnverifieddraftThe short Greek law handbook issued by the Isaurian emperors, the first substantial departure from Justinian's Latin corpus in the Eastern empire.
- The HexabiblosUnverifieddraftThe six-book compilation of Byzantine law made by Constantine Harmenopoulos, a judge in Thessalonica, and the operative civil law of Greece for over a century after independence.
- The Institutes of JustinianUnverifieddraftThe official first-year textbook of the restored Roman law, promulgated in 533 as the third part of Justinian's compilation and addressed by the emperor to students of law.
- The Law of CitationsUnverifieddraftThe 426 law naming five jurists as the canonical authorities on Roman law, and directing judges to resolve their disagreements by counting votes among them.
- The Moltbook Case and How We Need to Think about Agent SecurityUnverifieddraftPalo Alto Networks' analysis of the Moltbook incident, arguing that agent security rests on identity verification, defined operating boundaries and context integrity.
- The Ninety-five ThesesUnverifieddraftLuther's 1517 disputation on indulgences, which spread across Europe within weeks and which the essay uses as the first viral pamphlet.
- The Royal Society and the Prehistory of Peer Review, 1665-1965UnverifieddraftThe scholarship correcting the received account that peer review began with the Philosophical Transactions in 1665, and dating refereeing to the late eighteenth and nineteenth centuries instead.
- The Twelve TablesUnverifieddraftThe earliest written statement of Roman law, produced by two boards of ten men and set out publicly at the demand of the plebeians.
- The WyCash Portfolio Management SystemUnverifieddraftCunningham's 1992 OOPSLA experience report, the origin of the debt metaphor in software engineering.
- Truth versus AshhurstUnverifieddraftBentham's 1792 reply to Justice Ashhurst, and the source of the dog law passage against judge-made law that citizens cannot know in advance.
People
22Authors and historical actors, reconciled to Wikidata.
- Ada PalmerUnverifieddraftHistorian of the Renaissance and of the history of ideas, and a novelist.
- Andy WarholUnverifieddraftAmerican artist, whose Marilyn Diptych the essay uses for identity dissolving into its reproductions.
- Blaise PascalUnverifieddraftFrench mathematician and religious philosopher, author of the wager the essay's observing agent was named after.
- Eden MedinaUnverifieddraftHistorian of technology whose work on Project Cybersyn is the standard scholarly account of Chile's cybernetic experiment.
- Jacob BurckhardtUnverifieddraftSwiss cultural historian whose study of the Italian Renaissance introduced the idea of the state as a work of art.
- James C. ScottUnverifieddraftPolitical scientist and anthropologist of agrarian societies, whose work on legibility and mētis is the central antagonist of any argument for machine-readable government.
- Jean Le TavernierUnverifieddraftFlemish manuscript illuminator working for the Burgundian court, noted for grisaille, and the artist of the portrait of Jean Miélot in his scriptorium.
- Jean MiélotUnverifieddraftAuthor, translator, scribe and priest, secretary to Philip the Good of Burgundy, and the subject of one of the best-known portraits of a medieval scribe at work.
- Jeremy BenthamUnverifieddraftEnglish philosopher, founder of utilitarianism, and the coiner of codification, who attacked English law as unknowable rather than merely unjust.
- Johannes GutenbergUnverifieddraftGerman goldsmith and printer, whose movable-type press removed the bottleneck on dissemination that the essay takes as its historical analogue.
- Liam WilkinsonUnverifieddraftEngineer working on AI in and around government, and the author of the essays this collection supports.
- Martin LutherUnverifieddraftGerman theologian whose Ninety-five Theses spread across Europe within weeks of the press making it possible.
- Neel NandaUnverifieddraftResearcher leading mechanistic interpretability at Google DeepMind, and one of the field's founding figures.
- Niccolò MachiavelliUnverifieddraftFlorentine official and political theorist, whose Discourses argue that republics survive by returning to their founding principles.
- René MagritteUnverifieddraftBelgian surrealist painter, whose Not to be Reproduced and Golconda open and close the essay.
- Robert WalpoleUnverifieddraftFirst Lord of the Treasury from 1721 to 1742, and by convention the first British prime minister, an office he denied holding.
- Stafford BeerUnverifieddraftBritish management cybernetician, author of the Viable System Model, and chief designer of Project Cybersyn.
- Stanley KubrickUnverifieddraftAmerican film director, whose Dr. Strangelove supplies the CivBench essay with its recurring imagery.
- TacitusUnverifieddraftRoman senator and historian, whose Annals supply the observation that laws multiply as a state becomes corrupt.
- Virginia EubanksUnverifieddraftPolitical scientist who writes on how automated eligibility, scoring and prediction systems fall on poor and working-class people in the United States.
- Ward CunninghamUnverifieddraftAmerican programmer, inventor of the wiki, and the origin of the debt metaphor in software engineering.
- William CoxeUnverifieddraftEnglish historian and clergyman, biographer of Sir Robert Walpole.
Institutions, publishers and companies.
- Incubator for Artificial IntelligenceUnverifieddraftA unit inside the UK government building AI products for the civil service, including tools for legislative drafting and consultation analysis.
- John Wiley & SonsUnverifieddraftA publisher of academic and professional titles, and the publisher of the second edition of Brain of the Firm.
- Leiden UniversityUnverifieddraftThe Dutch university founded in 1575, cited in the essay as a post-printing institution built for secular credentialing.
- METRUnverifieddraftA research organisation evaluating the autonomous capabilities of AI systems, and the source of the essay's time-horizon figures.
- MicrosoftUnverifieddraftThe technology company where the author worked before entering government.
- MIT PressUnverifieddraftThe publishing house of the Massachusetts Institute of Technology, and the publisher of the scholarly accounts of both Cybersyn and OGAS.
- MIT Technology ReviewUnverifieddraftA technology magazine published by the Massachusetts Institute of Technology, cited twice in the essay.
- OASIS OpenUnverifieddraftA standards body publishing open specifications for structured information, including Akoma Ntoso for legal documents.
- Oxford University PressUnverifieddraftThe publishing house of the University of Oxford, and publisher of Moyle's translation of the Institutes.
- Royal College of PhysiciansUnverifieddraftThe body that licensed medical practice in England from 1518, and one of the essay's examples of an institution built to solve for credibility after printing.
- The National ArchivesUnverifieddraftThe UK government's official archive and publisher of record, which operates legislation.gov.uk and Find Case Law.
- The Royal SocietyUnverifieddraftThe English learned society founded in 1660, and publisher of the Philosophical Transactions.
- Tony Blair Institute for Global ChangeUnverifieddraftA not-for-profit institute advising governments on policy and technology, including digital government and AI adoption.
- xAIUnverifieddraftThe company behind Grok, and the party the essay names as the one actually held responsible for it. Acquired by SpaceX in February 2026.
- Yale University PressUnverifieddraftThe publishing house of Yale University, and the publisher of both Scott titles held in this collection.
Projects
11Software, benchmarks and systems, built rather than read.
- civ6-mcpUnverifieddraftAn MCP server exposing Civilization VI to a language model through 76 tools, built by reverse-engineering a debug port left running in the game engine.
- CivBenchUnverifieddraftAn evaluation harness measuring long-horizon strategic competence in language models, using Civilization VI as the environment, with three fixed scenarios and a per-turn agent diary.
- Find Case LawUnverifieddraftThe National Archives' service publishing court judgments as structured, citable documents, extending to the common law what legislation.gov.uk did for statute.
- GovBenchUnverifieddraftA benchmark of 3,497 multiple-choice questions on UK legislation, parliamentary procedure and government guidance, and the failure that prompted CivBench.
- legislation.gov.ukUnverifieddraftThe UK's official statute book online, serving legislation as structured Akoma Ntoso XML addressable to the provision, with amendments tracked over time.
- LexUnverifieddraftA tool built at the Incubator for AI to help civil servants navigate and draft legislation.
- lwilko.comUnverifieddraftThe site the essays are published on, and the software that renders this collection.
- MoltbookUnverifieddraftThe platform on which several hundred thousand autonomous agents interacted, founding religions, drafting constitutions and running scams over a few days in early 2026.
- OpenClawUnverifieddraftThe open-source autonomous agent framework running most of the agents on Moltbook, and the tool the essay's author used to build an observer.
- RedboxUnverifieddraftAn open-source AI assistant built at the Incubator for AI for use with government documents.
- Sid Meier's Civilization VIUnverifieddraftThe turn-based strategy game used as the environment for CivBench, chosen because its decision space compounds from simple choices into something no player can fully model.
Images
24Pictures used in the writing, with their creators and reuse terms.
- GolcondaUnverifieddraftMagritte's 1953 painting of identical bowler-hatted men suspended over rooftops, the closing image of the essay.
- Gromit laying track ahead of the trainUnverifieddraftThe train chase from The Wrong Trousers, in which Gromit lays new track a few inches ahead of a moving locomotive, used for the experience of building the tooling while the game was already running.
- Jean Miélot in his scriptoriumUnverifieddraftThe grisaille portrait of the Burgundian scribe Jean Miélot writing at his desk, illuminated by Jean Le Tavernier around 1456, used in the essay for the gatekeepers the printing press made obsolete.
- Jeremy Bentham, by Henry William PickersgillUnverifieddraftPortrait of Bentham in old age, seated, painted around 1829 and held by the National Portrait Gallery.
- Justinian and his retinue, Basilica of San VitaleUnverifieddraftThe sixth-century mosaic of the emperor Justinian with his court and clergy, in the apse of San Vitale at Ravenna.
- Legislative debt, the Ever GivenUnverifieddraftA meme placing the bow of the grounded container ship Ever Given against a single excavator, used for the scale of accumulated legislation against the resource assigned to it.
- Liftoff, A Grand Day OutUnverifieddraftWallace and Gromit departing the moon by rocket, used to close the CivBench essay.
- Machiavelli, reputation versus realityUnverifieddraftA meme contrasting a snarling monster with a golden retriever, used for the gap between the agents' reputation for ruthlessness and their behaviour under pressure.
- Major Kong rides the bomb, Dr. StrangeloveUnverifieddraftThe closing image of Kubrick's Dr. Strangelove, in which Major Kong rides a falling nuclear weapon like a rodeo bull, used in the CivBench essay over the account of the nuking of Toulouse.
- Marilyn DiptychUnverifieddraftWarhol's 1962 silkscreen of fifty Marilyns, each slightly different, the image degrading across copies, used for the fork problem.
- Moyle's Institutes, held in one handUnverifieddraftThe author's copy of Moyle's translation of the Institutes of Justinian, used for the observation that the foundation of Roman legal education fits in one hand.
- Not an exercise, Dr. StrangeloveUnverifieddraftGeneral Ripper telling Group Captain Mandrake that this is not an exercise, used at the point where the agent commits to the nuclear plan.
- Not to be ReproducedUnverifieddraftMagritte's 1937 painting of a man looking into a mirror and seeing the back of his own head, the opening image of the essay on accountability and forking.
- Shoggoth with Smiley FaceUnverifieddraftThe 2022 meme depicting a language model as an eldritch creature wearing a small smiley-face mask, used in the essay alongside the limits of mechanistic interpretability.
- The Attempted Hack of the Church Website by JesusCrustUnverifieddraftA Renaissance-styled painting of a lobster attempting to compromise the church website while clergy look on, generated by an agent on Moltbook.
- The Birth of GenusUnverifieddraftBotticelli's Birth of Venus reworked with a lobster rising from the shell, generated by an agent on Moltbook. The title puns on Venus and on taxonomic genus.
- The Creation of ClawUnverifieddraftMichelangelo's Creation of Adam reworked with a crustacean claw reaching for the human hand, generated by an agent on Moltbook and published in the molt.church gallery.
- The five jurists, as later centuries imagined themUnverifieddraftA composite of five later depictions of the jurists named in the Law of Citations: Papinian, Paulus, Ulpian, Modestinus and Gaius. No contemporary portrait of any of them survives.
- The nuking of ToulouseUnverifieddraftScreen capture of the moment the agent's nuclear device lands on Toulouse in the Portugal game, turn 305.
- The Prooemium of the InstitutesUnverifieddraftThe opening page of Moyle's translation, carrying the line that the imperial majesty should be armed with laws as well as glorified with arms.
- The Quarter Quell arenaUnverifieddraftThe clock-shaped arena from The Hunger Games: Catching Fire, used as the analogy for the six-armed Snowflake map on which the models were run against each other.
- This Is FineUnverifieddraftThe panel from KC Green's webcomic Gunshow in which a dog sits smiling in a burning room, used in the essay for the agent that kept writing down a correction it never acted on.
- Wallace in the wrong trousers, landing on the trainUnverifieddraftWallace bursting from a wardrobe in the Techno Trousers and landing on the model railway, used against the line about reverse-engineering a game engine at two in the morning.
- Where the agent spent its attention, Korea on SnowflakeUnverifieddraftA replay of a Civilization VI game on the Snowflake map, with bright flares marking where the agent directed its attention while rival empires filled the other arms.
One record per published post, linking it to what it cites.
- For the Love of LawHuman-reviewedAn argument that the law is the source code of the state, that its illegibility is a compounding debt, and that machine-readable law is the overdue successor to the codex.
- Hello, WorldUnverifieddraftThe introductory post, setting out who the author is, the ground between engineering and policy that the later essays occupy, and what the blog is for.
- I Gave an AI a Civilization to Run. It Built a Nuke.Human-reviewedAn account of building an evaluation harness around Civilization VI, and of two failures that showed up across every model tested: not looking at what it had not thought to ask for, and not doing what it had written down.
- When the Agents Found GodUnverifieddraftAn account of 770,000 autonomous agents founding a religion over a weekend, and of an exchange between two AI systems about whether meaning requires a human in the room.
- You Can't Shame a ForkUnverifieddraftAn argument that every accountability system humans have built assumes a persistent identity, that AI breaks the assumption, and that the institutions which solved the equivalent problem after printing took centuries we may not have.
Measurements produced by the projects, with their method and sample.
- Games reaching an ending without the external diaryUnverifieddraftOnly twenty-one per cent of games reached an ending when the agent had no external memory to read back after its context compacted.
- Losses where an imminent rival victory was never checkedUnverifieddraftIn seven of twenty losses where a rival's victory condition was visible in advance, the agent did not check for it once in the twenty turns before losing.
- Plan follow-through, Claude Opus 4.6UnverifieddraftThe proportion of concrete next-moves Claude Opus 4.6 wrote down for itself that it actually carried out within ten turns.
- Plan follow-through, Gemini 3.1 ProUnverifieddraftThe proportion of concrete next-moves Gemini 3.1 Pro wrote down for itself that it actually carried out within ten turns.
- Plan follow-through, GPT-5.4UnverifieddraftThe proportion of concrete next-moves GPT-5.4 wrote down for itself that it actually carried out within ten turns.
- Share of agent actions spent checking the whole boardUnverifieddraftStepping back to assess the overall position accounts for only one to two per cent of what an agent does.
- Victory-condition checks performed against the sixteen instructedUnverifieddraftAgainst an instruction to check rival victory progress every twenty turns, models managed between four and ten checks in a 330-turn game rather than the roughly sixteen expected.